Artificial Intelligence for Enhancing Indoor Air Quality in Educational Environments: A Review and Future Perspectives

Abstract

Indoor Air Quality (IAQ) in educational environments is a critical determinant of students’ health, well-being, and learning performance, with inadequate ventilation and pollutant accumulation consistently associated with respiratory symptoms, fatigue, and impaired cognitive outcomes. Conventional monitoring approaches—based on periodic inspections or subjective perception—provide only fragmented insights and often underestimate exposure risks. Artificial intelligence (AI) offers a transformative framework to overcome these limitations through sensor calibration, anomaly detection, pollutant forecasting, and the adaptive control of ventilation systems. This review critically synthesizes the state of AI applications for IAQ management in educational environments, drawing on twenty real-world case studies from North America, Europe, Asia, and Oceania. The evidence highlights methodological innovations ranging from decision tree models integrated into large-scale sensor networks in Boston to hybrid deep learning architectures in New Zealand, and regression-based calibration techniques applied in Greece. Collectively, these studies demonstrate that AI can substantially improve predictive accuracy, reduce pollutant exposure, and enable proactive, data-driven ventilation management. At the same time, cross-case comparisons reveal systemic challenges—including sensor reliability and calibration drift, high installation and maintenance costs, limited interoperability with legacy building management systems, and enduring concerns over privacy and trust. Addressing these barriers will be essential for moving beyond localized pilots. The review concludes that AI holds transformative potential to shift school IAQ management from reactive practices toward continuous, adaptive, and health-oriented strategies. Realizing this potential will require transparent, equitable, and cost-effective deployment, positioning AI not only as a technological solution but also as a public health and educational priority.

1. Introduction

The building sector is responsible for approximately 30–40% of global final energy consumption and nearly 30% of energy-related CO₂ emissions [1,2,3,4]. Consequently, research and policy have largely emphasized energy efficiency measures, renewable integration, and the deployment of smart building technologies [5,6,7,8,9]. To this end, sustainable buildings have become a cornerstone of global strategies to mitigate climate change, reduce energy demand, and enhance human well-being [10,11]. Yet, sustainability also encompasses the health and comfort of occupants, making indoor environmental quality a critical dimension of building performance. Among its components, indoor air quality (IAQ) is of particular concern because it directly influences human health, productivity, and cognitive function [12,13].

Educational environments require special attention. Children spend up to 90% of their time indoors, and schools are among the most densely occupied building types [14]. Poor IAQ in classrooms has been consistently linked to respiratory illnesses, asthma, allergies, fatigue, and impaired cognitive outcomes [15,16,17]. Empirical evidence from European and North American schools shows that carbon dioxide (CO₂) concentrations frequently exceed the recommended 1000 ppm threshold, with many classrooms reporting values above 2000 ppm due to inadequate ventilation [18,19]. High levels of particulate matter [20,21], volatile organic compounds (VOCs) [19], ozone [22], and nitrogen dioxide [23] are also common in urban schools, further compromising children’s health. These findings underline that IAQ is not only a comfort issue but also a public health priority and a key determinant of sustainable school design.

Unlike other indoor settings, such as offices, hospitals, or residential buildings, educational environments present a unique convergence of challenges for indoor-air-quality management. Classrooms typically accommodate high occupant densities with limited control over ventilation by the users themselves—conditions that intensify pollutant accumulation and CO₂ buildup during teaching hours [17]. Moreover, children are physiologically more vulnerable than adults to air pollutants, and their cognitive performance has been shown to deteriorate under elevated CO₂ concentrations and poor air conditions [24,25]. From a technological perspective, most schools operate without advanced Building Management Systems (BMS), relying instead on manual ventilation or rudimentary controls, which constrains the deployment of conventional smart-building solutions. In addition to these technical and operational aspects, the adoption of AI in schools also raises important social and ethical considerations. Issues such as data privacy, transparency, and stakeholder trust among teachers, parents, and administrators require careful attention to ensure responsible implementation and long-term acceptance of AI-based IAQ systems. In this context, Artificial Intelligence (AI) offers a promising pathway to bridge this gap by enabling low-cost, data-driven monitoring and adaptive control, tailored to the dynamic occupancy and behavioral patterns of educational settings. These features distinguish AI applications for IAQ in schools from those designed for workplaces or healthcare facilities, where infrastructure, resources, and regulatory frameworks differ substantially.

Poor IAQ in classrooms has been consistently linked to respiratory illnesses, asthma, allergies, fatigue, and impaired cognitive outcomes [15,16,17]. Empirical evidence from European and North American schools shows that CO₂ concentrations frequently exceed the recommended 1000 ppm threshold, with many classrooms reporting values above 2000 ppm due to inadequate ventilation [18,19]. High levels of particulate matter (PM_2.5) [20,21], volatile organic compounds (VOCs) [19], ozone [22], and nitrogen dioxide [23] are also common in urban schools, further compromising children’s health. These findings underline that IAQ is not only a comfort issue, but also a public-health priority and a key determinant of sustainable school design.

It is important to note that most existing AI applications for IAQ management in educational environments have concentrated on a limited number of measurable pollutants—principally carbon dioxide, particulate matter (PM_2.5/PM₁₀), and volatile organic compounds (VOCs). This focus reflects the practical availability of sensor data, the strong relationship of these variables with ventilation efficiency and occupancy, and their relevance to students’ comfort and cognitive performance [26,27,28,29,30,31,32,33]. By contrast, the use of AI for monitoring biological contaminants, bacteria, or viral aerosols remains limited due to challenges in real-time sensing, data quality, and ethical constraints associated with personal exposure monitoring. Nevertheless, recent advances in optical and spectroscopic sensing integrated with deep-learning models, coupled with improved ventilation analytics, suggest that AI-assisted detection of airborne pathogens could become feasible in future school-based studies [34,35].

Traditional IAQ monitoring methods face significant shortcomings. Periodic inspections offer only episodic snapshots of classroom conditions, while reliance on subjective perception often leads to underestimation of pollutant levels [36]. Even with the deployment of continuous sensor networks, technical challenges such as calibration drift, measurement noise, and heterogeneity across devices undermine reliability. This complexity calls for analytical approaches capable of managing high-frequency, multivariate, and dynamic datasets that characterize real-world classroom environments.

Artificial intelligence (AI) has emerged as a promising framework to address these challenges. Classical machine learning (ML) techniques—such as decision trees, support vector machines (SVMs), and random forests—have demonstrated strong predictive capability in pollutant trend estimation and classification of IAQ states [5]. More advanced deep learning (DL) architectures, including convolutional neural networks (CNNs), recurrent neural networks (RNNs), long short-term memory (LSTM) models, and autoencoders, extend these capabilities by automatically learning spatiotemporal dependencies, filtering noise, and enhancing anomaly detection. These methods enable a range of tasks critical for IAQ management: (i) sensor calibration using regression and feature engineering to correct biases in low-cost devices; (ii) pollutant forecasting (e.g., CO₂, particulate matter with an aerodynamic diameter of 2.5 μm or less, PM_2.5) to support preemptive ventilation control; (iii) anomaly detection in time-series to flag system malfunctions or atypical occupancy; and (iv) multi-objective optimization of heating, ventilation, and air conditioning (HVAC) systems, balancing IAQ improvements with energy efficiency.

Recent reviews have proposed a systematic framework that classifies AI applications for air quality monitoring into five domains: sensor calibration, anomaly detection, air quality index estimation, short-term forecasting, and integrated control [37,38]. This framework underscores both the opportunities and the challenges of deploying AI, particularly with respect to data reliability, scalability, and system integration. By organizing diverse applications into coherent categories and evaluating their strengths and limitations, this body of work demonstrates how AI can move IAQ assessment from episodic and reactive approaches toward continuous, predictive, and adaptive management.

Despite these advances, the application of AI to IAQ in educational environments remains fragmented and underdeveloped. First, most reported studies are localized pilot projects confined to individual schools or small samples under specific climatic and infrastructural conditions [39,40]. Such narrow scopes restrict generalizability and make it difficult to evaluate the scalability of AI-based solutions across diverse educational contexts. Second, while many models achieve high predictive accuracy in controlled settings such as [39,41], few address issues of long-term sustainability [42]. Performance often deteriorates without frequent retraining, and persistent problems of sensor calibration and data drift remain unresolved. These limitations undermine the robustness and reliability of AI systems in real-world deployments. Third, the integration of AI with legacy HVAC and Building Management Systems (BMS) has received limited attention [43,44], even though automated ventilation control depends on such interoperability in most schools. Without seamless integration, AI tools risk functioning as isolated analytics rather than as actionable decision-support systems. Fourth, the social and ethical dimensions of AI adoption—including privacy protection, data security, transparency, and trust among teachers, parents, and administrators—are seldom addressed in technical studies [45,46]. Yet, these considerations are critical for acceptance in sensitive environments such as schools. Finally, to date, there has been no systematic review consolidating international evidence on AI for IAQ in schools. Existing reviews tend to emphasize energy efficiency [47,48] or general air quality [24], leaving a gap in understanding how methodological advances, practical challenges, and contextual constraints intersect in educational settings.

The objective of this review is to critically examine the applications of AI for IAQ management in educational environments, with a focus on both methodological innovation and practical deployment. Specifically, the review seeks to:

(a)

Synthesize methodological advances in AI-based IAQ monitoring, prediction, and control, including the use of (ML), (DL), and hybrid models;

(b)

Assess outcomes across diverse geographical and socio-technical contexts, drawing on twenty representative international case studies that span North America, Europe, Asia, and Oceania;

(c)

Identify systemic barriers—technical (e.g., data scarcity, model generalizability), economic (e.g., cost of deployment and maintenance), and ethical (e.g., privacy and trust)—that constrain the broader adoption of AI in schools;

(d)

Highlight pathways for future research and implementation, emphasizing scalability, sustainability, and equity in educational settings.

The methodology adopted for the literature review followed a selective but comprehensive strategy rather than a systematic database search. This approach was chosen to reflect the interdisciplinary and emergent nature of research on AI applications for IAQ in educational environments, where publications often appear across diverse outlets and formats. To ensure methodological transparency and reproducibility, the literature search was conducted in Scopus, Web of Science, and ScienceDirect databases covering the period 2013–2025, using combinations of keywords such as “artificial intelligence,” “machine learning,” “deep learning,” “indoor air quality,” and “educational buildings.”

The main criteria guiding inclusion were: (i) relevance to AI applications for IAQ in schools; (ii) methodological rigor and empirical validation; (iii) diversity of approaches, ranging from classical ML to advanced DL and hybrid models; and (iv) practical applicability, including real-world case studies and integration with HVAC or smart campus systems. Selected studies were then categorized by pollutant type, AI technique, and application objective (e.g., forecasting, anomaly detection, or integrated control). In doing so, the review provides not only a state-of-the-art synthesis of current practice but also a transparent and structured framework for advancing AI-enabled IAQ management in educational environments.

