Interpretable Emotion Estimation in Indoor Remote Work Environments via Environmental Sensor Data

Abstract

Indoor environmental factors such as CO₂ concentration, temperature, and humidity can significantly influence individuals’ emotional states and productivity. This study continuously collected environmental data using wireless sensors and emotional data from wearable devices in an office-like remote-work setting. Machine learning models, including Random Forest and Gradient Boosting Decision Tree, were developed and interpreted using SHAP (Shapley Additive Explanations). The proposed models achieved estimation accuracies above 85%. SHAP analysis revealed that CO₂ concentration, temperature, and humidity were influential factors in predicting pleasant or unpleasant states. These findings demonstrate the feasibility of real-time, data-driven emotion estimation and provide insights into the design of indoor environments that foster comfort and mental well-being.

1. Introduction

The diversification of modern work styles, particularly the rise of remote work, has highlighted the critical role of indoor environments in affecting a worker’s mental state. Unfavorable environmental conditions can negatively impact productivity and lead to an increase in human error. Numerous studies have shown that indoor factors such as temperature, humidity, lighting, ventilation, and acoustics significantly influence mental health and productivity [1]. For instance, noise and lighting have been reported to affect comfort and work performance among office workers [2], with a series of studies confirming the indoor environment’s impact on well-being [3,4,5,6,7,8,9,10,11]. However, most of this research has relied on questionnaires or controlled laboratory tests, which struggle to capture the complex, dynamic relationship between environmental factors and an individual’s mental state in real-world settings over time.

In the fields of psychology and cognitive science, methods for understanding mental states can be broadly categorized into two groups: those based on subjective self-reports (e.g., questionnaires) and those based on objective physiological data (e.g., heart rate, brain waves, or perspiration). While questionnaires can effectively capture subjective psychological aspects, they often lack objectivity and quantitativeness. In contrast, methods using physiological data provide an objective and quantitative basis for analyzing the relationships between physiological indicators and mental states [12,13,14,15,16].

Previous research has explored the impact of indoor factors, such as elevated CO₂ levels, on human performance [17,18,19] and physiological functions [20] through controlled experiments. However, such studies often rely on indirect methods like self-reported symptoms of sick building syndrome [21] or standard sleepiness scales [22], which may not accurately reflect unconscious mental states. Furthermore, studies conducted in highly controlled climate chambers [23,24,25] have provided foundational insights but faced challenges in capturing the complex interplay of multiple environmental factors that occur in realistic, dynamic settings.

To address these limitations, our study introduces a comprehensive system that extends our previous work [26,27,28]. The system is designed for multimodal, real-time monitoring of both environmental factors and an individual’s emotional state in a realistic indoor environment. We utilize wearable devices to acquire biometric data for emotion estimation and a wireless sensor network to collect environmental data. This approach allows us to overcome the limitations of traditional methods by providing an objective, continuous, and integrated dataset.

A key contribution of this work is the application of interpretable machine learning—specifically, SHAP (SHapley Additive exPlanations)—to analyze the developed machine learning models. Unlike previous black-box models that could only infer mental states from environmental data with high accuracy [26,27,28], SHAP enables us to quantify and interpret the specific influence of each environmental factor on an individual’s emotional state. This allows us to move beyond simple estimation and provide direct, actionable insights into which factors affect pleasantness or unpleasantness.

In summary, this study addresses the gap in existing research by offering a system for the continuous collection of multimodal data in a real-world setting. The primary contributions of this work include the following:

• A novel system for continuous, multimodal data collection on both indoor environmental factors and human emotional states;
• The application of an interpretable machine learning approach (SHAP) to identify and quantify the specific environmental factors that influence a person’s sense of pleasantness or unpleasantness;
• A comprehensive analysis of the dynamic relationship between indoor factors, such as CO₂ concentration, temperature, and humidity, and emotional well-being.

