A Review of Indoor Thermal Comfort Studies on Older Adults in China

Abstract

This review systematically examines research on indoor thermal comfort for older adults conducted in China since 2010. It highlights several existing research gaps, including the lack of a systematic understanding of environmental and individual influences, limitations of thermal comfort models, challenges in their optimization, and inadequate integration of intelligent technologies. Results indicate that environmental factors usually exert a greater impact on the elderly’s neutral temperature than individual factors. Thermal comfort models differ in predictive accuracy, data requirements, and applicability. The adaptive predicted mean vote (aPMV) model is better suited for group-level assessments. Machine learning (ML) models, featuring high flexibility and accuracy, are more appropriate for personalized predictions. In addition, physiological parameters could play a critical role in thermal assessments. When integrated with ML models, physiological parameters could further improve predictive accuracy. When integrated into artificial intelligence (AI) and Internet of Things (IoT) systems, forehead and back skin temperatures could act as early-warning indicators during heat exposure, while lower-limb temperatures are more indicative of thermal discomfort during cold exposure. Overall, this review summarizes current progress and limitations, offering a reference for the development of user-friendly modeling and intelligent temperature-control systems.

1. Introduction

1.1. Background on Thermal Comfort Among Older Adults in China

China is undergoing a demographic transformation, with a rapidly growing elderly population. According to the United Nations’ definition [1] and the Chinese law on the protection of the rights and interests of the elderly [2], individuals aged 60 or older are classified as elderly. As of 2021, China had approximately 264 million people aged 60 and above, accounting for 18.7% of its population of 1.4 billion. This figure is expected to exceed 400 million by 2035, marking China’s transition into a deeply aged society [3,4].

With aging, older adults become less mobile and spend more time indoors [5]. In China, approximately 64% of elderly urban residents live in aging residential buildings over 20 years old, often featuring suboptimal design and insufficient thermal comfort [6]. The majority of older Chinese adults living in rural areas reside in houses without professionally designed insulation, resulting in significant fluctuations in indoor thermal conditions [7]. In addition to building insulation issues, accelerated climate change has heightened the risk of thermal discomfort among the elderly. In southern China, extreme heat events have become increasingly frequent. Many older adults use air conditioning less frequently due to lower incomes and frugal habits, increasing their susceptibility to heat stress and associated health problems [8]. Similarly, cold environments in northern China also pose significant health threats. Older adults are particularly vulnerable to freezing [9]. Cardiovascular disease, a major cause of death, reaches its peak in winter [10]. Moreover, as aging progresses, especially with disabilities or cognitive decline, the ability to perceive and express thermal comfort needs weakens [11]. Overall, the quality of the indoor thermal environment is crucial not only for health and comfort but also for overall well-being [12,13,14]. However, the current thermal environment design standards do not sufficiently reflect the specific needs of the elderly population and provide limited guidance for age-friendly design [15,16,17,18,19].

1.2. Limitations of Existing Thermal Comfort Research

The present indoor thermal environment assessment standard GB 50736-2012 [20] mainly relies on the PMV and predicted percentage of dissatisfied (PPD) indices, which define acceptable temperature ranges for summer and winter in mechanically conditioned environments [21,22]. The PMV model has limited applicability for evaluating thermal comfort in older adults, primarily due to the lack of data from older adults [12]. Consequently, it has several limitations, such as insufficient consideration of age-related physiological changes and limited evaluation of health risks [15,23,24]. Compared with younger adults, older adults exhibit reduced thermal sensitivity and weaker physiological responses, necessitating larger temperature changes to perceive equivalent sensations [21,25,26,27]. They also have limited environmental regulation capacity, often relying on passive strategies such as opening windows or adjusting clothing, rather than active cooling [28,29,30,31,32]. These age-related differences lead to distinct neutral temperatures and narrower comfort ranges among older adults [33,34,35].

In addition, thermal comfort variations exist not only between older and younger adults but also among older adults themselves. These differences are mainly shaped by a combination of environmental and individual factors, including climatic zone, urban–rural context, physical condition, and adaptive behaviors [36,37,38]. In recent years, although numerous studies have examined thermal environments across various groups and regions, the findings remain fragmented and inconsistent. Differences are often presented in isolation within individual studies, lacking systematic integration.

To address the limitations of the PMV model, various evaluation methods have been proposed. Modified PMV models, such as the adaptive PMV and the Griffiths method, have demonstrated improved predictive performance. Compared with the PMV and modified PMV models, ML models offer distinct advantages in integrating multiple parameters and capturing nonlinear relationships, substantially improving prediction accuracy. Nonetheless, several challenges remain. First, selecting appropriate ML algorithms and tuning their parameters is complex [39], and the lack of model standardization reduces generalizability. Second, ML models require large, high-quality datasets, which are difficult to obtain and accurately label for elderly populations [40].

1.3. Objectives and Contributions of This Review

Overall, this review is organized around a framework comprising three stages: understanding, evaluating, and improving thermal comfort. First, this review summarizes the environmental and individual factors influencing the thermal comfort of older adults. Then, it compares existing thermal comfort models in terms of their structural characteristics, required sample size, and predictive accuracy, and further examines how these models account for the influencing factors. Finally, it proposes specific directions for model optimization and the development of intelligent temperature control applications. The specific objectives are as follows:

(1)

Identify and clarify how environmental and individual factors influence the thermal comfort of older adults.

(2)

Evaluate and compare the applicability of existing thermal comfort models to identify their strengths, limitations, and contexts of optimal use.

(3)

Propose specific directions for model optimization and future intelligent applications.