The innovation of this review lies in its integrative and critical perspective. While most prior studies have examined AI in buildings primarily for energy efficiency or in general indoor environments [44,47], few have addressed the specific challenges of schools, where children’s heightened vulnerability, high occupancy rates, and limited resources necessitate tailored approaches [39,40]. This review advances the field by offering a comparative synthesis of real-world AI applications in educational settings, explicitly linking technical performance metrics—such as forecasting accuracy, anomaly detection, and adaptive control—with broader systemic issues of scalability, equity, and privacy. Beyond framing AI as a technological solution, the review positions its adoption as both a public health imperative and an educational priority. By examining AI within the interconnected domains of sustainability, health, and education, this review extends beyond conventional technical surveys to systematically evaluate the potential of an integrated approach to IAQ management. The novelty of this work, therefore, lies not only in synthesizing algorithmic advances but also in demonstrating that AI’s impact should be evaluated based on its capacity to deliver equitable, transparent, and sustainable improvements to learning environments. In doing so, this review charts clear directions for future research and practice, identifying pathways to advance from fragmented pilot studies toward globally scalable and impactful solutions.

The remainder of this paper is structured as follows: Section 2 provides an overview of IAQ challenges in educational environments; Section 3 reviews the main categories of AI methods (ML, DL, and hybrid models) applied to IAQ; Section 4 synthesizes findings from twenty representative case studies worldwide; and finally, Section 5 offers a critical discussion of implications, limitations, and future directions.

2. Indoor Air Quality in School Environments: Key Considerations and Determinants

Indoor air quality (IAQ) refers to the condition of indoor air in relation to occupants’ health, comfort, and performance, encompassing pollutant concentrations, odors, and the adequacy of ventilation [49]. In school environments, this concept acquires heightened importance because children spend extended hours indoors, exhibit higher inhalation rates per body weight compared to adults, and are physiologically more vulnerable to environmental stressors. A substantial body of evidence links inadequate IAQ in classrooms to respiratory illnesses, allergy symptoms, absenteeism, and impaired cognitive functions, thereby influencing both health and learning outcomes [14,24,50,51,52].

These challenges are compounded by structural and operational characteristics of schools, such as high occupant density, limited ventilation rates, and aging infrastructures with outdated heating, ventilation, and air-conditioning (HVAC) systems [53]. Furthermore, many educational buildings were constructed with limited consideration of modern energy and IAQ standards, leading to situations where efforts to improve ventilation and pollutant removal directly conflict with energy conservation goals [16]. As a result, ensuring satisfactory IAQ in schools requires not only identifying the predominant pollutants and their sources but also evaluating the building’s ability to balance air exchange, filtration efficiency, and energy performance [54]. This dual perspective places IAQ management in schools at the intersection of public health and smart building design, underlining the need for innovative approaches, such as sensor-based monitoring and AI-driven optimization, to deliver safe, healthy, and sustainable learning environments [55,56,57].

2.1. Classification and Sources of Indoor Air Pollutants

Indoor air pollutants in schools originate from both indoor emission sources and outdoor infiltration, with their impacts often amplified by high occupant density, inadequate ventilation rates, and outdated building infrastructures [22,23,58,59,60,61]. Among the most widely studied indicators of indoor air quality, carbon dioxide (CO₂) serves as both a contaminant of concern and a widely used proxy for ventilation adequacy [62]. CO₂ is primarily generated by human respiration, with additional contributions from combustion-based heating systems. In classrooms with insufficient ventilation, concentrations frequently surpass recommended thresholds of 1000 ppm established by international standards such as ASHRAE 62.1 and EN 16798 [49,63]. Prolonged exposure to elevated CO₂ levels has been associated with symptoms including headaches, fatigue, and drowsiness, as well as with measurable decrements in students’ concentration, decision-making, and overall cognitive performance [40,53,64]. Importantly, strategies to reduce CO₂ concentrations through increased ventilation often impose significant energy penalties, particularly in climates requiring substantial heating or cooling, thereby illustrating the persistent trade-off between IAQ management and energy efficiency in educational buildings [56].

Particulate Matter (PM_2.5 and PM₁₀) constitutes a critical pollutant group in school environments, comprising airborne particles with aerodynamic diameters below 2.5 μm and 10 μm, respectively. These particles remain suspended for extended periods and can penetrate deeply into the respiratory tract, where they are associated with adverse cardiovascular and respiratory outcomes [15]. In addition to fine and coarse fractions, ultrafine particles (UFPs, <0.1 μm) are increasingly recognized as a concern due to their ability to translocate into the bloodstream and exert systemic health effects [57,65].

In schools, PM originates from a combination of outdoor sources—notably traffic-related emissions and resuspension of playground dust—and indoor sources such as cleaning activities, combustion appliances, chalk use, and resuspension from floors and furniture. Elevated concentrations of PM_2.5 and PM₁₀ have been consistently linked to increased incidence of asthma symptoms, reduced lung function, and higher absenteeism among children, who are physiologically more vulnerable to inhaled pollutants [57,64].

The mitigation of PM exposure in classrooms typically relies on increased ventilation or filtration efficiency, but both approaches carry significant energy implications. Enhanced ventilation dilutes indoor concentrations but increases heating and cooling demand, while advanced filtration technologies improve IAQ at the expense of higher fan energy use. This duality underscores the necessity of optimizing PM control strategies within an integrated IAQ–energy management framework [66].

Volatile Organic Compounds (VOCs) represent a diverse group of carbon-based chemicals that readily evaporate at room temperature and are frequently detected in school environments [67]. Common sources include cleaning products, paints, adhesives, flooring materials, and furnishings, with formaldehyde—a major constituent of pressed wood products such as desks and cabinets—being one of the most prevalent and well-documented indoor VOCs [68]. Acute exposure to VOCs can cause mucosal irritation, headaches, dizziness, and fatigue, whereas chronic exposure has been associated with more severe health outcomes, including asthma development, nasopharyngeal cancer, and myeloid leukemia [40,64]. Children are especially vulnerable because of their higher inhalation rates relative to body weight and their physiologically immature detoxification systems [69]. The continuous low-level release of VOCs from construction materials and consumer products results in cumulative exposures that pose risks to both health and learning performance. Strategies to mitigate VOC levels typically involve source control (selecting low-emission building materials and furnishings) and ventilation enhancement, yet these measures often entail an energy penalty. Increased ventilation raises heating and cooling demand, while advanced filtration or sorption technologies elevate operational energy use [70]. This duality underscores the importance of integrating material selection, ventilation design, and IAQ monitoring within a broader framework of energy-efficient building operation.

Biological contaminants constitute a major determinant of indoor air quality in schools, where high occupancy density and variable maintenance practices create favorable conditions for microbial growth and transmission. Fungal contamination is particularly common in damp environments, arising when relative humidity exceeds 60% or when water damage compromises walls, ceilings, carpets, or books [71]. Exposure to mold spores has been consistently associated with asthma exacerbation, allergic responses, and respiratory symptoms, with children and individuals with pre-existing conditions being the most vulnerable populations [15,72].

Beyond fungi, bacteria and viruses readily circulate in crowded classrooms through both airborne droplets and contact with contaminated surfaces [73]. Pathogens of concern include Streptococcus pneumoniae, Rhinovirus, influenza viruses, and, more recently, SARS-CoV-2, whose airborne transmission highlighted the central role of ventilation and filtration effectiveness in infection control [34,74]. Inadequate ventilation, poor humidity regulation, and insufficient HVAC maintenance exacerbate microbial accumulation and persistence, whereas interventions such as mechanical ventilation upgrades, high-efficiency filtration, and humidity control have been shown to mitigate transmission risks.

The presence of biological pollutants not only undermines student health but also results in increased absenteeism among pupils and staff, thereby reducing overall learning outcomes and institutional productivity. Importantly, effective mitigation strategies often require higher ventilation and filtration rates, which can substantially increase energy demand. This reinforces the need for integrated IAQ–energy management frameworks that leverage advanced monitoring, predictive modeling, and smart building operation to maintain healthy indoor environments in schools without compromising energy efficiency [53,66].

In addition to CO₂, PM, and VOCs, gaseous pollutants such as nitrogen dioxide (NO₂) and tropospheric ozone (O₃) represent significant concerns in school environments [22,23]. NO₂ originates predominantly from outdoor traffic-related emissions, with additional contributions from unvented gas appliances and combustion-based heating systems indoors. Elevated NO₂ levels have been consistently associated with airway inflammation, asthma exacerbation, and reduced lung function in children [22]. By contrast, O₃ is largely introduced from outdoor air, although it can also be generated indoors by certain electronic devices and cleaning technologies. Exposure to tropospheric ozone has been linked to eye and throat irritation, impaired pulmonary function, and worsening of asthma symptoms [23]. Both pollutants highlight the strong dependence of indoor air quality on ambient outdoor conditions and building ventilation dynamics. Schools situated near major roads or in urban pollution hotspots are especially vulnerable, as pollutant infiltration often coincides with inadequate building envelope performance and insufficient filtration. Moreover, ozone readily reacts with indoor VOCs, forming secondary pollutants such as formaldehyde and ultrafine particles, further compounding health risks [67,75]. Mitigation strategies—including enhanced filtration, demand-controlled ventilation, and selective air intake scheduling—can effectively reduce exposures but frequently increase energy demand, underlining the need for integrated IAQ–energy management solutions [76].

A synthesis of the principal pollutant groups relevant to school environments, along with their dominant sources and associated health and performance outcomes, is presented in Table 1. The table highlights the broad spectrum of contaminants typically encountered in classrooms and shows how school-specific conditions—such as high occupant density, intensive use of materials, and insufficient ventilation—can substantially exacerbate exposures. By systematically linking pollutant categories with their health and cognitive effects, Table 1 offers a structured framework for understanding the mechanisms through which indoor contaminants contribute to both acute symptoms and long-term risks in students and staff. Moreover, this synthesis emphasizes that pollutant management in schools cannot be decoupled from building operation: strategies to reduce exposure often influence energy performance, reinforcing the need for integrated approaches that jointly address IAQ, health, and sustainability objectives.

Table 1. Major pollutant categories relevant to school environments, their typical indoor and outdoor sources, and associated health and performance impacts on students and staff.