2. Materials and Methods

An overview of this study is illustrated in Figure 1, which is structured into three distinct parts:

Figure 1. Overview of of the sytem proposed in this work. (Source: authors’ own study.)

• Data Acquisition: Various indoor environmental data related to human cognitive functions and mental states (pleasantness/unpleasantness) are acquired by edge computersvia wireless sensor networks. The data are then segmented into training and test datasets for the machine learning model.
• Model Construction and Evaluation: Utilizing the training data, a machine learning model is developed, where the environmental data serve as the input and the mental state (pleasantness/unpleasantness) serves as the output. Post construction, the accuracy of the model is assessed using the test data.
• SHAP Analysis: SHAP is employed to construct a regression model from the training model. This model aims to determine which indoor environmental values impact emotions (pleasantness/unpleasantness) in the original machine learning model, as represented by SHAP values, thereby adding interpretability to the model.

The novelty of this study lies in its unified approach, which continuously collects environmental and emotional data from multiple sensors. By applying SHAP to analyze the machine learning model, this research objectively and interpretably reveals the relationships between dynamic environmental factors and individual emotional states in an indoor work environment. This approach contributes to the quantitative clarification of key influencing factors, a feat that goes beyond the limitations of conventional subjective questionnaires or one-off experiments, which often fail to capture the continuous, nuanced, and multifaceted nature of these relationships.

2.1. Environmental Sensor

The multiple types of environmental data (units) relevant to human cognitive functions used in constructing the estimation model are listed in Table 1, along with the number of sensors deployed.

Table 1. Specifications of the environmental sensors used in the experiment.

Environmental Data	Sensor (Model, Principle)
Temperature (°C)	DHT11—Measures temperature
Humidity (%)	via resistance change and humidity via capacitance change.
CO2 concentration (ppm)	MH-Z19C—Detects CO2 by infrared absorption at 4.26 $\mathsf{\mu }$m.
Illuminance (lux)	TSL25721—Photodiodes sensitive to visible + IR and IR only.
Visible light (output, 0–1023)
Infrared light rays (output, 0–1023)
Full light rays (visible light + infrared rays)
Intensity of blue light (output, 0–1023)	CDS393-B—CdS cell with blue filter; resistance varies with light intensity.
Sound level (output, 0–1023)	DFR0034—Vibration of electrodes alters capacitance.
Smell (output, 0–1023)	TGS2450—Conductivity change in response to sulfur compounds.
Ultrasonic distance (cm)	HC-SR04—Distance measured by ultrasonic reflection time.
Barometric pressure (hPa)	BME280—Thin-film resistance change due to pressure deformation.

The following measures were implemented to ensure reliable sensor data acquisition. A wireless sensor network was configured in a star topology using Xbee, enabling real-time data collection at 10 s intervals. In parallel, the NEC Emotion Analysis Solutions system [29] was used to ensure consistency in measurements of mental states.

2.2. Emotional Recognition

To measure participants’ emotions, the “Emotion Analysis Solutions” systemof NEC [29] was utilized. This system interprets pulse and barometric data from wearable devices to visualize mental states [30,31], allowing for the determination of subjects’ arousal levels and valence values from heart-rate variation data. The study integrates “Silmee TMW22”, a wristband-type wearable device, with a cloud-based analysis platform featuring emotion recognition technology, enabling the real-time collection and storage of subjects’ barometric data. In this research, “Silmee TMW22” is employed for barometric data acquisition, and the valence values (pleasantness/unpleasantness) derived from the “Emotion Snalysis Solutions” system of NEC are utilized as the emotional data.

2.3. Model Building

In this study, machine learning techniques are employed to develop an emotion estimation model. Specifically, random forest (RF) and gradient boosting decision tree (GBDT) are utilized as machine learning models, with all parameters, except those stated, maintained at their default settings in the respective libraries. RF and GBDT were selected due to their superior estimation performance. Furthermore, these tree-based methods are particularly compatible with SHAP analysis, enabling an intuitive interpretation of feature contributions.