The framework of this review is shown in Figure 1. Notably, although all cases considered in this review are from China, the comprehensive research framework is generalizable and can be applied to other regions and countries.

2. Method

2.1. Literature Identification

This review followed the PRISMA guidelines. The review included studies published from 2010 onwards. Since then, improvements in living conditions and indoor environments in China have changed residential and daily living patterns [41,42,43], making earlier studies potentially less representative of current thermal environments for older adults. Searches were performed in the Web of Science, Scopus, and China National Knowledge Infrastructure (CNKI) databases. Only peer-reviewed journal articles were included. For CNKI, only articles published in Peking University Core Journals or indexed in the Chinese Social Sciences Citation Index were selected to ensure research quality. The review focused on studies conducted in mainland China, Hong Kong, and Taiwan.

The search strategy was based on three core concepts: (1) elderly population (e.g., elderly, aged, older adults, senior); (2) thermal comfort–related terms (e.g., thermal comfort, thermal requirement, thermal sensation, thermal perception, thermal adaptation, thermal satisfaction, thermal preference, thermal environment); and (3) exclusion of outdoor studies.

A total of 453 records were identified, including 423 from international databases and 30 from CNKI. After removing 155 duplicates, 298 articles remained for screening. Based on titles and abstracts, 181 irrelevant articles were excluded, leaving 117 full-text articles for eligibility assessment. Among these, the reasons for exclusion were as follows: 1 study conducted in outdoor environments, 16 studies lacking thermal sensation or physiological data under thermal exposure, and 5 studies that defined the elderly population as younger than 60 years old. Finally, 95 studies met the inclusion criteria. The screening process is illustrated in Figure 2 using the PRISMA flow diagram. The PRISMA Checklist for this review is provided in the Supplementary Materials.

2.2. Keywords Frequency Analysis

Based on the selected literature, a keyword co-occurrence map was generated using VOSviewer version 1.6.20. Figure 3 illustrates that the core keyword “thermal comfort” is centrally positioned and frequently co-occurs with terms such as elderly population, indoor environment, physiology, surveys, and air temperature. This indicates a focused research effort on the subjective and physiological responses of older adults to indoor thermal conditions. In recent years, emerging keywords such as ML, predictive models, thermal sensation votes, and health risks have reflected a shift toward data-driven modeling and health-oriented assessment. Meanwhile, frailty, cold climate zones, and rural housing show increasing attention to vulnerable subgroups and region-specific conditions, especially in colder or rural settings. The field is evolving from traditional surveys and statistical analyses toward integrated, multidimensional approaches emphasizing smart modeling, precision targeting, and contextual sensitivity. These thematic trends provide a foundation for the structure of this review.

3. Results

3.1. Environmental Factors Influencing Thermal Comfort in Older Adults

3.1.1. Climate Zones

China spans a vast geographical range from north to south, resulting in significant climatic diversity, as shown in Figure 4. China’s thermal climate zones are classified according to the GB 50176-2016 [44], and are divided into five categories: severe cold zone (SCZ), cold zone (CZ), hot summer and cold winter zone (HSCWZ), hot summer and warm winter zone (HSWWZ), and mild zone (MZ). Due to the significant temperature differences between winter and summer across these climate zones, the thermal environment, thermal comfort, thermal sensation, and adaptive behaviors of older people vary considerably [7,45,46,47]. For example, studies in Taiwan [5] and Shanghai [37] found that clothing adjustment is the primary adaptive strategy in both summer and winter, although it is more effective in summer than in winter. Meanwhile, another study showed that clothing adjustment is most responsive to temperature changes during transitional seasons, while in summer and winter, extreme conditions or limited adjustment options reduce its correlation with temperature [48]. In addition, elderly individuals exhibit significant differences in their tolerance and sensitivity to heat and cold, which collectively shape diverse patterns of thermal comfort across various climate zones.

Figure 5 presents the relationships between thermal sensation vote (TSV) and indoor temperature among elderly residents in urban areas across four climate zones: CZ in Xi’an [48], HSCWZ in Shanghai [37], MZ in Guiyang [49], and HSWWZ in Taiwan [5], during winter and summer. Overall, seasonal and climatic differences are evident. The regression slopes are generally steeper in summer, indicating that older people are more sensitive to high temperatures. Among all cities, Shanghai’s neutral temperature is the highest in summer and the lowest in winter, with a temperature difference of 8.8 °C, likely due to the hot and humid summer conditions and the lack of central heating in winter, which may contribute to stronger thermal adaptability. In Xi’an, both the neutral temperature and thermal sensitivity are at moderate levels. In Taiwan, the neutral temperatures in winter and summer are nearly identical, with a difference of only 2 °C.

3.1.2. Urban–Rural Differences

Urban–rural differences significantly influence the thermal comfort of older adults, primarily due to variations in lifestyle, heating systems, and building insulation [50]. Figure 6 shows the typical building forms and heating facilities in urban and rural settings. Urban dwellings generally have better insulation, helping maintain stable indoor temperatures throughout the year [50]. In contrast, most rural homes lack professional insulation design, and only a few economically developed regions have adopted basic insulation measures [7]. As a result, rural buildings experience faster heat loss in winter and greater heat gain in summer [7,50,51,52,53].

Heating conditions also differ substantially between urban and rural areas. Urban households often benefit from district central heating, while rural residents rely on decentralized systems such as individual water heaters and traditional kang stoves. A survey conducted in Beijing revealed that 85% of urban elderly households were connected to municipal heating networks, whereas 70% of rural households used individual systems [38]. A field study conducted in Hunan found that 60.4% of urban older adults used independent radiators, while 39.6% used electric heaters or air conditioners; in contrast, all rural elderly relied solely on simple “warm buckets” [54]. Consequently, indoor temperatures in winter were typically 5–6 °C higher in urban homes than in rural ones, where the use of coal or biomass stoves led to larger temperature fluctuations [38,54,55]. During summer, rural homes were about 2–4 °C warmer than urban homes [48,50,56].

