Presentation
Decoding Mental Health Through Wearable Signals Using Machine Learning Approaches
SessionPoster Session 1
DescriptionIntroduction:
Mental health conditions are common and consequential: more than one in five U.S. adults experience a mental illness annually, and worldwide the burden exceeds one billion people. The impact is also substantial, including increased disability, higher mortality, and major economic costs. In clinical and public health practice, mental health is most often assessed through standardized screening and clinical interviews. Binary risk stratification (e.g., screen-positive vs. screen-negative) offers operational advantages for surveillance and triage. Instruments such as the PHQ-9 (≥10 for moderate depression) and GAD-7 (≥10 for moderate anxiety) align with validated severity thresholds, guiding care pathways and enabling consistent monitoring. Yet limitations remain: self-report screeners are efficient and validated but episodic, subject to recall bias, and prone to false positives, while structured interviews, though diagnostic gold standards, are resource-intensive and time-consuming. These gaps have motivated interest in complementary, data-driven approaches to mental health monitoring.
Recent research emphasizes the value of behavioral and physiological data streams for mental health assessment. Physical activity and energy expenditure reveal movement patterns linked to depression and anxiety risk; sleep duration and quality capture disturbances commonly associated with poor mental health; heart rate and other autonomic measures reflect stress reactivity, whereas sedentary behavior indicates vulnerability. Importantly, these measures can be tracked continuously with wearable devices, opening opportunities for mental health assessment to extend beyond clinical settings into daily life.
More recently, artificial intelligence (AI) techniques, particularly machine learning (ML), have been applied to wearable data to reveal behavioral and physiological patterns that traditional statistical methods may overlook. For example, wearable-based models for depression detection have reported accuracies approaching 0.89, demonstrating promise. However, important challenges remain: many studies rely on small, non-representative samples, device heterogeneity complicates comparability, recruitment often skews toward younger and healthier populations, and most analyses remain cross-sectional, limiting insight into temporal dynamics. Addressing these issues will require larger, more diverse cohorts, standardized protocols, and rigorous external validation before wearable-based ML models can be reliably deployed in real-world care.
This study aims to advance the field by leveraging the All of Us Research Program, a large-scale federally funded initiative linking electronic health records, survey responses, and wearable data from volunteer participants across the U.S. Our work aims to classify mental health status (“positive vs. negative”) using wearable-derived features and to compare performance across canonical machine learning algorithms. In particular, we evaluate the extent to which passively sensed physical activity and energy expenditure can accurately distinguish mental health outcomes and discuss their implications for human factors studies.
Approach:
This study leverages the database from the All of Us Research Program. Among the participants with the Fitbit dataset (15,620 of 838,000 participants, the largest of its kind currently available), two data sources were analyzed: (i) self-rated mental health and (ii) daily physical activity features.
Mental health was assessed with a five-point Likert item (poor–excellent), recorded into binary classes: negative (poor, fair) and positive (good, very good, excellent). The extracted cohort included 837 negative (11.5%) and 6,455 positive (88.5%) cases. To address imbalance, all negative cases were retained and positives were randomly under-sampled to a 1:1.5 ratio, yielding a final dataset of 2,092 participants (837 negative, 1,255 positive).
Physical activity and energy expenditure was represented by ten Fitbit-derived daily measures: (1) activity calories, (2) basal metabolic rate (BMR), (3) calories out, (4) elevation, (5) fairly active minutes, (6) lightly active minutes, (7) marginal calories, (8) sedentary minutes, (9) steps, and (10) very active minutes. Features were extracted for the seven days preceding each survey, capturing temporal dynamics that may reflect fluctuations in activity predictive of mental health. Age was included as a covariate, with results showing that participants with positive mental health were, on average, older than those with negative mental health.