2.3.1. RF

RF operates by constructing multiple decision trees based on training data and making predictions of new data (test data) by averaging the outcomes of each decision tree. For this experiment, we used sklearn, an open-source Python 3.13 library. The RF parameters determined through the grid search are listed in Table 2. In particular, max_depth represents the depth of the decision trees, whereas n_estimator indicates their quantity.

Table 2. RF’s Parameter settings.

Parameter	Value
random_state	42
max_depth	10
n_estimator	30

2.3.2. GBDT

GBDT is an algorithm that generates multiple decision trees from training data, sequentially corrects errors in these trees, and iteratively builds a model. This experiment employed LightGBM, another open-source Python library. The GBDT parameters obtained from grid search are outlined in Table 3.

Table 3. GBDT’s parameter settings.

Parameter	Value
num_boost_round	100
early_stopping_rounds	10

2.4. SHAP

The proposed system analyzes the developed machine learning model using SHAP (SHapley Additive exPlanations), as introduced by Lundberg et al. at NIPS 2017 [32]. SHAP is a game theory-based method that provides interpretable insights into machine learning model outputs through additive feature attribution. In this method, the output of the model is expressed as the sum of the contributions of each feature, thereby allowing for an understanding of how each feature influences the prediction.

Given an input vector ($x=(x_1,x_2,\dots ,x_a)$) and a model ($f\left(x\right)$), the SHAP framework approximates an output with an interpretable model ($g\left(x^{\prime }\right)$) using a simplified input ($x^{\prime }$), expressed as follows:

$$f\left(x\right)=g\left(x^{\prime }\right)={\varphi }_0+\sum _{i=1}^M{\varphi }_i{x}_i^{\prime }$$

(1)

where ${\varphi }_i$ represents the contribution of each feature to the prediction. These SHAP values are uniquely determined based on the desirable properties of local accuracy, missingness, and consistency.

In this study, we apply TreeSHAP, a variant of SHAP optimized for tree-based models such as Random Forest (RF) and Gradient Boosting Decision Trees (GBDT), which are used in our system. This enables us to estimate the influence of indoor environmental features on individuals’ emotional states (pleasantness or unpleasantness).

By examining the SHAP-value distributions corresponding to environmental features, we identify which features have the greatest impact on mental state and approximate the threshold values where their effect becomes significant.

For detailed mathematical formulation of SHAP, including its derivation from cooperative game theory and full expression of the Shapley value calculation, please refer to Appendix A.

2.5. Experiment

2.5.1. Participants and Experiment Design

This study was conducted in a residential house within the Chiba University experimental facility. A total of ten healthy, non-smoking participants (five couples), aged 20 and over, were recruited to reside in the facility for 3 nights and 4 days. Participants were selected based on a strict set of inclusion and exclusion criteria to ensure a representative and controlled sample for the study.

• Inclusion Criteria:
  • A healthy couple or partnershipaged 20 or older.
  • Available for a 3-night, 4-day stay at the experimental house at Chiba University.
  • Capable of working indoors (remote working) during the experiment (participants or their partners could also be registered as participants).
  • Low sensitivity to chemical substances, as determined by the Quick Environmental Exposure and Sensitivity Inventory (QEESI) questionnaire.
  • No prior history of symptoms related to sick building syndrome, as determined by the questionnaire.
  • No known allergies.
• Exclusion Criteria:
  • A history of smoking (if either partner had a history of smoking, the couple was excluded).
  • A diagnosis of chemical sensitivity (if either partner had a diagnosis, the couple was excluded).

Prior to the experiment, all participants provided written informed consent. All procedures were approved by the Institutional Review Board of Chiba University (Receipt No: M10032).