Reference	Pollutant	Primary Sources in Schools	Main Health and Performance Impacts
[18,19]	Carbon dioxide (CO2)	Occupant respiration, inadequate ventilation, combustion from heating	Fatigue, drowsiness, impaired concentration, reduced cognitive performance
[59]	PM (PM2.5/PM10)	Chalk dust, resuspension of settled particles, outdoor traffic and exhaust fumes, indoor cleaning activities	Respiratory tract irritation, asthma exacerbation, increased absenteeism
[19,67,77,78]	VOCs	Cleaning products, paints, adhesives, furniture, carpets, wooden materials	Headaches, allergic reactions, mucosal irritation, long-term carcinogenic potential
[79,80]	Fungi (mold)	Elevated humidity, water damage, poor maintenance and cleanliness	Allergies, asthma onset and attacks, respiratory symptoms
[34,73]	Bacteria and viruses	Occupants, contaminated surfaces, airborne droplets	Respiratory infections, influenza, COVID-19, school absenteeism
[22,75]	Tropospheric Ozone (O3)	Outdoor infiltration from ambient air (particularly in urban areas with high photochemical smog), Indoor generation from certain devices, Secondary chemical reactions indoors	Eye and airway irritation, asthma aggravation, reduced attention, absenteeism
[23]	Nitrogen Dioxide (NO2)	Outdoor traffic emissions, Indoor combustion sources, Proximity to parking areas or bus drop-off zones.	Respiratory irritation, asthma exacerbation, reduced lung function, absenteeism

While the presence of indoor pollutants is a central concern, the overall quality of classroom air is equally governed by environmental, operational, and structural determinants that mediate exposure dynamics. Ventilation strategy, thermal and moisture conditions, building envelope performance, emission characteristics of construction materials, occupancy density, and HVAC operation and maintenance interact in complex ways to shape pollutant concentrations and their associated health outcomes [81,82]. Importantly, these same parameters also influence energy demand, underscoring the need to evaluate IAQ within the broader context of sustainable building performance [83]. Systematically addressing these determinants is therefore critical for the design of resilient, energy-efficient, and health-promoting learning environments. Table 2 synthesizes the principal factors affecting IAQ in schools and outlines targeted interventions aimed at mitigating risks while supporting both student well-being and institutional sustainability.

The determinants outlined in Table 2 demonstrate that indoor air quality in schools is shaped not only by the presence of pollutants but also by the operational, environmental, and structural characteristics of the building. Inadequate ventilation remains one of the most critical drivers of elevated CO₂ and particulate concentrations, particularly in densely occupied, naturally ventilated classrooms [15,17,49,57]. Thermal and humidity regulation is equally essential, as deviations from recommended ranges not only compromise thermal comfort but also promote microbial growth and survival, thereby amplifying respiratory health risks [16,84]. Building materials and cleaning practices act as additional emission sources, with furnishings, paints, adhesives, and detergents identified as major contributors of VOCs and allergens [15,16]. Moreover, poor maintenance of HVAC systems diminishes filtration efficiency, encourages microbial proliferation, and facilitates the accumulation of chemical and biological contaminants [49,64]. Taken together, these findings highlight the need for integrated IAQ management strategies that couple technological interventions—such as advanced filtration, humidity control, and demand-controlled ventilation—with behavioral and policy measures, including low-emission material selection, pollutant source reduction, and systematic maintenance protocols. Importantly, because many of these interventions directly affect building energy demand, IAQ management must be embedded within a broader sustainability framework that balances health protection, energy efficiency, and climate objectives.

Table 2. Key environmental and structural determinants of indoor air quality in school buildings, together with their descriptions and recommended intervention measures to sustain healthy learning environments.

Reference	Key Determinant	Description	Recommended Measures
[19,51,52,58,64]	Pollutant load	Particulate matter, volatile organic compounds, allergens (e.g., mold, dust mites), and chemical residues.	Use air purifiers; limit the use of high-emission cleaning products and chemical agents.
[16,85]	Thermal and moisture conditions	Elevated temperature and humidity favor microbial growth, while excessively low temperatures can cause respiratory discomfort	Maintain indoor temperature within comfort ranges; regulate relative humidity between 30–60% using humidifiers/dehumidifiers.
[15,17,35]	Ventilation efficiency	Adequate aeration removes pollutants and contaminants while supplying oxygenated air.	Implement sufficient natural or mechanical ventilation; install and maintain high-efficiency particulate filters.
[15,16]	Pollution sources	Emissions from smoking, cleaning products, building materials, furniture, and appliances degrade IAQ.	Prohibit indoor smoking; select low-emission, eco-certified materials; store cleaning agents safely.
[49,64]	Building operation and maintenance	Proper HVAC design, operation, and cleanliness directly influence IAQ levels.	Conduct regular HVAC inspection and maintenance; clean or replace filters periodically.

Table 2. Key environmental and structural determinants of indoor air quality in school buildings, together with their descriptions and recommended intervention measures to sustain healthy learning environments.

Reference	Key Determinant	Description	Recommended Measures
[19,51,52,58,64]	Pollutant load	Particulate matter, volatile organic compounds, allergens (e.g., mold, dust mites), and chemical residues.	Use air purifiers; limit the use of high-emission cleaning products and chemical agents.
[16,85]	Thermal and moisture conditions	Elevated temperature and humidity favor microbial growth, while excessively low temperatures can cause respiratory discomfort	Maintain indoor temperature within comfort ranges; regulate relative humidity between 30–60% using humidifiers/dehumidifiers.
[15,17,35]	Ventilation efficiency	Adequate aeration removes pollutants and contaminants while supplying oxygenated air.	Implement sufficient natural or mechanical ventilation; install and maintain high-efficiency particulate filters.
[15,16]	Pollution sources	Emissions from smoking, cleaning products, building materials, furniture, and appliances degrade IAQ.	Prohibit indoor smoking; select low-emission, eco-certified materials; store cleaning agents safely.
[49,64]	Building operation and maintenance	Proper HVAC design, operation, and cleanliness directly influence IAQ levels.	Conduct regular HVAC inspection and maintenance; clean or replace filters periodically.

2.2. Impacts of IAQ on Health and Educational Performance

Poor IAQ in schools has been consistently associated with adverse health outcomes and impaired academic performance, with children representing a particularly vulnerable population due to their immature respiratory and immune systems, higher ventilation rates per body weight, and longer daily occupancy indoors [86]. Exposure to pollutants such as PM_2.5, PM₁₀, CO₂, and VOCs increases both the prevalence and severity of asthma, allergies, and other respiratory conditions, often manifested as sneezing, nasal congestion, eye irritation, coughing, and dermatological symptoms [40,87]. Acute exposures, especially elevated CO₂ concentrations in inadequately ventilated classrooms, are frequently associated with fatigue, headaches, dizziness, and discomfort, which undermine students’ physical well-being and directly reduce attention, concentration, and decision-making capacity [15].

Beyond short-term symptoms, poor IAQ has been shown to contribute to increased absenteeism, reduced standardized test scores, and diminished classroom engagement, thereby exerting measurable effects on educational outcomes [66,88]. These findings underline that the consequences of inadequate IAQ extend well beyond health risks, shaping both the learning efficiency of students and the overall productivity of school systems. These health burdens translate directly into educational performance. Students experiencing respiratory or pollutant-related symptoms are more likely to miss school, disrupting learning continuity and long-term academic progress. Even in the absence of absenteeism, poor IAQ exerts measurable effects on cognitive function: high CO₂ levels reduce alertness, concentration, and decision-making accuracy [72], while exposure to particulate matter and VOCs further impairs attention span and task completion [15,40]. The cumulative evidence demonstrates that inadequate IAQ simultaneously compromises children’s health and their capacity to learn, underscoring air quality management as a prerequisite not only for safeguarding well-being but also for sustaining academic performance in educational settings [72]. Figure 1 illustrates the pathways linking IAQ determinants to health outcomes and educational achievement.

Figure 1. Conceptual representation of the main sources and categories of indoor air pollutants in school environments.

2.3. Regulatory Framework and Standards for IAQ in School Environments

Ensuring adequate indoor air quality (IAQ) in schools is widely recognized as a fundamental prerequisite for safeguarding student health, well-being, and academic performance, and is therefore embedded within a range of international and national regulatory frameworks. Although these frameworks differ in scope, specificity, and enforcement, they share the overarching objective of defining acceptable pollutant thresholds and establishing protocols for monitoring and managing air quality in educational environments [69].

In the United States, ASHRAE has established performance-based standards for schools, most notably in ASHRAE Standard 62.1, which prescribes a minimum outdoor airflow rate of 10 L/s per person (equivalent to ~10 L/min per student), alongside criteria for air filtration, humidity regulation, and CO₂ concentration control, in order to ensure both health protection and thermal comfort [85]. Complementing these technical standards, the U.S. Environmental Protection Agency (EPA) developed the Indoor Air Quality Tools for Schools program, which provides a structured framework for pollutant monitoring, ventilation management, and stakeholder engagement, thereby facilitating the translation of regulatory guidance into operational practice within educational facilities [64].

At the European level, EN 16798-1 [63] specifies ventilation requirements for non-residential buildings, including classrooms, with reference to both per-person airflow rates and indoor CO₂ thresholds relative to outdoor concentrations, while the World Health Organization (WHO, 2010) [19] has issued guideline values for key pollutants such as formaldehyde, benzene, NO₂, and PM_2.5. These frameworks highlight the dual challenge of achieving adequate IAQ while controlling the energy implications of ventilation and filtration, a balance that remains particularly difficult in aging school infrastructures with limited retrofitting capacity.

At the international level, the WHO has issued air quality guidelines that are widely referenced in the management of IAQ in schools, with recommended thresholds of 10 µg/m³ for PM_2.5 and 20 µg/m³ for PM₁₀ (annual mean values) [21]. These guidelines underscore the heightened susceptibility of children to air pollution, linking exposure to fine particulate matter with increased respiratory morbidity and long-term health risks. In parallel, performance-based ventilation standards such as those set by ASHRAE and the European Standard EN 16798-1 specify a maximum indoor CO₂ concentration of 1000 ppm, which is commonly adopted as a benchmark for adequate classroom ventilation [85].

Within the European Union, regulatory attention to IAQ has expanded through both legislative and technical instruments. The revised Energy Performance of Buildings Directive (Directive 2018/844/EU [89]) explicitly incorporates IAQ as a requirement for healthy indoor environments, with Article 13 encouraging the monitoring of key pollutants in high-occupancy spaces such as classrooms. Complementing this, the European Standard EN 16798 specifies performance-based ventilation requirements, including a minimum outdoor airflow of 7 L/s per person and an indoor CO₂ concentration not exceeding 1000 ppm [63]. By coupling pollutant control with energy efficiency objectives, these measures reflect a growing policy emphasis on integrated approaches that safeguard occupant health while supporting sustainability targets [19,89].

Despite the establishment of regulatory frameworks, substantial challenges and limitations remain in practice. Compliance is often hampered by financial constraints, outdated infrastructure, and inconsistent enforcement mechanisms, particularly in older or underfunded schools where resources for retrofitting are limited. Furthermore, most current regulations adopt a static, prescriptive approach that does not adequately reflect the complexity of pollutant dynamics or the variability of classroom occupancy and use. Crucially, existing standards rarely integrate emerging technologies such as AI-enabled real-time monitoring, predictive modeling, and adaptive ventilation control, which offer considerable potential for achieving dynamic, cost-effective, and energy-efficient IAQ management [13]. Addressing these gaps will require not only more rigorous implementation of existing requirements but also the systematic incorporation of smart, data-driven strategies into regulatory and operational practice, thereby aligning IAQ management with broader objectives of health protection, energy efficiency, and long-term sustainability in school environments.