Differences in lifestyle further shape the patterns of thermal comfort. Urban older adults primarily engage in indoor activities and can adjust their clothing accordingly, such as removing outer layers or switching to short-sleeved shirts. In contrast, rural older adults, especially those living in mountainous regions, frequently move between outdoor farm work and indoor environments, which exposes them to larger temperature variations [50]. As a result, rural older adults usually wear thicker or multilayered clothing, and their average clothing insulation value is significantly higher than that of urban residents [38,54,55,57,58]. These differences in building insulation, heating systems, and daily behavior contribute to distinct thermal sensitivities and neutral temperatures between urban and rural older adults.

Figure 7 compares the relationship between TSV and indoor temperature among older adults living in urban and rural areas of Xi’an [48,56,58], Hunan [54], and Guilin [50] during winter and summer. The results indicate that, except for Xi’an in winter, where the urban–rural neutral temperature difference is relatively large at 7.98 °C, the differences in other cases are all below 1 °C. In Xi’an during summer, the urban–rural neutral temperature difference is only 0.13 °C. This pattern may be explained by the strong thermal adaptation of rural older adults, which reduces the urban–rural difference and results in similar neutral temperatures in most cases.

3.2. Individual Factors Influencing Thermal Comfort in Older Adults

3.2.1. Gender

Research on gender differences remains limited. Most studies indicate that gender has no significant impact on overall thermal sensation, with only minor differences in neutral temperatures between elderly males and females during summer and winter [5,37,59,60,61,62,63]. However, older women tend to be more sensitive to high temperatures and humidity fluctuations, experiencing greater thermal discomfort and reduced concentration, whereas they are generally more comfortable under cooler conditions. In contrast, older men generally demonstrate greater heat tolerance [11]. A field study conducted in Sichuan further confirmed this gender disparity, showing that the proportion of males reporting an “extremely hot” sensation was significantly lower than that of females [51].

3.2.2. Age Groups

Age has been shown to influence thermal preferences among older adults [64], with different age groups exhibiting distinct adaptive behaviors, preferred neutral temperatures, and thermal comfort ranges [22].

Studies conducted in rural areas of Guanzhong [65] and Inner Mongolia [66] during the winter categorized older adults into three age groups: 60–69 years, 70–79 years, and 80 years and above. Results showed that with increasing age, cold tolerance declined, while preferred temperatures increased and thermal sensitivity decreased. Older adults over 80 tended to wear thicker clothing and preferred warmer environments, reflected in higher neutral temperatures and lower comfort thresholds.

A study conducted in the urban areas of Baoding during the summer categorized older adults into three age groups: 60–69 years, 70–89 years, and 90–99 years [67]. It was found that with increasing age, thermoregulatory functions declined, resulting in reduced thermal sensitivity. Meanwhile, the upper limits and widths of the temperature and humidity comfort zones expanded significantly, suggesting a broader environmental tolerance among the oldest group. Another summer study conducted in urban Baoding classified older adults into three groups: 60–74 years, 75–88 years, and 89 years and above. Similarly, thermal sensitivity decreased with age, while environmental tolerance increased in the oldest group [68].

Figure 8 compares the relationship between TSV and indoor temperature across age groups in these studies. The differences in neutral temperature among various age groups are relatively small, ranging from 0.41 °C to 1.45 °C. The overall trend shows that the neutral temperature rises with age, while sensitivity to environmental changes declines.

3.2.3. Frailty Levels

Frailty, a common geriatric syndrome involving multisystem functional decline, directly reduces thermoregulatory capacity. The commonly used frailty phenotype proposed by Fried et al. [69] classifies individuals according to five criteria. These include unintentional weight loss of at least 4.5 kg in the past year not caused by dieting or illness, reduced grip strength within the lowest 20 percent for sex and BMI group, fatigue or exhaustion assessed by two items from the revised Center for Epidemiologic Studies Depression Scale, slow walking speed based on a 15-foot walking time within the slowest 20 percent for sex and height group, and low physical activity reflected by weekly energy expenditure within the lowest 20 percent for sex group. Individuals who meet none of these criteria are classified as non-frail, those who meet one or two as pre-frail, and those who meet three or more as frail. Frailty levels could reflect physiological differences among older adults [70]. Older adults with varying levels of frailty exhibit significant differences in physiological regulation, behavioral adjustments, and thermal perception and preferences.

As for physiological regulation, pre-frail older adults show reduced thermoregulatory responsiveness compared to non-frail older adults. Pre-frail older adults exhibit slower stabilization of skin temperature, especially in the limbs, during transitions from warm to neutral environments [71]. Their sweating responses are weaker, with faster reduction during cooling and delayed onset during heating [71]. Due to their impaired cardiovascular responses, they exhibit elevated heart rates and delayed pulse index adjustment when temperature changes exceed 5 °C [71]. Additionally, they are more sensitive to air movement in warm environments, especially around the head, lower legs, and feet, and less tolerant of higher airspeeds [72].

Regarding behavioral regulation, frailty levels have a significant influence on thermal comfort behaviors in older adults. During winter, individuals with higher frailty levels report greater discomfort, rely more heavily on heating devices, and tend to keep their windows closed to maintain indoor thermal stability [73]. They also wear more insulating clothing and engage in fewer daily activities, reducing their metabolic rate [73]. In summer, they rarely use air conditioners or fans, instead opting for increased clothing and reduced outdoor activity to avoid exposure to temperature fluctuations [70].