2.5.2. Experimental Environment and Data Collection

This study utilized a hybrid system combining wearable devices and wireless sensor networks to continuously collect both personal vital data and indoor environmental data. This setup was designed to capture the dynamic interplay between a participant’s emotional state and the surrounding environment during extended indoor work periods. Figure 2 displays the layout of this facility, with red circles marking the locations of the wireless environmental sensors in each room. These sensors collected 12 types of indoor environmental data, as detailed in Table 1 and Table 4.

Figure 2. Floor plan of the experimental house, which includes a bedroom, a working room, and a living room. (Source: authors’ own study.)

Table 4. Concrete dataset of the environmental feature data we used for model construction.

Time	Temp.	Hum.	Light (Inf.)	Light (Full)	Light (Vis.)	Ill.	Blue	Soumd	Smell	Ultra.	Barometric	CO2	Emotion
09:18:20	23.4	61	83	319	236	719	841	43	1017	171.9	1010	432	HAPPY
09:18:30	23.4	62	83	322	239	733	847	47	1016	173.16	1010	432	HAPPY
09:18:40	23.4	61	83	322	239	733	847	48	1017	174.08	1010	432	HAPPY
09:18:50	23.3	61	84	323	239	729	852	59	1017	172.75	1010	432	HAPPY
09:19:00	23.4	61	84	323	239	729	848	44	1018	173.2	1010	431	HAPPY

For personal emotional data, we used the Silmee TMW22 wearable device, which estimates a person’s emotional state based on fluctuations in their pulse and skin temperature. A central server was used to collect, manage, and synchronize all sensor and wearable data. To facilitate data matching, participants were required to fill out a behavior chart, which was used to correlate their location and activities within the facility with the time-stamped sensor data. Data collection was paused during periods of sleep, outings, and bathing to focus on emotional states during active, indoor periods.

3. Results

3.1. Model Development and Validation Method

For both GBDT and RF, models were constructed using training data labeled with the subjects’ emotions and the 12 types of indoor environmental data detailed in Table 1. Emotional labeling was primarily conducted using the Emotions Analysis Solutions method [29], which estimates four types of emotion (HAPPY, STRESSED, RELAXED, and SAD) based on fluctuations in pulse and skin temperature.

The dataset, as a whole, may not be strictly time-series data; however, because it includes time-series components, the potential temporal dependencies must be considered in the validation method. In this study, the dataset is divided into segments of fixed length (e.g., groups of five consecutive samples), and these segments serve as the unit of partitioning. Specifically, 80% of the segments are allocated to the training set, while the remaining 20% are assigned to the test set. Crucially, the chronological order of the data within and across segments is maintained to ensure that temporal neighbors of the test data are excluded from the training set.

3.2. Emotion Estimation

Table 5 and Table 6 present the accuracy of the emotion estimation models for each participant ID, including precision, recall, F1 score, accuracy, and the number of samples. While the primary evaluation focuses on test data performance, results for both training and testing are included to assess potential overfitting.

Table 5. Detailed performance of the Random Forest model. This table presents a comprehensive analysis of the Random Forest model’s classification performance on both the training and test datasets. Results are disaggregated for each participant (ID1–ID6) and include key metrics such as precision, recall, F1 score, and accuracy. The table also specifies the number of samples for each emotion class (Pleasure, Unpleasantness) to provide context for the performance metrics. While a high average training accuracy of 95.2% indicates strong learning, the robust test accuracy of 87.8% confirms the model’s generalization capability.