2.4. IAQ in Educational Environments: Challenges, Innovations, and Policy Prospects

Despite decades of research, achieving adequate IAQ in educational environments remains a persistent and systemic challenge. Recurrent issues include non-standardized ventilation practices, frequent exceedances of pollutant thresholds, and limited policy prioritization, particularly in older or under-resourced educational facilities [34,57,86]. These challenges are compounded by the absence of harmonized, child-specific exposure limits and the lack of comprehensive long-term monitoring frameworks, which together constrain efforts to establish robust, evidence-based links between IAQ conditions, health outcomes, and educational performance.

At the same time, technological and analytical innovations provide promising avenues for improvement. Advances in low-cost sensor networks and real-time monitoring platforms are increasingly being deployed to track pollutant concentrations in classrooms, with growing efforts to integrate IAQ metrics with student health and cognitive performance indicators [90]. In parallel, statistical and machine learning models have quantified the effects of CO₂, PM, and other pollutants on outcomes such as attention, fatigue, and academic productivity, reinforcing the central role of IAQ in promoting both health and educational equity [91]. However, these initiatives remain fragmented and largely experimental, with limited validation across diverse climatic and socioeconomic contexts and insufficient incorporation into binding regulatory and design frameworks. Without systematic integration into school building standards and operational protocols, the potential of these innovations to deliver sustainable, scalable, and equitable improvements in IAQ will remain underutilized.

Taken together, the current body of evidence underscores the need for a coordinated, technology-enabled approach to IAQ management in schools. This requires not only the deployment of smart ventilation systems, sensor-based monitoring platforms, and predictive control strategies, but also the systematic alignment of IAQ objectives with broader health, education, and sustainability policies at both national and supranational governance levels. To support this integration, Table 3 synthesizes recent insights from the literature, structuring them into focus areas, innovation highlights, critical gaps, and future prospects, thereby providing a strategic roadmap for research, policy development, and practical implementation in school environments.

Table 3. Strategic overview of IAQ research and practice in school environments, highlighting focus areas, recent innovations, identified gaps, and future prospects for implementation.

Reference	Focus Area	Innovation Highlight	Identified Gaps	Future Prospects
[24]	Natural ventilation and CO2 management	Classroom-level CO2 thresholds applied in naturally ventilated schools	Lack of enforcement and standardization for natural ventilation practices	Integration of smart alerts and continuous CO2 feedback systems
[13,51,86]	Health impacts of pollutants in schools	Evidence linking IAQ to pediatric respiratory and allergic outcomes	Absence of child-specific IAQ exposure thresholds	Development of health-integrated IAQ criteria in building and education policies
[87]	Ventilation strategies and monitoring	Deployment of low-cost sensor-based ventilation control	Lack of harmonized IAQ monitoring protocols across schools	Establishment of standardized sensor networks for large-scale monitoring
[69]	Indoor environment and learning outcomes	Integration of IAQ metrics with cognitive and academic performance indicators	Limited availability of long-term outcome data	Longitudinal studies linking IAQ to learning achievements and curriculum design
[64,92,93,94]	Sensor technologies for IAQ	Advances in sensor calibration for deployment in schools	Absence of unified protocols for sensor placement and validation	Creation of open-access IAQ dashboards and data-sharing frameworks
[74]	Public health implications in EU schools	Regional mapping of IAQ inequalities	Low policy prioritization in disadvantaged regions	EU-level mandates and funding schemes to reduce IAQ disparities
[87]	CO2 and cognition in naturally ventilated schools	Statistical modeling of CO2 effects on student performance	Over-reliance on CO2 as the sole IAQ indicator	Hybrid ventilation strategies integrating multi-pollutant assessment
[68,87]	Ventilation and cognitive performance	Quantified effects of IAQ on brain function and task performance	Limited field validation across diverse climatic and socio-economic contexts	Development of neurodevelopmental IAQ indices for schools
[69]	Productivity and IAQ	Meta-analyses linking CO2 thresholds with student productivity	Insufficient attention to equity-related outcomes	Incorporation of IAQ metrics into indicators of educational access and equality
[64]	Governance and policy frameworks	EPA’s IAQ Tools for Schools provides an operational framework	Voluntary implementation with no binding legal effect	Introduction of mandatory IAQ audits supported by federal or state funding
[57,91]	IAQ–energy trade-offs and smart management	AI-driven demand-controlled ventilation and predictive IAQ–energy modeling	Limited integration of IAQ and energy metrics in existing standards	Systemic frameworks combining IAQ monitoring, adaptive ventilation, and energy efficiency for sustainable school operation

The synthesis presented in Table 3 shows that, although research on IAQ in educational environments has expanded substantially, progress remains fragmented and uneven across thematic domains. Natural ventilation and CO₂ monitoring are among the most frequently studied strategies; however, their effectiveness is limited by the lack of enforceable performance standards and the persistent over-reliance on CO₂ as a proxy for IAQ, despite its inability to capture chemical and biological exposures [24,87]. Efforts to establish links between pollutant exposure, child-specific health outcomes, and cognitive performance show considerable promise but are hindered by the absence of harmonized thresholds and the limited availability of longitudinal epidemiological datasets [13,69]. Similarly, while sensor-based monitoring, statistical modeling, and data-driven ventilation strategies offer transformative potential, their impact is curtailed by the lack of standardized calibration protocols, interoperability frameworks, and coordinated large-scale deployment, which restrict both comparability across studies and scalability in practice [64,87].

A recurrent theme is the imbalance between technological innovation and policy adoption. Programs such as the EPA’s IAQ Tools for Schools [64] provide structured governance frameworks, yet their voluntary nature and lack of enforcement limit their systemic impact. Within Europe, uneven policy prioritization and resource allocation continue to perpetuate disparities in IAQ, with disadvantaged schools disproportionately affected [74]. This reinforces the need for EU-level mandates, harmonized child-specific exposure standards, and dedicated funding mechanisms to ensure equitable protection.

Crucially, IAQ in schools cannot be considered in isolation from energy performance objectives. Strategies such as increased ventilation and advanced filtration improve pollutant control but often elevate heating, cooling, and fan energy demand. The absence of integrated frameworks to reconcile this trade-off highlights a major research and policy gap. The new row in Table 3 emphasizes that AI-driven demand-controlled ventilation and predictive IAQ–energy modeling represent a promising pathway to address this dual challenge. Embedding such approaches into building operation protocols and regulatory standards will be essential to achieve healthy, energy-efficient, and sustainable school environments.

Overall, Table 3 underscores the urgency of bridging the gap between scientific evidence, technological innovation, and regulatory enforcement. A systemic approach that combines smart monitoring technologies, predictive and adaptive ventilation control, child-focused exposure guidelines, and binding governance frameworks, while simultaneously integrating energy efficiency considerations, will be essential to translate current knowledge into effective, scalable, and sustainable practice in schools.

3. Artificial Intelligence Approaches for IAQ Assessment in Educational Environments

Artificial intelligence (AI) is increasingly regarded as a transformative framework for indoor air quality (IAQ) assessment in educational environments, where exposure is closely tied to health outcomes and learning performance. Unlike conventional statistical or rule-based methods, AI can integrate and analyze heterogeneous data sources—including pollutant concentrations, meteorological drivers, ventilation rates, and dynamic occupancy profiles—to capture the nonlinear interactions that govern IAQ variability [95,96,97]. For example, recent reviews of neural network and machine learning models in school settings highlight their superiority over linear approaches in capturing CO₂ variation under fluctuating occupancy and ventilation schedules [39]. Likewise, García-Pinilla et al. [26] demonstrated that ML-based models outperform simple methods for longer-term CO₂ forecasting in school classrooms. Machine learning (ML) and Deep Learning (DL) approaches, in particular, have been shown to enhance the accuracy of pollutant forecasting, enable anomaly detection in sensor networks (e.g., LSTM-autoencoder models achieving >99% accuracy in school IAQ time series [98]), and support adaptive control strategies for HVAC systems, including DL-driven fault detection and diagnostics with F-measure values exceeding 0.97 [5,6,99,100,101]. Such capabilities are especially relevant in educational buildings, where ventilation demand often fluctuates rapidly and where traditional steady-state models fail to capture transient exposure conditions. Nevertheless, the application of AI to IAQ in schools remains constrained by challenges such as limited availability of long-term, high-resolution datasets, potential overfitting of models trained on small or site-specific samples, lack of model generalization across different climatic and building contexts [39], and difficulties in ensuring model interpretability for practical building management [96]. Addressing these limitations is critical if AI is to evolve from a predictive tool toward a reliable decision-support system for sustainable and health-oriented educational environments. The hierarchical structure of AI, ML, and DL, and their respective roles in IAQ modeling, is illustrated in Figure 2.

Figure 2. Integration of ML and DL models within AI frameworks.

3.1. Machine Learning Methods

Unlike traditional statistical approaches, which often assume linear relationships, ML can capture nonlinear dependencies among diverse environmental and operational variables, providing more reliable predictions of pollutant behavior [102,103]. In general, ML methods can be grouped into four main categories according to how they learn from data [5]: (a)supervised learning, which relies on labeled datasets to establish explicit input–output mappings and is commonly applied to tasks such as pollutant classification or concentration forecasting [104], (b) unsupervised learning, which, by contrast, works with unlabeled data to identify latent structures or clusters—for instance, grouping classrooms by similar pollution profiles [105,106], (c) semi-supervised learning which bridges the two by leveraging a small set of labeled data together with a much larger body of unlabeled observations [107], and finally, (d) reinforcement learning, which uses iterative interaction between an agent and its environment to optimize long-term outcomes [108].

Among supervised approaches, Support Vector Machines (SVMs), Decision Trees (DTs), k-nearest neighbors (k-NNs), and Artificial Neural Networks (ANNs) are the most widely applied for short-term pollutant forecasting, exposure classification, and anomaly detection [109,110,111,112,113,114].

Support Vector Machines (SVMs) have demonstrated strong performance in classifying classroom air quality conditions—such as “good,” “moderate,” or “poor”—using input features including CO₂ concentration, particulate matter levels, and occupancy-related variables [113]. The method constructs an optimal separating hyperplane between classes by maximizing the margin, formulated as [105]:

minimize (1/2)‖w‖² subject to y_i (w·x_i + b) ≥ 1 ∀i

(1)

where x_i is the feature vector, y_i is the class label, w is weight vector defining the orientation of the separating hyperplane, b is bias term, and the constrain y_i (w·x_i + b) ≥ 1 ensures correct classification of all training samples with maximum margin.