We then trained five supervised machine learning models, namely logistic regression (LR), support vector machines (SVM), decision tree (DT), random forest (RF), and extreme gradient boosting (XGBClassifier). Each participant was represented by seven-day activity sequences (10 × 7 features) plus age. Models were tuned using five-fold cross-validation and evaluated on a held-out test set using area under the ROC curve (AUC), accuracy, precision, recall, and F1 score as the performance metric.
Findings:
The results revealed clear differences in model performance. LR, SVM, and RF achieved comparable levels of discrimination, each reaching an AUC of approximately 0.75. These outcomes indicate that both linear methods (LR and SVM) and an ensemble-based tree method (RF) were able to capture meaningful patterns in the wearable-derived features for distinguishing between positive and negative mental health states. XGBClassifier, another ensemble-based approach, performed slightly better, achieving an AUC of 0.77. Although the improvement over other models was modest, it suggests that gradient boosting can exploit non-linear interactions and complex feature relationships more effectively. In contrast, the DT model lagged behind, with an AUC of 0.63, reflecting weaker discriminative ability and confirming the known tendency of single-tree classifiers to overfit and perform poorly on unseen data.
When evaluating other performance metrics, XGBClassifier again achieved the highest overall performance. It achieved the highest F1 score (0.77) and recall (0.88). This indicates effectiveness in correctly identifying participants with good mental health while maintaining a balance between sensitivity and precision, thus keeping the false positive rate manageable. Such reliability is important in public health and clinical contexts, where accurate stratification supports appropriate allocation of resources and follow-up care. RF also achieved comparable results (F1 score = 0.77 and recall = 0.84) and recorded the highest accuracy (0.71) and precision (0.71). By balancing precision, recall, and F1, it provided a robust alternative to XGBClassifier among the ensemble models. LR and SVM showed nearly identical performance, demonstrating the robustness of linear models, whereas DT showed the weakest accuracy and F1 score, indicating limited generalizability.
Taken together, these findings suggest that ensemble-based methods (RF and XGBClassifier) offer distinct advantages over simpler classifiers by leveraging multiple decision rules and capturing complex, non-linear patterns in the data.
Takeaways From Human Factors Perspectives:
Wearable data combined with ML models offer promising applications in human factors research by enabling continuous, unobtrusive monitoring of mental health in naturalistic settings. Such assessments can support the design of adaptive systems that detect early signs of mental strain and guide interventions in safety-critical domains such as healthcare, transportation, and aviation. In healthcare, daily-life monitoring of activity and energy expenditure may aid in identifying poor mental health symptoms, tracking recovery, and supporting adherence to treatment plans. In transportation, recognizing positive or negative mental health states can help flag risks such as stress, anxiety, or other emotional regulation that impair attention and decision-making, enabling timely interventions to improve road safety.
From a human factors perspective, these findings underscore both opportunities and challenges in integrating wearable-based mental health assessment into practice. Ensemble models such as XGBClassifier and RF demonstrate strong predictive performance, yet their limited interpretability remains a critical consideration for human-centered system design. Transparency and user trust are essential if such models are to meaningfully inform decision-making or guide interventions. Beyond detecting positive mental health states, the capacity to capture true negatives—accurately identifying poor mental health—is equally important to ensure that at-risk individuals are not overlooked. Looking forward, future work could also incorporate additional features such as sleep quality and sedentary behavior, which may capture richer behavioral dynamics, as well as temporal dependencies across weekly cycles or circadian rhythms to model fluctuations more effectively. To advance human-centered systems, further considerations include enhancing interpretability for both clinicians and end users, ensuring representativeness across populations, safeguarding privacy, and designing feedback mechanisms that are supportive rather than stigmatizing.
Mental health conditions are common and consequential: more than one in five U.S. adults experience a mental illness annually, and worldwide the burden exceeds one billion people. The impact is also substantial, including increased disability, higher mortality, and major economic costs. In clinical and public health practice, mental health is most often assessed through standardized screening and clinical interviews. Binary risk stratification (e.g., screen-positive vs. screen-negative) offers operational advantages for surveillance and triage. Instruments such as the PHQ-9 (≥10 for moderate depression) and GAD-7 (≥10 for moderate anxiety) align with validated severity thresholds, guiding care pathways and enabling consistent monitoring. Yet limitations remain: self-report screeners are efficient and validated but episodic, subject to recall bias, and prone to false positives, while structured interviews, though diagnostic gold standards, are resource-intensive and time-consuming. These gaps have motivated interest in complementary, data-driven approaches to mental health monitoring.