In terms of thermal perception and subjective preferences, older adults with higher frailty levels tend to exhibit greater thermal sensitivity, prefer higher neutral temperatures, and have narrower thermal comfort ranges [70,73].

Table 1 shows that increasing frailty is associated with a stronger preference for warmer environments, higher neutral temperatures, and narrower comfort ranges. The neutral temperature rises by 2.1 °C with increasing frailty, while the comfort range shifts upward and narrows by about 1–2 °C.

Table 1. Thermal comfort characteristics of older adults with different frailty levels.

	Non-Frail Older Adults	Pre-Frail Older Adults	Frail Older Adults
Neutral temperature	25.8 °C in summer	26.9 °C in summer	27.9 °C in summer
Sensitivity of TSV to temperature (Slope)	0.1139 in summer	0.2001 in summer	0.3399 in summer
Temperature difference triggering significant TSV differences	>5 °C	>3 °C	NA
Comfortable temperature range	24.0–30.0 °C in summer; 18–22 °C in winter	24.3–29.3 °C in summer; 19–23 °C in winter	25.9–29.3 °C in summer; 20–24 °C in winter

Table 1. Thermal comfort characteristics of older adults with different frailty levels.

	Non-Frail Older Adults	Pre-Frail Older Adults	Frail Older Adults
Neutral temperature	25.8 °C in summer	26.9 °C in summer	27.9 °C in summer
Sensitivity of TSV to temperature (Slope)	0.1139 in summer	0.2001 in summer	0.3399 in summer
Temperature difference triggering significant TSV differences	>5 °C	>3 °C	NA
Comfortable temperature range	24.0–30.0 °C in summer; 18–22 °C in winter	24.3–29.3 °C in summer; 19–23 °C in winter	25.9–29.3 °C in summer; 20–24 °C in winter

3.3. Evaluation Methods

3.3.1. Modified PMV Models

The widely used PMV model is based on specific experimental conditions and does not adequately account for regional or age-related differences [21,66,74,75,76]. Therefore, based on living habits and adaptive capacity, several modified models and methods have been developed.

The aPMV is an adaptive thermal comfort model that introduces a correction factor (λ) to the traditional PMV to better account for thermal adaptation. According to GB/T 50785-2012 [77], λ is set at 0.21 for PMV ≥ 0 (warm conditions) and −0.49 for PMV < 0 (cold conditions). However, these fixed values are often inaccurate and vary significantly across regions, as shown in Table 2. Winter aPMV λ values range from −0.49 to −0.26, and summer from 0.065 to 0.272, indicating greater variability in cold adaptation than in heat adaptation.

$$\mathrm{aPMV}={\displaystyle \frac{\mathrm{PMV}}{1+\mathrm{\lambda }·\mathrm{PMV}}}$$

(1)

Besides these modified PMV models, Griffiths’ method was used to estimate the neutral temperature. Unlike regression-based TSV methods, it determines the neutral temperature by assuming a linear relationship between TSV and operative temperature, combined with a predefined thermal sensitivity constant.

$$T_n=T_{op}+{\displaystyle \frac{\left(0-\mathrm{TSV}\right)}{\mathrm{\alpha }}}$$

(2)

where T_n means neutral temperature, and T_op means operative temperature (°C).

For instance, Jiao et al. [37] used a Griffiths constant of 0.33 to calculate the neutral temperature for older adults in Shanghai. The results showed a winter neutral temperature of 14 °C, which was 2.6 °C lower than that derived from the TSV regression method, and a summer neutral temperature of 28 °C, which was 2.6 °C higher [37]. This discrepancy highlights the importance of considering thermal adaptation in older adults.

Apart from these, modified PMV models based on personalized regulation have also been proposed, which could more accurately reflect the unique physiological traits and daily habits of older adults. For example, based on the relatively fixed activity patterns and daily routines of older adults at home, Miao et al. [78] proposed a time-weighted PMV model, PMVt, incorporating temporal scales and habitual trajectories. This model calculates an integrated thermal comfort index by weighting the PMV values of different spaces according to the duration of stay and the proportion of valid records in each space [78]. The model demonstrated strong performance, achieving a high correlation with TSV (R² = 0.82) [78].

Additionally, considering the impact of age, Zheng et al. [79] proposed the mPMV model, which incorporates a logistic function with age and TSV as key correction parameters. The model achieved improved prediction performance by calculating both the age-related coefficient and the age-temperature correlation.

Table 2. Comparison of modified PMV models.

Model/Method	City or Region (Climate Zone)	Sample Size (Subjects)	Key Parameters	Neutral Temperature or Comfortable Range (°C)	Paper
aPMV, Griffiths Method	Shanghai (HSCWZ)	672	α = 0.33	Winter: 14 °C; Summer: 28 °C	[37]
aPMV	Xiangxi (HSCWZ)	92	Winter λ = −0.26	Comfort temperature range 16.7–27.1 °C	[74]
aPMV	Suining, Sichuan (HSCWZ)	177	Summer λ = 0.065	Acceptable temperature range 21.41–27.61 °C	[51]
aPMV, Griffiths Method	Weihai (CZ)	203	Cold λ = −0.38; Warm λ= 0.272; α = 0.33	Griffiths 21.63 °C	[16]
aPMV	Hefei (HSCWZ)	720 (60)	Summer λ = 0.21; Winter λ = −0.49	NA	[45]
mPMV	Baoding (CZ)	1535 (44)	Age coefficient BT (0.0012–0.0676)	27.8–28.3 °C	[79]
PMVt	Shanghai (HSCWZ)	447 (15)	Time-weighted coefficient	25.5 °C	[78]

Table 2. Comparison of modified PMV models.