DTs are valued for their interpretability, as they explicitly identify the dominant drivers of pollutant exceedances, such as occupancy density or inadequate ventilation [110,111,112]. DTs add value through interpretability, as they identify dominant drivers of exceedances (e.g., occupancy density, ventilation regime), making them particularly suitable for building management applications that demand transparency [110,112].

At each node, the algorithm selects the variable and threshold that minimize an impurity measure, most commonly the Gini index [105,111]:

G = 1 − Σ (p_k²)

(2)

where p_k is the proportion of samples belonging to class k in a given node, k is the number of classes, and G is the impurity measure (0 = perfectly pure node, higher values = more mixed node).

k-NNs is a non-parametric algorithm that classifies or predicts outcomes by comparing a new observation with the k most similar instances in the training dataset. Similarity is typically quantified using a distance metric, most commonly the Euclidean distance [115]:

$$\mathrm{d}\left({\mathrm{x}}_{\mathrm{i}},{\mathrm{x}}_{\mathrm{j}}\right)=\mathrm{s}\mathrm{q}\mathrm{r}\mathrm{t}\left(\mathrm{s}\mathrm{u}{\mathrm{m}}_{\mathrm{m}=1}^{\mathrm{M}}{\left({\mathrm{x}}_{\mathrm{i},\mathrm{m}}-{\mathrm{x}}_{\mathrm{j},\mathrm{m}}\right)}^2\right)$$

(3)

where $x_i,x_j$ are feature vectors, M is the number of features, $\mathrm{d}\left(x_i,x_j\right)$ is the Euclidean distance between two observations i and j, and k is the number of nearest neighbors used to classify or predict.

In classroom applications, k-NN supports real-time anomaly detection by identifying deviations from previously observed sensor patterns, which allows for timely corrective actions—such as adjusting ventilation rates—before pollutant levels exceed health-related thresholds [1,5,6,101,109,114].

Artificial Neural Networks (ANNs), inspired by the structure of biological neurons, have been increasingly employed for IAQ prediction because of their ability to approximate nonlinear relationships between multiple input variables (e.g., occupancy, temperature, ventilation rates, outdoor meteorological conditions) and output responses (e.g., pollutant concentrations, IAQ categories). A neuron in a feed-forward ANN computes its output as [104,105]:

$$\mathrm{y}=\mathrm{f}\left(\mathrm{s}\mathrm{u}{\mathrm{m}}_{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{w}}_{\mathrm{i}}{\mathrm{x}}_{\mathrm{i}}+\mathrm{b}\right)$$

(4)

where $x_i$ are the input features, $w_i$ are the connection weights, b the bias term, $f$ the activation functions, and $y$ the predicted or estimated output.

Feed-forward ANNs trained with backpropagation are frequently applied to tasks such as pollutant forecasting and short-term IAQ classification [26,39]. These models have, for instance, been used to predict CO₂ variation in classrooms with fluctuating occupancy schedules.

Beyond classification tasks, regression-oriented ML techniques are increasingly employed to model pollutant dynamics and to examine their associations with both indoor and outdoor determinants. These methods are particularly relevant in naturally ventilated schools, where CO₂, particulate matter, and volatile organic compounds often display pronounced temporal variability shaped by occupancy density, building envelope characteristics, and local meteorological conditions [27,102]. Recent studies have further strengthened this research direction by linking ML-based pollutant forecasts with indicators of student health and cognitive performance, suggesting that accurate prediction can enable timely interventions aimed at reducing absenteeism and improving learning outcomes [87,103].

Nevertheless, several challenges continue to limit the scalability and robustness of ML applications in IAQ management. A primary constraint is the scarcity of large, high-quality training datasets, as most school-based investigations are based on short-term monitoring or restricted sample sizes, which undermines model generalizability [40,102,116]. Data uncertainties introduced by sensor calibration issues and variable measurement quality further increase the risk of systematic bias. In addition, models trained in specific climatic zones or building typologies often perform poorly when transferred to different contexts, highlighting the fragility of current approaches. Addressing these shortcomings will require coordinated initiatives to establish harmonized monitoring protocols, publicly available benchmark datasets, and rigorous validation frameworks that can guarantee reproducibility and transferability across diverse educational environments.

3.2. Deep Learning Approaches

DL constitutes a major methodological advancement in IAQ research, particularly suited to the high-dimensional datasets produced by continuous sensor networks and environmental monitoring platforms [28,117]. In contrast to conventional machine learning, which often depends on manual feature engineering, DL architectures are capable of learning hierarchical feature representations directly from raw data, thereby uncovering hidden patterns and nonlinear dependencies that traditional methods frequently fail to capture [116,118]. This feature is particularly relevant in school environments, where pollutant concentrations are shaped by rapidly changing occupancy levels, intermittent ventilation, and variable outdoor infiltration.

Among DL techniques, Convolutional Neural Networks (CNNs) have been increasingly employed in IAQ studies for their capacity to extract spatial and temporal features from multivariate time-series data. The operation of a convolutional layer can be expressed as [5,119]:

$$F_{i,j}={\sum }_m{\sum }_n\left(I_{i+m,j+n}·K_{m,n}\right)+b$$

(5)

where $F_{i,j}$ is the output at spatial location $\left(i,j\right)$, $I$ is the input feature map, $K$ is the convolutional kernel (or filter), $I_{i+m,j+n}$ is the local receptive field of the input over which the kernel is applied, $b$ is the bias term added after convolution, and ${\mathsf{\Sigma }}_m{\mathsf{\Sigma }}_n$ is the summation across the kernel dimensions.

By processing IAQ sensor streams, CNN-based models have achieved high predictive accuracy in forecasting CO₂ and PM_2.5–PM₁₀ levels, while also identifying pollution hotspots linked to overcrowding, dust resuspension, or insufficient ventilation [39,120,121,122,123].

Recurrent Neural Networks (RNNs) represent another family of DL models, particularly effective for sequential data. Their recursive architecture enables the modeling of temporal dependencies, making them highly suitable for pollutant forecasting where daily and weekly cycles dominate IAQ dynamics. The hidden state update of a standard RNN is defined as [5,124]:

$$h_t=f\left(W_h{h}_{t-1}+W_x{x}_t+b_h\right)$$

(6)

where $h_t$ is the hidden state at time t, $x_t$ is the input, $W_h$ and $W_x$ are weight matrices, b is the bias, and f(⋅) is the activation function. Long Short-Term Memory (LSTM) networks extend this formulation by introducing memory cells and gating mechanisms that allow the retention of long-range dependencies, overcoming the vanishing gradient problem typical of conventional RNNs [125]. In school environments, LSTM-based models have been applied to forecast pollutant accumulation and dispersion cycles, supporting anticipatory ventilation strategies that minimize exposure during critical hours of the day [114,126,127].

LSTM networks extend conventional RNNs by introducing a dedicated cell state that preserves information across time steps. This state is regulated by three gates—input, forget, and output—which selectively update, discard, or propagate information, thereby enabling the network to retain relevant temporal dependencies while discarding redundant patterns. Such a structure effectively mitigates the vanishing and exploding gradient problems commonly observed in standard RNN training. Gated Recurrent Units (GRUs) adopt a similar gating mechanism but use a more compact architecture [128]. Specifically, GRUs merge the input and forget gates into a single update gate while retaining a reset gate, thus reducing the number of trainable parameters and computational overhead. Despite their simpler structure, GRUs have demonstrated comparable performance to LSTMs in time-series forecasting tasks, making them particularly attractive for IAQ prediction in resource-constrained environments such as school monitoring systems [27,129].

Although DL methods consistently outperform classical ML models in terms of predictive accuracy, robustness to noisy inputs, and capacity to integrate heterogeneous environmental, meteorological, and occupancy data [28,116,117], their adoption in educational environments remains constrained. Persistent barriers include the scarcity of long-term, high-quality IAQ datasets, the significant computational resources required for training and operation, and the limited interpretability of model outputs, which restricts their utility for practical building management [114,116]. Addressing these challenges will require the creation of open-access benchmark datasets tailored to school environments, the design of computationally efficient DL models suitable for real-time operation in resource-limited settings, and the integration of explainable AI (XAI) approaches capable of translating complex outputs into actionable insights for educators, facility managers, and policymakers.

Table 4 summarizes key advantages and drawbacks of major AI methods—including Support Vector Machine (SVM), Decision Tree (DT), Artificial Neural Network (ANN), Recurrent Neural Network (RNN), Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU).

Table 4. Strengths and limitations of common artificial intelligence (AI) algorithms applied to indoor air quality (IAQ) prediction and management in educational environments.

AI Algorithm/Reference	Strengths	Limitations	Main Results
svm/ [113]	High accuracy with small or medium-sized datasets; effective in high-dimensional spaces; robust to overfitting when kernel functions are well-tuned.	Performance deteriorates with large datasets; computationally expensive in multi-class problems; requires careful kernel and parameter selection.	High detection precision for known samples; effective detection of unknown anomalies; weighted SVM improved classification on imbalanced datasets; one-class SVM effective for unsupervised data.
dt/ [110,111,112]	Simple and interpretable model; fast training and low computational cost; suitable for classification and feature importance analysis.	Prone to overfitting with noisy data; low generalization ability; limited in modeling complex nonlinear relationships.	- Accurate IAQ and thermal comfort classification with over 90% prediction accuracy [99] - Strong predictive reliability for nonlinear datasets [100] - Effective classification accuracy in educational data applications [101]
knn/ [115]	Simple and easy to implement; non-parametric and adaptable to nonlinear relationships; performs well with small datasets; useful as a benchmark for regression and classification of IAQ data.	Computationally expensive with large datasets; performance degrades in high-dimensional spaces; sensitive to noise and to the choice of distance metric and k value.	KNN demonstrated high classification accuracy and asymptotic error rate approaching twice the Bayes error
ANN/ [5,104,105]	Captures complex nonlinear dependencies among variables; adaptable to dynamic and multivariate datasets; performs well in pollutant prediction and control optimization.	Requires large, well-curated datasets; training can be computationally expensive; may act as a “black box” with limited interpretability.	Theoretical foundations confirm ANN’s universal function approximation capability and strong generalization performance for nonlinear multivariate datasets
RNN [5,124]	Models temporal dependencies and sequential data; effective for time-series prediction such as CO2 and PM fluctuations.	Suffers from vanishing/exploding gradient problems; limited ability to capture long-term dependencies without enhancements.	Theoretically, RNNs enable dynamic sequence learning by propagating temporal context through recurrent connections, allowing representation of time-dependent patterns. However, their training stability is constrained by gradient vanishing in long sequences
CNN [5,119]	Excels at extracting spatial features and pattern recognition; suitable for processing sensor arrays, images, or spectroscopic data.	Requires high computational power; needs large training datasets; less effective for temporal or sequential data.	CNNs, originally developed for visual recognition tasks, have evolved into a foundational architecture for environmental sensing and IAQ analysis. Their hierarchical feature extraction capability enables the identification of complex spatial patterns within sensor or imaging data, highlighting their potential for advanced pollutant mapping and smart-building applications.
LSTM [5,117,126]	Overcomes vanishing gradient issue; excellent for long-term time-series forecasting; captures complex temporal dynamics in pollutant concentrations.	Computationally demanding; requires fine-tuning of hyperparameters; slower training compared to simpler models.	LSTM networks address the vanishing gradient limitations of traditional RNNs by incorporating gated mechanisms that regulate information flow across time steps. Their capacity to capture long-term temporal dependencies makes them highly effective for forecasting pollutant dynamics and indoor air quality variations, particularly in environments characterized by complex temporal fluctuations and delayed system responses.
GRU [5,128]	Simplified variant of LSTM with fewer parameters; faster training and lower computational cost; performs comparably on many time-series tasks.	May be less accurate than LSTM in modeling very long sequences; reduced interpretability compared to simpler models.	GRUs offer a streamlined alternative to LSTMs by merging the forget and input gates into a single update mechanism, thereby reducing computational complexity while maintaining comparable performance. Their efficiency and faster convergence make GRUs particularly suitable for real-time IAQ monitoring and control applications, where predictive accuracy must be balanced with processing speed and energy constraints.