Train						Test
ID1		Precision	Recall	F1 Score	Number	ID1		Precision	Recall	F1 Score	Number
	Pleasure	0.94	0.80	0.87	1353		Pleasure	0.84	0.69	0.76	296
	Unpleasure	0.92	0.98	0.95	3187		Unpleasure	0.90	0.95	0.92	844
	Macro avg	0.93	0.89	0.91	4540		Macro avg	0.87	0.82	0.84	1140
	Accuracy		0.93				Accuracy		0.88
ID2		Precision	Recall	F1 Score	Number	ID2		Precision	Recall	F1 Score	Number
	Pleasure	0.94	0.99	0.97	3507		Pleasure	0.87	0.94	0.90	853
	Unpleasure	0.97	0.81	0.88	1138		Unpleasure	0.79	0.61	0.69	312
	Macro avg	0.96	0.90	0.93	4645		Macro avg	0.83	0.77	0.79	1165
	Accuracy		0.95				Accuracy		0.85
ID3		Precision	Recall	F1 Score	Number	ID3		Precision	Recall	F1 Score	Number
	Pleasure	0.97	0.95	0.96	1972		Pleasure	0.93	0.86	0.90	493
	Unpleasure	0.97	0.98	0.98	3548		Unpleasure	0.93	0.96	0.95	887
	Macro avg	0.97	0.97	0.97	5520		Macro avg	0.93	0.91	0.92	1380
	Accuracy		0.97				Accuracy		0.93
ID4		Precision	Recall	F1 Score	Number	ID4		Precision	Recall	F1 Score	Number
	Pleasure	0.94	0.89	0.92	1150		Pleasure	0.75	0.68	0.72	256
	Unpleasure	0.96	0.98	0.97	3230		Unpleasure	0.91	0.93	0.92	844
	Macro avg	0.95	0.94	0.94	4380		Macro avg	0.83	0.81	0.82	1100
	Accuracy		0.96				Accuracy		0.87
ID5		Precision	Recall	F1 Score	Number	ID5		Precision	Recall	F1 Score	Number
	Pleasure	0.96	0.95	0.96	2496		Pleasure	0.92	0.92	0.92	685
	Unpleasure	0.93	0.93	0.93	1564		Unpleasure	0.83	0.84	0.83	335
	Macro avg	0.94	0.94	0.94	4060		Macro avg	0.88	0.88	0.88	1020
	Accuracy		0.95				Accuracy		0.89
ID6		Precision	Recall	F1 Score	Number	ID6		Precision	Recall	F1 Score	Number
	Pleasure	0.95	0.97	0.96	2528		Pleasure	0.86	0.91	0.88	655
	Unpleasure	0.94	0.92	0.93	1552		Unpleasure	0.82	0.74	0.77	370
	Macro avg	0.95	0.94	0.94	4080		Macro avg	0.84	0.82	0.83	1025
	Accuracy		0.95				Accuracy		0.85

Table 6. Detailed performance of the GBDT model. This table provides a comprehensive overview of the GBDT model’s performance, mirroring the structure of RF. The results, shown for each participant (ID1–ID6), are based on the same key metrics (precision, recall, F1 score, and accuracy) and sample counts. The GBDT model achieved a high average training accuracy of 95.3% and a strong test accuracy of 86.5%, demonstrating its effectiveness in distinguishing between pleasant and unpleasant emotional states.