Model/Method	City or Region (Climate Zone)	Sample Size (Subjects)	Key Parameters	Neutral Temperature or Comfortable Range (°C)	Paper
aPMV, Griffiths Method	Shanghai (HSCWZ)	672	α = 0.33	Winter: 14 °C; Summer: 28 °C	[37]
aPMV	Xiangxi (HSCWZ)	92	Winter λ = −0.26	Comfort temperature range 16.7–27.1 °C	[74]
aPMV	Suining, Sichuan (HSCWZ)	177	Summer λ = 0.065	Acceptable temperature range 21.41–27.61 °C	[51]
aPMV, Griffiths Method	Weihai (CZ)	203	Cold λ = −0.38; Warm λ= 0.272; α = 0.33	Griffiths 21.63 °C	[16]
aPMV	Hefei (HSCWZ)	720 (60)	Summer λ = 0.21; Winter λ = −0.49	NA	[45]
mPMV	Baoding (CZ)	1535 (44)	Age coefficient BT (0.0012–0.0676)	27.8–28.3 °C	[79]
PMVt	Shanghai (HSCWZ)	447 (15)	Time-weighted coefficient	25.5 °C	[78]

3.3.2. ML Models

ML models are computational frameworks that learn patterns from existing data and use the learned knowledge to make predictions on unseen data. ML has been applied in thermal comfort studies to develop group-based and personalized comfort models. It enables more accurate predictions than traditional PMV models and can be used to optimize air conditioning control, improve indoor environments, and enhance energy efficiency [80]. Furthermore, ML models can automatically uncover complex relationships and optimize predictions [81]. These models possess a self-updating ability, allowing them to adjust and fine-tune their predictions across diverse environmental and contextual scenarios, thereby ensuring better alignment with the evolving nature of thermal comfort [82].

Table 3 summarizes studies that applied ML in thermal comfort research for the Chinese elderly. Based on the study setting, these studies can be categorized into climate chamber experiments and field studies. Climate chamber studies provide a highly controlled environment, enabling a detailed examination of the mechanisms between environmental and physiological parameters. Field studies are conducted in real-life settings, taking into account the adaptive behaviors of older adults. ML models can incorporate a broader range of influencing factors and be tailored to different population groups.

In addition, the sample sizes varied significantly across studies, ranging from as few as 135 [83] to as many as 3440 [84]. Algorithm choices have also diversified, ranging from traditional methods to ensemble learning and neural networks, with XGBoost and Random Forest (RF) frequently showing superior accuracy.

Models can be categorized into four types based on the input parameters: those that use only environmental parameters, those that use only physiological parameters, those that combine environmental and physiological parameters, and those that further incorporate human-related factors. The latter two types have become increasingly common, reflecting a shift in thermal comfort research from ‘environmental determinism’ to a human–environment interaction model. Most studies used 3-point or 7-point TSV for output parameters. In terms of performance metrics, while accuracy remains widely used, many studies have adopted more detailed indicators such as precision, recall, F1 score, ROC curve, AUC, and even regression metrics like R², MAE, and MSE. This shift reflects a move toward more comprehensive model evaluation, focusing on correctness and error types, model stability, and real-world applicability.

Table 3. ML-based thermal comfort studies for the elderly in China.

Paper	Study Setting	Subjects (Sample Size)	Algorithm	Input Parameters	Output Parameter(s)	Generalization Test Method	Performance Metric(s)
[8]	Climate chamber	38 (NA)	GBDT, AdaBoost, XGBoost	Environmental, physiological, and heat exchange parameters	7-point TSV	80% of the dataset for training and 20% for testing	R2, MAE, and MSE
[84]	Field study (summer)	44 (3440)	AB, DT, GNB, KNN, RF, SVM, and XGBoost	Environmental, physiological, and human-related parameters	3-point TSV	80% of the dataset for training and 20% for testing	Precision, recall, accuracy (76%), ROC curve, AUC, and F1 score
[85]	Field study (summer)	14 (1389)	LR, DT, KNN, and SVM	Environmental and physiological parameters	3-point TSV	80% of the dataset for training and 20% for testing	Precision, accuracy (70%), ROC curve, AUC, and F1 score
[86]	Field study (summer)	20 (1865)	AB, RF, LR, ANN, and NB	Environmental parameters	3, 5, 7, and 9-point TSV	80% of the dataset for training and 20% for testing	Recall, accuracy (92.4%), ROC curve, AUC, and F1 score
[87]	Climate chamber	13 (964)	BP and RBF	Environmental and physiological parameters	3-point TSV	80% of the dataset for training and 20% for testing	Accuracy (87.82%)
[88]	Climate chamber	5 (NA)	SVM, RF, KNN, MLP-ANN	Physiological parameters	3-point TSV	10-fold cross-validation	Accuracy (81.2%), F1 score, ROC curve, and AUC
[81]	Field study (year-round)	1040 (724)	RF	Environmental and human-related parameters	3-point TSV	80% of the dataset for training and 20% for testing	Accuracy (56.6%) Accuracy (81.2%)
	Climate chamber	18 (372)	RF	Physiological parameters
[83]	Field study (winter)	35 (135)	ANN, LoR, SVM, KNN, DT, and NB	Environmental and physiological parameters	7-point TSV	80% of the dataset for training and 20% for	Accuracy (67.6%)

[GBDT]: gradient boosting decision tree; [AdaBoost]: adaptive boosting; [XGBoost]: eXtreme gradient boosting; [AB]: adaptive boosting; [DT]: decision tree; [GNB]: gaussian naive bayes; [KNN]: k-nearest neighbors; [RF]: random forest; [SVM]: support vector machine; [LR]: logistic regression; [ANN]: artificial neural network; [NB]: naive bayes; [BP]: back propagation neural network; [RBF]: radial basis function neural network; [MLP-ANN]: multi-layer perceptron artificial neural network; [LoR]: logistic regression.