3.3. Hybrid AI Models

Hybrid AI frameworks are increasingly recognized as effective solutions for IAQ monitoring in educational buildings, as they combine the complementary strengths of ML and DL to enhance robustness, generalizability, and predictive accuracy. By integrating classical algorithms such as SVMs and DTs with advanced architectures including CNNs and RNNs, hybrid approaches are able to process heterogeneous data streams comprising CO₂, PM_2.5, PM₁₀, VOCs, and bioaerosols [103,130,131]. This integration enables simultaneous tasks such as anomaly detection, pollutant forecasting, and adaptive control of ventilation or air purification systems [40,96,132,133], thereby linking predictive analytics with automated decision-making in real time.

Several hybrid strategies have been reported in the literature. One configuration integrates SVMs with CNNs or RNNs, exploiting the discriminative capacity of SVMs for feature separation and anomaly detection while CNNs and RNNs capture spatial and temporal dependencies within IAQ data. This architecture has demonstrated effectiveness for pollutant classification and short-term forecasting in school classrooms [95,134]. A second strategy couples DTs with deep neural networks (DNNs), leveraging the interpretability of DTs to identify critical pollutant thresholds while employing DNNs to model complex nonlinear relationships between environmental drivers and indoor concentrations [135,136]. A third category involves CNN–RNN hybrids, where CNNs extract local features from sensor streams and RNNs (particularly LSTM networks) model temporal dynamics. This dual-stage design has been shown to improve forecasting accuracy in high-density classrooms where pollutant fluctuations are driven by rapid occupancy changes and variable ventilation [39,114,126].

The synthesis illustrated in Figure 3 and Table 5 confirms that hybrid AI approaches address several limitations of stand-alone ML or DL methods. Reliability is strengthened through ensemble mechanisms that reduce bias and variance across heterogeneous classroom conditions [103,132]. Effectiveness is further enhanced when classical ML techniques are applied for feature preprocessing or dimensionality reduction, thereby reducing the risk of overfitting and alleviating the intensive data requirements of DL [28,103]. Computational efficiency is also improved: lightweight ML algorithms can perform rapid preprocessing, while deeper architectures handle more complex feature extraction, enabling real-time responsiveness where decision latency directly affects student exposure [39,137].

Figure 3. Conceptual framework of hybrid AI models integrating machine learning (ML) and deep learning (DL).

Table 5. Advantages and implementation strategies of hybrid AI models for IAQ monitoring in schools.

References	Description	Advantages	Implementation Strategy
[103,132]	Integration of traditional ML with DL models enhances robustness across varying classroom conditions.	Reliability	Ensemble outputs from multiple algorithms to reduce bias and improve stability.
[28,103]	Classical ML methods reduce the data demands of deep models.	Effectiveness	Preprocessing and feature reduction with ML before DL training.
[39,137]	Combines lightweight ML with high-precision DL.	Analysis Speed	Parallel use of fast ML classifiers with DL networks for real-time operation.
[126,138]	Models adapt to shifts in occupancy and environment.	Adaptability	Online or incremental learning for continuous updating.
[109,116]	Statistical preprocessing mitigates unreliable sensor signals.	Noise Reduction	Data cleaning and smoothing to handle noisy or incomplete datasets.
[95,114]	Detects hazardous IAQ deviations beyond normal ranges.	Anomaly Detection	Integration of unsupervised learning (e.g., autoencoders) for early detection of outliers.
[39,87]	Forecasts improve continuously as new data arrives.	Real-Time Improvement	Implementation of reinforcement learning and incremental parameter updates.

Adaptability represents another critical advantage. Online and incremental learning mechanisms allow hybrid systems to maintain predictive accuracy under shifting environmental or occupancy regimes [126,138]. Hybrid models also exhibit resilience to noisy or incomplete sensor data by incorporating statistical preprocessing and denoising techniques [109,116]. Moreover, the integration of unsupervised components such as autoencoders facilitates early anomaly detection [95,114], while reinforcement learning modules enable continuous refinement of predictive policies in response to new data [39,87]. Collectively, these attributes position hybrid AI as a promising pathway toward adaptive and automated IAQ management in schools.

Nonetheless, large-scale deployment remains constrained by critical barriers. The absence of standardized IAQ datasets hinders model benchmarking and generalization across diverse school contexts. The interpretability of hybrid models also remains limited, raising concerns about trust and practical uptake in building operations.

Finally, integration into existing school infrastructures requires not only technical advances but also policy incentives and resource allocation. Addressing these challenges through explainable AI frameworks [96], the development of open-access IAQ datasets [40] and incentive-driven implementation strategies [133] will be essential to transition hybrid AI methods from experimental validation to sustainable deployment in educational environments.

ML, DL, and hybrid AI approaches each provide unique contributions to IAQ management in educational environments. ML offers interpretability and modest data requirements, DL captures spatiotemporal dynamics with superior accuracy, and finally, hybrid systems integrate these strengths to achieve robustness and adaptability. Selecting the appropriate method depends on data availability, computational resources, and the balance between accuracy and interpretability required for decision support. Together, these approaches represent a pathway toward intelligent, adaptive, and health-oriented IAQ management in schools.

4. AI Applications for Indoor Air Quality in Educational Environments

AI is increasingly being deployed to improve IAQ in educational environments, where children’s health, comfort, and cognitive performance are especially vulnerable to pollutant exposure. Unlike traditional approaches that rely on periodic inspections, subjective perception, or static ventilation schedules, AI-based systems provide continuous monitoring, predictive forecasting, and adaptive control of indoor environments. To capture the current state of research and implementation, twenty representative case studies were reviewed, covering applications in North America, Europe, Asia, and Oceania. These are summarized in Table 6, which consolidates information on methodological approaches, monitored parameters, deployment scale, and key outcomes. Figure 4 provides a world map overview of the twenty case study locations, highlighting their geographic distribution across North America, Europe, Asia, and Oceania. The case studies span a wide range of applications, from large-scale sensor networks to small pilot projects and privacy-preserving smart classroom frameworks. Collectively, these cases illustrate both the potential of AI to strengthen IAQ management in educational environments and the persistent barriers—such as data scarcity, model transferability, and long-term operational sustainability—that constrain widespread adoption.

Figure 4. Geographic distribution of the AI-based IAQ case studies in educational environments, Spanning North America, Europe, Asia, and Oceania.

Table 6 provides a quantitative synthesis of twenty representative case studies applying AI techniques for IAQ management in educational environments. In addition to qualitative insights, the table compiles numerical indicators such as R², RMSE, MAE, prediction accuracy, and error reduction percentages, allowing direct comparison of algorithmic performance across different pollutants, climatic contexts, and scales of implementation. This quantitative perspective complements the descriptive review by identifying performance ranges, methodological tendencies, and scalability constraints that characterize the current state of AI-enabled IAQ management. The case studies reviewed in Table 6 collectively demonstrate both the opportunities and limitations of applying AI to IAQ management in schools.

A first insight concerns scalability. Large deployments such as Boston, with more than 3600 sensors across 4400 classrooms [139,140], and the German network spanning 329 classrooms [141], confirm the technical feasibility of AI-driven monitoring at scale. These systems achieved measurable reductions in CO₂ concentrations and enabled real-time fault detection, yet they also revealed structural barriers: high installation and maintenance costs, dependence on robust digital infrastructure, and unequal adoption capacity in lower-resource schools. While scalability is therefore achievable, its equitable application remains uncertain.

Methodological innovation has been another defining feature across studies. Hybrid deep learning frameworks, such as the LSTM–autoencoder in Dunedin [98], achieved anomaly detection accuracy above 99%, outperforming classical models such as k-NN and fuzzy clustering.

Other approaches, including RF–TPE–LSTM in Central China [29], SVR with feature engineering in Athens [30,64], temporal convolutional networks in Navarra [26], and BO–EMD–LSTM in North China [31], further advanced predictive performance, often achieving R² values close to 0.9 for CO₂ or PM_2.5 forecasting. Comparative analyses in Codsall, UK [27] and Riga [129] highlighted the growing role of GRU-based architectures, which combined predictive accuracy with lower computational costs, reinforcing their value for energy-efficient HVAC control. Taken together, these methodological advances confirm the capacity of AI models to capture pollutant dynamics and anticipate exposure peaks, enabling proactive ventilation control. At the same time, their dependence on site-specific data raises concerns about generalizability and interpretability, which remain unresolved.

A further dimension concerns the discrepancy between subjective perception and objective measurement of IAQ. Studies in Portugal [90], the UK [32], and Finland [142] revealed systematic underestimation of pollutant exceedances by teachers and staff, even in classrooms where CO₂ and particulate matter levels regularly surpassed recommended thresholds. Feedback systems based on IoT devices reduced average CO₂ concentrations by nearly 20% [32], but behavioral constraints, such as reduced ventilation during cold weather, limited their effectiveness. These findings highlight the inadequacy of perception-driven management and underline the value of AI-based transparency in guiding both behavioral adjustments and institutional decision-making.

Beyond prediction, AI is increasingly being embedded in HVAC optimization strategies. In Seoul [143], integrated neural networks coupled with heuristic multi-objective optimization achieved up to 16% energy savings while maintaining IAQ, while in Hong Kong [144], real-time occupancy detection combined with CFD and fuzzy logic enabled dynamic balancing of thermal comfort and air quality. Similarly, Bayesian gray-box models in Montreal [145] leveraged continuous CO₂ data to infer ventilation rates and guide targeted interventions. These applications illustrate how AI can align health protection with energy performance, though their computational demands and system integration requirements may limit broader adoption in the near term.