Train						Test
ID1		Precision	Recall	F1 Score	Number	ID1		precision	Recall	F1 Score	Number
	Pleasure	0.91	0.83	0.87	1288		Pleasure	0.75	0.65	0.70	361
	Unpleasure	0.94	0.97	0.95	3252		Unpleasure	0.85	0.90	0.87	779
	Macro avg	0.92	0.90	0.91	4540		Macro avg	0.80	0.77	0.78	1140
	Accuracy		0.93				Accuracy		0.82
ID2		precision	Recall	F1 Score	Number	ID2		Precision	Recall	F1 Score	Number
	Pleasure	0.96	0.99	0.98	3488		Pleasure	0.92	0.95	0.93	872
	Unpleasure	0.97	0.89	0.93	1157		Unpleasure	0.84	0.74	0.79	293
	Macro avg	0.96	0.94	0.95	4645		Macro avg	0.88	0.85	0.86	1165
	Accuracy		0.96				Accuracy		0.90
ID3		Precision	Recall	F1 Score	Number	ID3		Precision	Recall	F1 Score	Number
	Pleasure	0.97	0.97	0.97	1998		Pleasure	0.87	0.85	0.86	467
	Unpleasure	0.98	0.98	0.98	3522		Unpleasure	0.92	0.94	0.93	913
	Macro avg	0.98	0.98	0.98	5520		Macro avg	0.90	0.89	0.89	1380
	Accuracy		0.98				Accuracy		0.91
ID4		Precision	Recall	F1 Score	Number	ID4		Precision	Recall	F1 Score	Number
	Pleasure	0.94	0.84	0.89	1069		Pleasure	0.82	0.60	0.69	337
	Unpleasure	0.95	0.98	0.97	3311		Unpleasure	0.84	0.94	0.89	763
	Macro avg	0.94	0.91	0.93	4380		Macro avg	0.83	0.77	0.79	1100
	Accuracy		0.95				Accuracy		0.84
ID5		Precision	Recall	F1 Score	Number	ID5		Precision	Recall	F1 Score	Number
	Pleasure	0.96	0.96	0.96	2613		Pleasure	0.85	0.88	0.87	568
	Unpleasure	0.92	0.93	0.93	1447		Unpleasure	0.85	0.80	0.82	452
	Macro avg	0.94	0.94	0.94	4060		Macro avg	0.85	0.84	0.85	1020
	Accuracy		0.95				Accuracy		0.85
ID6		Precision	Recall	F1 Score	Number	ID6		Precision	Recall	F1 Score	Number
	Pleasure	0.96	0.96	0.96	2508		Pleasure	0.91	0.90	0.90	675
	Unpleasure	0.93	0.93	0.93	1572		Unpleasure	0.81	0.83	0.82	350
	Macro avg	0.94	0.95	0.95	4080		Macro avg	0.86	0.86	0.86	1025
	Accuracy		0.95				Accuracy		0.87

The average accuracy of the training data was 95.2% for RF and 95.3% for GBDT, whereas the average accuracy of the test data was 87.8% for RF and 86.5% for GBDT. Despite potential overfitting, both models demonstrated high accuracy as personal emotion estimation models. Furthermore, the precision, recall, and F1 score of the test data were approximately 0.90, signifying the requirement of high reliability.

3.3. Identification of Critical Environmental Factors Using SHAP

Figure 3 and Figure 4 illustrate the distribution of SHAP values for each environmental factor across the two models. Each subgraph (a–f) corresponds to a participant (ID1–ID6). The horizontal axis represents SHAP values: larger positive values indicate a stronger contribution toward unpleasant states, while negative values indicate contributions toward pleasant states. The color coding (red for high, blue for low) represents the magnitude of each environmental variable. The order of features, sorted by average absolute SHAP values, reflects their overall importance.

Figure 3. SHAP value distributions for Random Forest models across participants: (a) ID1; (b) ID2; (c) ID3; (d) ID4; (e) ID5; (f) ID6. Each point represents an environmental factor (red = higher values, blue = lower values). Positive SHAP values indicate stronger contributions toward unpleasant emotional states, while negative values indicate contributions toward pleasant states. CO₂ concentration and temperature consistently show strong influences across participants.

Figure 4. SHAP value distributions for GBDT models across participants: (a) ID1; (b) ID2; (c) ID3; (d) ID4; (e) ID5; (f) ID6. Interpretation is analogous to Figure 3. While CO₂ concentration and temperature remain key factors, humidity displays stronger inter-individual variation compared to Random Forest results.

3.3.1. Key Environmental Factors

Table 7 summarizes the frequency with which each factor ranked among the top four in terms of SHAP importance. CO₂ concentration and temperature consistently ranked highest across participants and models, indicating their dominant influence. Humidity and barometric pressure were also found to be significant but exhibited stronger individual variation. Factors such as light intensity, odor, and sound showed weaker or more inconsistent associations.

Table 7. Frequency of top-four SHAP-ranked features for RF and GBDT. Counts represent how many times each environmental factor appeared in the top four across IDs 1–6. CO₂ and temperature were consistently dominant, while humidity and barometric pressure had a moderate influence and light and odor exhibited weaker or individual-dependent effects.