Table 3. ML-based thermal comfort studies for the elderly in China.

Paper	Study Setting	Subjects (Sample Size)	Algorithm	Input Parameters	Output Parameter(s)	Generalization Test Method	Performance Metric(s)
[8]	Climate chamber	38 (NA)	GBDT, AdaBoost, XGBoost	Environmental, physiological, and heat exchange parameters	7-point TSV	80% of the dataset for training and 20% for testing	R2, MAE, and MSE
[84]	Field study (summer)	44 (3440)	AB, DT, GNB, KNN, RF, SVM, and XGBoost	Environmental, physiological, and human-related parameters	3-point TSV	80% of the dataset for training and 20% for testing	Precision, recall, accuracy (76%), ROC curve, AUC, and F1 score
[85]	Field study (summer)	14 (1389)	LR, DT, KNN, and SVM	Environmental and physiological parameters	3-point TSV	80% of the dataset for training and 20% for testing	Precision, accuracy (70%), ROC curve, AUC, and F1 score
[86]	Field study (summer)	20 (1865)	AB, RF, LR, ANN, and NB	Environmental parameters	3, 5, 7, and 9-point TSV	80% of the dataset for training and 20% for testing	Recall, accuracy (92.4%), ROC curve, AUC, and F1 score
[87]	Climate chamber	13 (964)	BP and RBF	Environmental and physiological parameters	3-point TSV	80% of the dataset for training and 20% for testing	Accuracy (87.82%)
[88]	Climate chamber	5 (NA)	SVM, RF, KNN, MLP-ANN	Physiological parameters	3-point TSV	10-fold cross-validation	Accuracy (81.2%), F1 score, ROC curve, and AUC
[81]	Field study (year-round)	1040 (724)	RF	Environmental and human-related parameters	3-point TSV	80% of the dataset for training and 20% for testing	Accuracy (56.6%) Accuracy (81.2%)
	Climate chamber	18 (372)	RF	Physiological parameters
[83]	Field study (winter)	35 (135)	ANN, LoR, SVM, KNN, DT, and NB	Environmental and physiological parameters	7-point TSV	80% of the dataset for training and 20% for	Accuracy (67.6%)

[GBDT]: gradient boosting decision tree; [AdaBoost]: adaptive boosting; [XGBoost]: eXtreme gradient boosting; [AB]: adaptive boosting; [DT]: decision tree; [GNB]: gaussian naive bayes; [KNN]: k-nearest neighbors; [RF]: random forest; [SVM]: support vector machine; [LR]: logistic regression; [ANN]: artificial neural network; [NB]: naive bayes; [BP]: back propagation neural network; [RBF]: radial basis function neural network; [MLP-ANN]: multi-layer perceptron artificial neural network; [LoR]: logistic regression.

3.3.3. Thermal Comfort Questionnaires

TSV is the most commonly used subjective method for evaluating thermal comfort. Figure 9 shows the commonly used 3-, 5-, 7-, and 9-point TSV scales, with the 7-point TSV scale being the most widely applied. These scales differ in the level of granularity and practical applicability.

Using the ML method, Zheng et al. [86] evaluated the performance of 3-, 5-, 7-, and 9-point TSV scales among older adults. Results showed that finer scales could capture more detailed differences in thermal sensation. For example, the 7-point TSV scale reflected subtle gradients but faced issues with ambiguous boundaries. In contrast, the 9-point TSV scale accurately described transitions from “slightly warm” to “very hot” but imposed a greater cognitive load [86]. In contrast, despite its simplicity, the 3-point TSV scale was easier for older adults to understand and complete, reducing the risk of response fatigue and inaccurate reporting [86]. This aligns with findings by Chang et al. [11], who noted that longer questionnaires increased impatience and resistance among older participants, potentially affecting data quality.

Zheng et al. [89] found that the 3-point TSV scale had the highest overall correlation with physiological indicators, outperforming the 5-, 7-, and 9-point TSV scales. This suggests that a simpler TSV scale aligns better with the cognitive capacity of older adults and enhances physiological consistency. In addition, the 3-point TSV scale can effectively address the class imbalance problem, especially when ML models struggle to distinguish between adjacent categories [90]. Several studies [81,84] have successfully mitigated this issue by converting 7-point TSV data into 3-point TSV. Moreover, its clarity facilitates integration into intelligent temperature-control systems: as shown in Figure 10, when the system detects an input of +1 (hot) or −1 (cold), it can automatically adjust toward the neutral state (0), enabling a more adaptive and user-friendly temperature control [89].

Besides numerical scales, thermal satisfaction is a practical subjective indicator for comfort assessment. It reflects users’ overall experience more directly and may further reduce cognitive burden compared to the 3-point TSV scale. For instance, a field study in Shanghai by Jiao et al. [63] used thermal satisfaction as the dependent variable (satisfied = 1, dissatisfied = 0). It developed a logistic regression model with individual and health characteristics as predictors. This method, being more intuitive for elderly participants, provided valuable insights for designing age-friendly living environments. However, given the decline in thermal perception among older adults and the potential mismatch between subjective comfort and physiological safety, perceived satisfaction may not necessarily indicate a healthy state [37].