Occupancy detection and smart campus platforms provide an additional pathway for enhancing IAQ management. The Smart UA platform in Alicante [146] applied ANN-based ventilation quality certificates with almost 98% accuracy, while MLP models in Pombal [147] predicted occupancy patterns with R² = 0.96, enabling more effective control strategies. Work in Florida [148] demonstrated how PCA–ANN models could link pollutant infiltration to envelope condition and proximity to traffic sources, showing the potential of AI to inform broader building management decisions.

Finally, issues of ethics, privacy, and contextual adaptation remain critical. Edge-based, privacy-preserving frameworks such as SITA [149] demonstrate that accurate IAQ management is possible without compromising data security, while studies in Beijing [150] highlight the necessity of context-specific strategies—for example, portable filters and controlled ventilation outperformed generic interventions under severe outdoor pollution. These cases underscore that AI solutions in educational buildings cannot be universally standardized but must be adapted to local climatic, infrastructural, and socio-economic realities.

In synthesis, AI applications in schools reveal a clear trajectory: from large-scale monitoring to sophisticated predictive modeling, integration with smart HVAC, and embedding within broader smart campus platforms. Across these contexts, AI consistently enhances predictive accuracy, anomaly detection, and adaptive control compared with traditional approaches. Yet systemic challenges persist, including data scarcity, calibration and reliability issues, weak transferability across settings, high implementation costs, limited interoperability with legacy systems, and enduring concerns over privacy and interpretability. Unless these barriers are addressed through standardized open datasets, explainable AI models, cost-effective integration strategies, and supportive governance frameworks, AI risks remaining confined to isolated pilots rather than scaling into mainstream IAQ management in educational environments.

Table 6. Real-world case studies on AI applications for IAQ management in schools, summarizing methods, monitored parameters, scale, and main outcomes.

Reference	Location/Year	AI Method	Parameters Monitored	Sample Size	Main Results/Critical Insights
[139,140]	Boston, USA (2023)	ML (Decision Trees)	CO2, PM2.5, PM10, CO, T, RH	4400 classrooms, 3659 sensors	(a) Demonstrated feasibility of large-scale IAQ monitoring (b) Reduced CO2 by 25–30% (c) Enabled real-time fault detection and improved health indicators (d) Highlighted value of teacher engagement in decision-making
[98]	Dunedin, New Zealand (2022)	Hybrid DL (LSTM + Autoencoder)	CO2	74 sensors, 247 k readings	(a) Achieved 99.5% anomaly detection accuracy (b) Outperformed k-NN and fuzzy clustering (c) Proved capacity of hybrid DL models to generalize patterns in IAQ data.
[30,64]	Athens Greece (2024)	SVR	PM2.5, CO, NO2, O3, CO2	1 classroom (25 students)	(a) Improved PM2.5 prediction (R2 from 0.6 → 0.9) (b) Achieved CO2 error < 20 ppm (c) Validated low-cost IAQ monitoring with strong calibration accuracy in small-scale deployment
[90]	Ponte de Sor, Portugal (2023)	Statistical Analysis + Teacher Surveys	CO2, PM10, T, RH	9 classrooms, 171 sessions	(a) Revealed frequent exceedances (46% T, 32% PM10, 27% CO2) (b) Exposed mismatch between perceived vs. measured IAQ; (c) Underlined importance of awareness and real-time feedback.
[26]	Navarra Spain (2022)	DL (TCN)&ML Forecasting	CO2	15 schools	(a) Proved superiority of TCNs for long-horizon forecasts (>30 min, R2 > 0.9) (b) Confirmed ML’s role in demand-driven ventilation control and proactive IAQ management.
[149]	Smart Classrooms (SITA), Asia (2023)	Privacy-preserving ML (SITA, edge AI)	CO2, PM, VOCs, T	IoT deployment	(a) Showed viability of privacy-by-design AI (b) Achieved accurate HVAC optimization with local, low-latency processing (c) Reinforced trust in AI adoption by safeguarding data security.
[150]	Beijing, China (2023)	AHP + ML-supported decision	PM2.5, CO2, TVOCs, T, RH	15 schools	(a) Identified portable filters + controlled ventilation as optimal under severe outdoor pollution (b) Demonstrated need for context-specific IAQ strategies guided by decision frameworks.
[142]	Finland (2017)	Supervised ML + participatory feedback	CO2, VOCs, T, RH, bioaerosols	6 schools + national program	(a) Confirmed systemic IAQ problems (frequent exceedances) (b) Emphasized shift from inspection-based to predictive AI monitoring (c) Highlighted role of transparency and trust in national health programs.
[141]	Lower Saxony, Germany (2021)	Continuous monitoring, trend analysis	CO2, noise, T, RH	329 classrooms, 50 schools	(a) Documented widespread CO2 exceedances (b) Revealed variability due to room design/ventilation (c) Made strong case for AI-driven alerts during pandemic conditions.
[32]	Guilford, UK (2024)	IoT-based visual and visual-acoustic CO2 feedback systems (real-time AI feedback)	CO2, PM2.5, PM10	1 classroom	(a) Visual alarms reduced CO2 by 20% (b) All PM concentrations remained within WHO limits (c) IoT feedback systems improve air quality
[146]	Alicante, Spain (2023)	Artificial Neural Networks (ANN)	CO2, Real-time occupancy, Environmental variables	University classrooms	(a) Achieved 97.8% accuracy in classifying ventilation conditions (b) Integrates CO2, Wi-Fi occupancy, and environmental variables (c) Demonstrates high reliability with minimal false positives/negatives
[27]	Codsall, UK (2025)	Machine Learning (ML) models: RNN, LSTM, GRU, CNN	CO2, PM, T, RH, Formaldehyde, environmental variables	Two classrooms (35 students each)	(a) All models achieved > 92% predictive accuracy (b) Models enabled adaptive HVAC control balancing IAQ& energy use (c) Provides a replicable, data-driven model for other schools and learning spaces
[31]	North China (2018–2019)	Hybrid model EMD (Empirical Mode Decomposition), LSTM, BO (Bayesian Optimization)	Indoor CO2 concentration (time-series data)	Long-term dataset covering one full academic year	(a) 55% reduction in MAE for predictions up to 30 min ahead (b) Maintained R2 > 95% across forecasts. (c) Demonstrated robustness in predicting nonlinear and fluctuating CO2 patterns.
[144]	Hong Kong (2022)	- YOLOv5 (computer vision, deep learning) - CFD simulation - Fuzzy logic control for dynamic HVAC adjustment	Occupant number & spatial distribution - Thermal comfort index (PMV) - Air temperature & air velocity	University classrooms	(a) Identified significant spatial variations in thermal comfort linked to occupancy patterns. (b) System could predict and stabilize PMV rapidly under dynamic conditions. (c) Improved thermal comfort while offering potential energy savings.
[143]	Seoul, South Korea (2021)	Integrated Neural Network (INN)	PMV, CO2, PM10	1 school	(a) INN-based strategy predicts PMV, CO2, and PM levels one control cycle ahead. (b) Maintains CO2 below 700 ppm and decreases PM exceedances. (c) Achieves up to ~9% heating and ~16% cooling energy savings under closed-window conditions. (d) Offers a robust, adaptive, energy-efficient approach suitable for dynamic school environments.
[29]	Central China (2022)	RF (Random Forest)-TPE Tree-structured Parzen Estimator -LSTM Hybrid model	CO2, PM, T, H, O2, Illumination, Indoor population	One university classroom monitored for ~1.5 months	(a) RF-TPE-LSTM outperformed other predictive models (MAE, RMSE, MAPE, R2). (b) Achieved R2 > 98% for 10-min-ahead CO2 forecasts. (c) Incorporating occupancy and environmental factors improves prediction accuracy.
[145]	Montreal, Canada (2020–2021)	Bayesian parameter estimation to infer ventilation rates, CO2 emission, and noise levels	CO2, Ventilation, Noise	2 classrooms	(a) Ventilation rates are often below recommended standards. (b) Suggested use of Supplementary Air Cleaning Devices to improve IAQ. (c) Established CO2 thresholds aligned with ASHRAE standards for aerosol transmission risk management. (d) Provides a robust, uncertainty-aware approach for optimizing ventilation strategies in schools.
[147]	Pombal, Portugal (2013)	Multi-Layer Perceptron (MLP) neural network	CO2, T, H	2 classrooms	(a) MLP model using humidity + CO2 achieved: Mean Squared Error = 1.99 R2 = 0.96 (p < 0.001) MAE = ~1 occupant (b) Demonstrates accurate occupancy reconstruction from environmental data. (c) Supports improved IEQ control and energy-efficient building management. (d) Validates ML approaches for dynamic occupancy estimation in classrooms.
[129]	Riga, Latvia (2024)	Machine Learning (ML) models: Prophet, Transformer, Kolmogorov–Arnold Networks (KAN), LSTM, GRU	CO2, T, H	128 sensors	(a) KAN and GRU models outperformed others; GRU was most computationally efficient. (b) Hyperparameter optimization improved forecasting accuracy. (c) Sensor clusterization + individual modeling enhanced both accuracy and efficiency. (d) Demonstrates a digital shadow framework for healthier, energy-efficient indoor environments in public buildings.
[148]	Florida, USA (2021)	Hybrid PCA (Principal Component Analysis)– LMBP (Levenberg–Marquardt Back propagation model	PM2.5, PM10, NO2, O3	Multiple building types: classrooms, offices, laboratories; continuous monitoring at 10 min intervals over two-week periods	(a) PCA-LMBP model outperformed conventional methods for predicting IAQ (b) Strong associations found between indoor pollutant levels and: Proximity to traffic Building envelope integrity (cracks, peeling paint) Outdoor pollutant infiltration (c) Provides insights for targeted ventilation, maintenance, and IAQ improvement strategies in educational facilities.

Table 6. Real-world case studies on AI applications for IAQ management in schools, summarizing methods, monitored parameters, scale, and main outcomes.