Random Forest		GBDT
Parameter	Count	Parameter	Count
CO2 Concentration	6	CO2 Concentration Temperature	6
Temperature	5
Barometric Pressure	4	Humidity	5
Humidity	3	Barometric Pressure Blue Light	2
Odor	2
Blue Light Full Ray Infrared Ray Ultrasonic Distance	1	Odor Infrared Ray Ultrasonic Distance	1
Visible Ray Sound level Illuminance	0	Visible Ray Full Ray Sound level Illuminance	0

3.3.2. CO₂ Concentration

Figure 5 and Figure 6 show the relationship between SHAP values and CO₂ levels. Across most participants, SHAP values were minimal, ranging between 400 and 600 ppm, suggesting a stable emotional state. Above 600 ppm, SHAP values increased, indicating a tendency toward negative emotions. This implies that lower CO₂ levels support psychological comfort, whereas elevated levels may impair well-being.

Figure 5. Relationship between CO₂ concentration and SHAP values for (a–f) participants ID1–ID6 using Random Forest. Stable emotional states are observed at 400–600 ppm (low SHAP values), while higher concentrations (≥600 ppm) correlate with increased SHAP values, suggesting negative emotional impact.

Figure 6. Relationship between CO₂ concentration and SHAP values for (a–f) participants ID1–ID6 using GBDT. Similar to the RF model, CO₂ causes stable and negative emotional states.

3.3.3. Temperature

Figure 7 and Figure 8 demonstrate the association between temperature and SHAP values. Stable, near-zero SHAP values were observed within the 23–26 °C range, indicating optimal comfort. Beyond this range, particularly above 26 °C, SHAP values increased, suggesting potential negative effects on emotional states. This threshold is consistent with prior findings on thermal comfort.

Figure 7. Relationship between temperature and SHAP values for (a–f) participants ID1–ID6 using Random Forest. SHAP values are stable near zero between 23 and 26 °C, indicating optimal comfort. Beyond 26 °C, SHAP values increase, reflecting a tendency toward discomfort.

Figure 8. Relationship between temperature and SHAP values for (a–f) participants ID1–ID6 using GBDT. Similar to the RF model, temperature causes comfort and uncomfortable states.

3.3.4. Humidity

Figure 9 and Figure 10 present the humidity–SHAP relationship. Unlike CO₂ and temperature, no consistent trend was observed, and individual differences were pronounced. Some participants exhibited high SHAP values at low humidity (dry conditions), whereas others showed sensitivity to higher humidity. This variation suggests that while humidity is a relevant factor, its psychological impact is highly dependent on the individual.

Figure 9. Relationship between humidity and SHAP values for (a–f) participants ID1–ID6 using Random Forest. Unlike CO₂ and temperature, no consistent trend is observed. Individual differences dominate, with some participants sensitive to low humidity (dry conditions) and others to high humidity.

Figure 10. Relationship between humidity and SHAP values for (a–f) participants ID1–ID6 using Random Forest. Similar to RF, individual differences dominate.

4. Discussion

4.1. Key Findings and Comparison with Prior Work

Our findings confirm that indoor environmental factors, particularly CO₂ concentration, temperature, and humidity, significantly influence a person’s mental state. While previous studies [25] have explored these relationships under highly controlled conditions, our work extends this research by developing a system for continuous, real-time monitoring in a more realistic remote work environment. This approach allowed us to capture the dynamic interplay of multiple factors that are challenging to measure with traditional methods like questionnaires and one-off experiments.

The models developed in this study achieved high estimation accuracy (>85% for both Random Forest and GBDT), validating our approach. More importantly, the SHAP analysis provided transparent insights into the decision-making process of these black-box models. This is a significant contribution, as it moves beyond simple prediction to provide a causality-like understanding of how specific environmental changes influence emotional well-being.