In conclusion, the selection of a TSV scale should be aligned with the purpose and structure of the model. For an ML model, a 3-point TSV scale is generally more appropriate. This choice addresses the cognitive decline and safety concerns of older adults, while also meeting the requirements of automated air conditioning control systems. In contrast, for modified PMV or regression-based models, a 7-point TSV scale is typically more suitable. Using a 3- or 5-point scale reduces granularity and tends to steepen the regression slope, whereas a 9-point scale may increase the cognitive burden and decision-making difficulty for elderly participants.

3.4. Application of Intelligent Technologies in Thermal Environment Regulation

AI technologies have shown great potential in optimizing thermal environments, as they can monitor indoor parameters in real time and automatically adjust temperature-control systems based on thermal comfort models. Temperature control has become a key direction for future research and practice on thermal comfort, with the IoT providing fundamental support for its implementation. Specifically, models can predict how different thermal environments affect thermal sensations. Wireless sensor networks provide continuous year-round monitoring of indoor conditions, overcoming the limitations of intermittent measurements or questionnaires and supporting safer aging in place for older adults [45,91,92]. As an example, Figure 11 illustrates one IoT-based framework that integrates indoor environmental parameters collected by sensors with the PMV model to jointly control the air-conditioning system, achieving intelligent temperature regulation [92]. The feasibility of an IoT-based thermal comfort model framework has been demonstrated, supported by its technical robustness, data-integration capability, interpretability, and practical value [92].

However, the application of ML models in field studies is currently limited, as shown in Table 3, with most studies focusing on a single season, either summer or winter. Only study [81] covers the entire year. Prediction accuracies of the models vary widely, ranging from 56.6% to 92.4%, with study [81] reporting 56.6%, study [83] 67.6%, and all other studies achieving at least 70%. Among these, study [83] had the smallest sample size (135), implying insufficient training data, which may explain its relatively lower prediction accuracy. In contrast, despite a relatively larger sample size (n = 724), study [81] achieved only 56.6% accuracy. This is likely due to its year-round coverage, as seasonal variations significantly affect thermal sensation (as observed in Section 3.1.1), making it difficult for a single model to capture annual patterns. Therefore, to improve the prediction of thermal sensation, employing separate models for each season, together with larger sample sizes, may be a more effective approach.

In addition, relevant practices still face challenges. On the one hand, it is necessary to strengthen data security and privacy protection mechanisms. On the other hand, some elderly individuals have a relatively low level of technology acceptance. Privacy protection can be achieved through both technical and regulatory means. At the technical level, sensitive data can be processed locally at gateways or edge devices, transmitting only aggregated features or model gradients to minimize the risks of data exposure and centralized storage [93]. At the regulatory level, a clear legal framework defining the rights and responsibilities of data subjects is needed to ensure compliance and accountability [94].

4. Discussion

As described above, thermal comfort in older adults is influenced by both environmental and individual factors, with varying degrees of impact. Among thermal comfort models, the aPMV and ML models are the most widely used for studying older adults, but their applicability differs considerably depending on the context. The suitability of questionnaires also varies. This section provides a comparative analysis of these integrated systems to offer a more comprehensive reference for practical applications.

4.1. Influencing Factors and Model Comparison for Older Adults

Based on Section 3.1, Section 3.2 and Section 3.3, Table 4 presents the effects of environmental and individual factors on neutral temperature. The impact of environmental factors varies widely, ranging from about 0.1 °C to nearly 10 °C, whereas the influence of individual factors is relatively small, all under 2.1 °C.

Table 4. Summary of key factors and their effects on neutral temperature.

Factors	Influence on Neutral Temperature (Range)	Trend
Environmental factors
Climate zones and seasons	2–8.8 °C	The winter–summer difference is largest in HSCWZ and smallest in HSWWZ.
Urban–rural differences	0.13–7.98 °C	Larger differences occur in regions with greater disparities in indoor environmental conditions.
Individual factors
Gender	No significant difference	Older women tend to be less tolerant of heat.
Age groups	0.41–1.45 °C	Neutral temperature generally increases with age.
Frailty	2.1 °C	Higher frailty levels are associated with higher neutral temperatures.

Table 4. Summary of key factors and their effects on neutral temperature.

Factors	Influence on Neutral Temperature (Range)	Trend
Environmental factors
Climate zones and seasons	2–8.8 °C	The winter–summer difference is largest in HSCWZ and smallest in HSWWZ.
Urban–rural differences	0.13–7.98 °C	Larger differences occur in regions with greater disparities in indoor environmental conditions.
Individual factors
Gender	No significant difference	Older women tend to be less tolerant of heat.
Age groups	0.41–1.45 °C	Neutral temperature generally increases with age.
Frailty	2.1 °C	Higher frailty levels are associated with higher neutral temperatures.

Figure 12 shows the distribution of studies across climate zones, seasons, and urban versus rural areas. The results show that research is unevenly distributed, mostly concentrated in CZ and HSWZ, with MZ receiving the least attention. Studies in SCZ have mainly focused on winter, while studies in HSWZ have focused on summer. However, with global warming, extreme summer heatwaves in SCZ have increased [95], and winter drops in HSWZ may cause significant cold stress for older adults [96]. Therefore, future research should focus on summer in SCZ and winter in HSWZ and expand studies in rural areas.

Models for thermal comfort vary in terms of input parameters, sample requirements, and applicability. In terms of input parameters, the aPMV includes air temperature, mean radiant temperature, relative humidity, air velocity, clothing insulation, and metabolic rate. Although it is adjusted using the λ coefficient, it still lacks systematic consideration of individual factors such as age, frailty levels, sex, and physiological status. In contrast, ML models offer greater flexibility in input selection, allowing the simultaneous inclusion of environmental factors, individual attributes, and physiological indicators, thereby providing richer predictive features.