Reference	Location/Year	AI Method	Parameters Monitored	Sample Size	Main Results/Critical Insights
[139,140]	Boston, USA (2023)	ML (Decision Trees)	CO2, PM2.5, PM10, CO, T, RH	4400 classrooms, 3659 sensors	(a) Demonstrated feasibility of large-scale IAQ monitoring (b) Reduced CO2 by 25–30% (c) Enabled real-time fault detection and improved health indicators (d) Highlighted value of teacher engagement in decision-making
[98]	Dunedin, New Zealand (2022)	Hybrid DL (LSTM + Autoencoder)	CO2	74 sensors, 247 k readings	(a) Achieved 99.5% anomaly detection accuracy (b) Outperformed k-NN and fuzzy clustering (c) Proved capacity of hybrid DL models to generalize patterns in IAQ data.
[30,64]	Athens Greece (2024)	SVR	PM2.5, CO, NO2, O3, CO2	1 classroom (25 students)	(a) Improved PM2.5 prediction (R2 from 0.6 → 0.9) (b) Achieved CO2 error < 20 ppm (c) Validated low-cost IAQ monitoring with strong calibration accuracy in small-scale deployment
[90]	Ponte de Sor, Portugal (2023)	Statistical Analysis + Teacher Surveys	CO2, PM10, T, RH	9 classrooms, 171 sessions	(a) Revealed frequent exceedances (46% T, 32% PM10, 27% CO2) (b) Exposed mismatch between perceived vs. measured IAQ; (c) Underlined importance of awareness and real-time feedback.
[26]	Navarra Spain (2022)	DL (TCN)&ML Forecasting	CO2	15 schools	(a) Proved superiority of TCNs for long-horizon forecasts (>30 min, R2 > 0.9) (b) Confirmed ML’s role in demand-driven ventilation control and proactive IAQ management.
[149]	Smart Classrooms (SITA), Asia (2023)	Privacy-preserving ML (SITA, edge AI)	CO2, PM, VOCs, T	IoT deployment	(a) Showed viability of privacy-by-design AI (b) Achieved accurate HVAC optimization with local, low-latency processing (c) Reinforced trust in AI adoption by safeguarding data security.
[150]	Beijing, China (2023)	AHP + ML-supported decision	PM2.5, CO2, TVOCs, T, RH	15 schools	(a) Identified portable filters + controlled ventilation as optimal under severe outdoor pollution (b) Demonstrated need for context-specific IAQ strategies guided by decision frameworks.
[142]	Finland (2017)	Supervised ML + participatory feedback	CO2, VOCs, T, RH, bioaerosols	6 schools + national program	(a) Confirmed systemic IAQ problems (frequent exceedances) (b) Emphasized shift from inspection-based to predictive AI monitoring (c) Highlighted role of transparency and trust in national health programs.
[141]	Lower Saxony, Germany (2021)	Continuous monitoring, trend analysis	CO2, noise, T, RH	329 classrooms, 50 schools	(a) Documented widespread CO2 exceedances (b) Revealed variability due to room design/ventilation (c) Made strong case for AI-driven alerts during pandemic conditions.
[32]	Guilford, UK (2024)	IoT-based visual and visual-acoustic CO2 feedback systems (real-time AI feedback)	CO2, PM2.5, PM10	1 classroom	(a) Visual alarms reduced CO2 by 20% (b) All PM concentrations remained within WHO limits (c) IoT feedback systems improve air quality
[146]	Alicante, Spain (2023)	Artificial Neural Networks (ANN)	CO2, Real-time occupancy, Environmental variables	University classrooms	(a) Achieved 97.8% accuracy in classifying ventilation conditions (b) Integrates CO2, Wi-Fi occupancy, and environmental variables (c) Demonstrates high reliability with minimal false positives/negatives
[27]	Codsall, UK (2025)	Machine Learning (ML) models: RNN, LSTM, GRU, CNN	CO2, PM, T, RH, Formaldehyde, environmental variables	Two classrooms (35 students each)	(a) All models achieved > 92% predictive accuracy (b) Models enabled adaptive HVAC control balancing IAQ& energy use (c) Provides a replicable, data-driven model for other schools and learning spaces
[31]	North China (2018–2019)	Hybrid model EMD (Empirical Mode Decomposition), LSTM, BO (Bayesian Optimization)	Indoor CO2 concentration (time-series data)	Long-term dataset covering one full academic year	(a) 55% reduction in MAE for predictions up to 30 min ahead (b) Maintained R2 > 95% across forecasts. (c) Demonstrated robustness in predicting nonlinear and fluctuating CO2 patterns.
[144]	Hong Kong (2022)	- YOLOv5 (computer vision, deep learning) - CFD simulation - Fuzzy logic control for dynamic HVAC adjustment	Occupant number & spatial distribution - Thermal comfort index (PMV) - Air temperature & air velocity	University classrooms	(a) Identified significant spatial variations in thermal comfort linked to occupancy patterns. (b) System could predict and stabilize PMV rapidly under dynamic conditions. (c) Improved thermal comfort while offering potential energy savings.
[143]	Seoul, South Korea (2021)	Integrated Neural Network (INN)	PMV, CO2, PM10	1 school	(a) INN-based strategy predicts PMV, CO2, and PM levels one control cycle ahead. (b) Maintains CO2 below 700 ppm and decreases PM exceedances. (c) Achieves up to ~9% heating and ~16% cooling energy savings under closed-window conditions. (d) Offers a robust, adaptive, energy-efficient approach suitable for dynamic school environments.
[29]	Central China (2022)	RF (Random Forest)-TPE Tree-structured Parzen Estimator -LSTM Hybrid model	CO2, PM, T, H, O2, Illumination, Indoor population	One university classroom monitored for ~1.5 months	(a) RF-TPE-LSTM outperformed other predictive models (MAE, RMSE, MAPE, R2). (b) Achieved R2 > 98% for 10-min-ahead CO2 forecasts. (c) Incorporating occupancy and environmental factors improves prediction accuracy.
[145]	Montreal, Canada (2020–2021)	Bayesian parameter estimation to infer ventilation rates, CO2 emission, and noise levels	CO2, Ventilation, Noise	2 classrooms	(a) Ventilation rates are often below recommended standards. (b) Suggested use of Supplementary Air Cleaning Devices to improve IAQ. (c) Established CO2 thresholds aligned with ASHRAE standards for aerosol transmission risk management. (d) Provides a robust, uncertainty-aware approach for optimizing ventilation strategies in schools.
[147]	Pombal, Portugal (2013)	Multi-Layer Perceptron (MLP) neural network	CO2, T, H	2 classrooms	(a) MLP model using humidity + CO2 achieved: Mean Squared Error = 1.99 R2 = 0.96 (p < 0.001) MAE = ~1 occupant (b) Demonstrates accurate occupancy reconstruction from environmental data. (c) Supports improved IEQ control and energy-efficient building management. (d) Validates ML approaches for dynamic occupancy estimation in classrooms.
[129]	Riga, Latvia (2024)	Machine Learning (ML) models: Prophet, Transformer, Kolmogorov–Arnold Networks (KAN), LSTM, GRU	CO2, T, H	128 sensors	(a) KAN and GRU models outperformed others; GRU was most computationally efficient. (b) Hyperparameter optimization improved forecasting accuracy. (c) Sensor clusterization + individual modeling enhanced both accuracy and efficiency. (d) Demonstrates a digital shadow framework for healthier, energy-efficient indoor environments in public buildings.
[148]	Florida, USA (2021)	Hybrid PCA (Principal Component Analysis)– LMBP (Levenberg–Marquardt Back propagation model	PM2.5, PM10, NO2, O3	Multiple building types: classrooms, offices, laboratories; continuous monitoring at 10 min intervals over two-week periods	(a) PCA-LMBP model outperformed conventional methods for predicting IAQ (b) Strong associations found between indoor pollutant levels and: Proximity to traffic Building envelope integrity (cracks, peeling paint) Outdoor pollutant infiltration (c) Provides insights for targeted ventilation, maintenance, and IAQ improvement strategies in educational facilities.

5. Concluding Remarks, Limitations and Future Challenges

This review has aimed to synthesize methodological advances, assess outcomes across diverse locations and settings, identify system barriers, and highlight future pathways.

To achieve these objectives, this review has examined the emerging role of AI in the assessment and management of IAQ in educational environments, synthesizing evidence from twenty case studies spanning different geographical regions and socio-technical contexts. The examination of outcomes has shown that AI has progressed from conceptual exploration to practical application, delivering measurable benefits in pollutant forecasting, anomaly detection, real-time fault diagnosis, and exposure mitigation. Large-scale deployments, such as the Boston initiative with more than 3600 sensors across 4400 classrooms [139,140] and the Lower Saxony network covering 329 classrooms [141], confirm the technical feasibility of AI-driven monitoring at scale. In parallel, smaller but methodologically innovative studies, including hybrid deep learning frameworks in Dunedin [98] and support vector regression with feature engineering in Athens [30], have shown that advanced models can outperform conventional methods, offering more reliable pollutant predictions and enabling proactive ventilation strategies. Collectively, these initiatives highlight the capacity of AI to serve as a practical instrument for safeguarding student health and enhancing educational outcomes.

Despite these advances, the transition from promising pilots to sustainable, system-wide adoption remains constrained by several barriers. Sensor reliability and calibration drift continue to undermine predictive accuracy, particularly in low-cost monitoring networks [30,90,151,152,153,154,155,156]. AI models often require periodic retraining to accommodate dynamic occupancy, HVAC variability, and evolving environmental standards [39,153,157,158], increasing both technical complexity and operational costs. Lack of interoperability with legacy HVAC and Building Management Systems (BMS) further impedes seamless integration, forcing reliance on parallel platforms [159,160]. Financial constraints—including installation, infrastructure, and long-term maintenance—pose additional barriers, especially for underfunded schools [90,156]. Finally, unresolved concerns regarding privacy, trust, and ethical acceptability [161] remain critical, even as edge-based frameworks such as SITA [149]. Edge-computing frameworks such as SITA offer potential pathways forward.

These challenges are not purely technical; they also carry significant social and equity implications. Schools with greater financial and technical capacity are better positioned to adopt advanced AI-driven IAQ solutions, while under-resourced institutions risk exclusion, thereby deepening existing inequalities in health and educational outcomes. Addressing these disparities requires reframing AI not only as a tool for technical optimization but also as a mechanism for promoting fairness, accessibility, and social responsibility.

Looking at future pathways, several priorities define the future research and policy agenda. First, standardized and open-access datasets are essential to support benchmarking, model validation, and cross-context generalizability. Second, advances in explainable AI (XAI) are needed to enhance transparency and foster trust among educators, parents, and policymakers. Third, innovation in low-cost yet reliable sensing technologies and cost-effective retrofitting strategies is vital for enabling equitable deployment in both new and existing school buildings. Finally, stronger integration with policy frameworks—including data governance, privacy protection, and accountability structures—will be crucial for sustainable implementation.

Ultimately, the adoption of AI for IAQ management in educational settings must be recognized as more than a technological intervention: it is a transformative public health and educational priority. Poor classroom air quality directly impacts student well-being, cognitive development, and long-term health. By embedding AI systems within broader strategies for school health, sustainability, and equity, their role can expand from fragmented pilots to globally scalable solutions.

Its long-term success, however, will depend not only on achieving algorithmic accuracy but also on overcoming systemic challenges of reliability, interoperability, cost, privacy, and trust. Addressing these interlinked issues will allow AI to deliver sustainable, transparent, and equitable improvements to learning environments, ensuring healthier conditions for future generations.