These findings underscore the strength of SHAP-based interpretation for clarifying the contribution of environmental parameters to emotional states. Conventional machine learning models typically offer only relative feature importance, making it difficult to quantify contributions or interpret correlations. In contrast, SHAP enables a more interpretable, participant-specific understanding of how indoor environmental conditions—particularly CO₂, temperature, and humidity—influence mental states.

Our results are consistent with prior studies on environmental factors and emotions [20,33,34,35,36,37,38,39,40]. For example, Bombail demonstrated that odors induce stress responses [35], while Shin et al. showed that lighting conditions influence emotions and brain activity [40]. Choi et al. [41] investigated how changes in indoor environmental quality affect task performance, whereas our study directly links environmental parameters to emotional states. By enabling multimodal, real-time monitoring, our approach provides a practical framework for designing adaptive work environments that support psychological well-being.

4.2. Interpretation of Key Findings

Our analysis, supported by SHAP values, showed that both CO₂ concentration and temperature have a consistent and predictable impact on mental state across most participants. Specifically, higher CO₂ concentrations and non-optimal temperatures were strongly correlated with predictions of “Unpleasantness.” This is consistent with physiological findings that high CO₂ levels can reduce cognitive performance [19] and evoke physical symptoms, which, in turn, affect mood.

In contrast, the influence of humidity was found to be highly inconsistent, with no clear trend across all participants. As demonstrated in Figure 9 and Figure 10, some individuals showed a negative correlation with high humidity, while others were more sensitive to low humidity. This suggests that humidity’s impact is highly subjective and may depend on individual physiological differences or personal preferences. This finding highlights a critical limitation of one-size-fits-all environmental guidelines and underscores the necessity of personalized environmental control systems.

4.3. Limitations and Future Work

This study is not without limitations. The sample size of six participants, while sufficient for a detailed case study, limits the generalizability of our findings. Future research should aim to validate these results with a larger and more diverse cohort. Furthermore, while we simplified emotional states to two categories (Pleasantness and Unpleasantness), incorporating a wider range of emotional states, such as stress or relaxation, could provide a more nuanced understanding.

In the future, we plan to deploy our system in a wider range of remote work and office environments. This will allow us to gather a more comprehensive dataset and develop models that can adapt to various building designs and occupant behaviors. We also aim to develop a real-time feedback system that can recommend personalized environmental adjustments to improve comfort and mental well-being.

5. Conclusions

This study introduced a system designed to analyze the impact of environmental data on individual pleasantness or unpleasantness. Through experimentation, we developed a model to estimate an individual’s mental state from indoor environmental data. By applying SHAP to approximate this model, we identified and analyzed the environmental factors that influence an individual’s sense of pleasantness or unpleasantness.

The experimental outcomes affirmed that each environmental factor plays a pivotal role in understanding the correlation and contribution to an individual’s mental state. These insights are of substantial importance for enhancing mental health and fostering comfortable working environments.

The study, while insightful, is limited by a small sample size and potential overfitting in machine learning models, indicating the need for broader participant samples in future research. Further exploration of individual environmental sensitivities and application across various settings would enhance the model’s universality.

Despite these limitations, the research significantly contributes to understanding of the impact of indoor environments on mental health, offering practical implications for the design of healthier, more productive spaces and underscoring its relevance in environmental psychology and cognitive science.

Our study advances the field by demonstrating the feasibility of continuously estimating mental states through an environmental monitoring system that captures multimodal environmental data. Unlike conventional approaches, which lack interpretability in understanding the relationship between environmental factors and human mental states and fail to identify the specific parameters involved, our system integrates machine learning and SHAP-based analysis to deliver real-time, interpretable insights into occupant well-being.

Based on our findings, it is recommended that indoor environments incorporate adaptive ventilation and temperature regulation systems to enhance occupants’ mental states.

Furthermore, given the limitations of our study, such as limited sample diversity, future work should involve research conducted in more diverse settings and over extended monitoring periods.