Regarding the sample size, the aPMV often relies on large samples (as shown in Table 2). In contrast, most ML models establish predictive relationships by relying on longitudinal data (as shown in Table 3).

Overall, the aPMV model is more suitable as a group-level thermal comfort model for building energy-efficiency design or regional evaluation [16,51], while ML models, which can account for individual attributes and physiological differences, are more appropriate for personalized thermal comfort prediction and intelligent control scenarios [8,83]. Table 5 provides a comparison of their differences.

Table 5. Comparative summary of aPMV and ML models.

Comparison Dimension	aPMV Model	ML Model
Input parameters	Fixed parameters include air temperature, mean radiant temperature, air velocity, clothing insulation, and relative humidity.	Flexible integration of environmental, individual, and physiological parameters.
Sample size	Hard to capture individual differences and relies on large samples.	Effective in capturing individual differences. Accuracy is mostly above 70%.
Output form	Outputs the 7-point TSV scale (−3 to +3), suitable for analyzing thermal sensation and comfort ranges.	Outputs are flexible, with the 3-point TSV scale (−1 to +1) being more suitable.
Application	Suitable for design standards and indoor thermal-comfort evaluation; Suitable for group-based models.	Suitable for intelligent control systems; Suitable for personalized models.

Table 5. Comparative summary of aPMV and ML models.

Comparison Dimension	aPMV Model	ML Model
Input parameters	Fixed parameters include air temperature, mean radiant temperature, air velocity, clothing insulation, and relative humidity.	Flexible integration of environmental, individual, and physiological parameters.
Sample size	Hard to capture individual differences and relies on large samples.	Effective in capturing individual differences. Accuracy is mostly above 70%.
Output form	Outputs the 7-point TSV scale (−3 to +3), suitable for analyzing thermal sensation and comfort ranges.	Outputs are flexible, with the 3-point TSV scale (−1 to +1) being more suitable.
Application	Suitable for design standards and indoor thermal-comfort evaluation; Suitable for group-based models.	Suitable for intelligent control systems; Suitable for personalized models.

4.2. Optimizing ML Models and Intelligent Temperature-Control Systems

Currently, although many ML studies include numerous physiological inputs, such as skin temperature, MST, eardrum temperature, heart rate, and blood pressure, there is still no clear guideline for older adults. Specifically, it remains unclear which parameters are essential and in which conditions they should be applied.

Skin temperatures at different body sites, as shown in Figure 13, correlate with thermal sensation and could serve as key physiological indicators in thermal comfort evaluation [87,97,98]. Mean skin temperature (MST) exhibits a strong positive correlation with operative temperature and TSV, making it a reliable indicator of subjective thermal perception [87,99]. However, MST responds slowly and requires over 50 min to stabilize after cold exposure and approximately 24 min after warm exposure [100,101], which limits its ability to capture rapid thermal changes.

Local skin temperature, particularly at the head, arms, and chest, exhibits notable differences in response to thermal stimuli and strongly influences MST and thermal perception [99]. Overall thermal sensation and comfort are largely determined by the coldest or hottest body part [97]. In warm environments, the head and back are the core regions where older adults exhibit the strongest thermal sensations and the highest weighting coefficients for overall thermal perception [25,97]. Conversely, in cold conditions, older adults’ heating preferences are typically lower than the maximum local heating demand, with the lower limbs often being the coldest body regions [97].

The temperature of the eardrum decreases markedly in response to cold exposure and recovers slowly, but remains below baseline even after returning to neutral conditions [101]. Heart rate tends to decrease during cold exposure [102]. However, the findings on heart rate changes under heat exposure remain inconsistent. One study reported a sharp increase in heart rate [103], while another found no significant change [101]. Blood pressure increases significantly during cold exposure and often remains elevated even after returning to neutral conditions, whereas changes under heat exposure are minimal [101,104].

In conclusion, MST and eardrum temperature recover slowly after cold exposure. Moreover, previous studies report inconsistent findings regarding changes in blood pressure and heart rate. Therefore, forehead and back skin temperatures are recommended for assessing thermal sensation and health in hot environments, while lower-limb temperatures are preferable indicators in cold conditions.

5. Conclusions

This review systematically examines studies on thermal comfort among older adults in China from 2010 to 2025. The key findings are described as follows:

(1)

Environmental and individual factors jointly shape the thermal comfort of older adults, but their impacts on neutral temperature vary considerably. Environmental factors usually exert stronger and more variable effects (as shown in Table 4). Current research remains unevenly distributed across urban and rural areas, climate zones, and seasons. Future research should focus more on summer in the SCZ and winter in the HSWZ, and expand investigations in rural regions.

(2)

Comparative findings indicate that the aPMV model has limited flexibility in capturing individual differences and relies on small-scale datasets, making it more appropriate as a group-level model. In contrast, ML models exhibit clear advantages in identifying individual variability and perform well with longitudinal datasets, making them more appropriate for personalized modeling. Moreover, integrating physiological parameters can further improve their accuracy, and the adoption of the 3-point TSV scale questionnaire facilitates better integration with intelligent temperature-control systems. A more detailed comparison of the two models is presented in Table 5.

(3)

Forehead and back skin temperatures can serve as indicative physiological parameters for predicting thermal sensation and providing health-related early warnings in warm environments, whereas lower-limb skin temperature is more indicative of cold discomfort. However, the real-world application of these physiological indicators remains limited. Future research should further incorporate these parameters into ML models to improve predictive performance and enhance intelligent temperature-control systems.