Recent research emphasizes the value of behavioral and physiological data streams for mental health assessment. Physical activity and energy expenditure reveal movement patterns linked to depression and anxiety risk; sleep duration and quality capture disturbances commonly associated with poor mental health; heart rate and other autonomic measures reflect stress reactivity, whereas sedentary behavior indicates vulnerability. Importantly, these measures can be tracked continuously with wearable devices, opening opportunities for mental health assessment to extend beyond clinical settings into daily life.
More recently, artificial intelligence (AI) techniques, particularly machine learning (ML), have been applied to wearable data to reveal behavioral and physiological patterns that traditional statistical methods may overlook. For example, wearable-based models for depression detection have reported accuracies approaching 0.89, demonstrating promise. However, important challenges remain: many studies rely on small, non-representative samples, device heterogeneity complicates comparability, recruitment often skews toward younger and healthier populations, and most analyses remain cross-sectional, limiting insight into temporal dynamics. Addressing these issues will require larger, more diverse cohorts, standardized protocols, and rigorous external validation before wearable-based ML models can be reliably deployed in real-world care.
This study aims to advance the field by leveraging the All of Us Research Program, a large-scale federally funded initiative linking electronic health records, survey responses, and wearable data from volunteer participants across the U.S. Our work aims to classify mental health status (“positive vs. negative”) using wearable-derived features and to compare performance across canonical machine learning algorithms. In particular, we evaluate the extent to which passively sensed physical activity and energy expenditure can accurately distinguish mental health outcomes and discuss their implications for human factors studies.
Approach:
This study leverages the database from the All of Us Research Program. Among the participants with the Fitbit dataset (15,620 of 838,000 participants, the largest of its kind currently available), two data sources were analyzed: (i) self-rated mental health and (ii) daily physical activity features.
Mental health was assessed with a five-point Likert item (poor–excellent), recorded into binary classes: negative (poor, fair) and positive (good, very good, excellent). The extracted cohort included 837 negative (11.5%) and 6,455 positive (88.5%) cases. To address imbalance, all negative cases were retained and positives were randomly under-sampled to a 1:1.5 ratio, yielding a final dataset of 2,092 participants (837 negative, 1,255 positive).
Physical activity and energy expenditure was represented by ten Fitbit-derived daily measures: (1) activity calories, (2) basal metabolic rate (BMR), (3) calories out, (4) elevation, (5) fairly active minutes, (6) lightly active minutes, (7) marginal calories, (8) sedentary minutes, (9) steps, and (10) very active minutes. Features were extracted for the seven days preceding each survey, capturing temporal dynamics that may reflect fluctuations in activity predictive of mental health. Age was included as a covariate, with results showing that participants with positive mental health were, on average, older than those with negative mental health.
We then trained five supervised machine learning models, namely logistic regression (LR), support vector machines (SVM), decision tree (DT), random forest (RF), and extreme gradient boosting (XGBClassifier). Each participant was represented by seven-day activity sequences (10 × 7 features) plus age. Models were tuned using five-fold cross-validation and evaluated on a held-out test set using area under the ROC curve (AUC), accuracy, precision, recall, and F1 score as the performance metric.
Findings:
The results revealed clear differences in model performance. LR, SVM, and RF achieved comparable levels of discrimination, each reaching an AUC of approximately 0.75. These outcomes indicate that both linear methods (LR and SVM) and an ensemble-based tree method (RF) were able to capture meaningful patterns in the wearable-derived features for distinguishing between positive and negative mental health states. XGBClassifier, another ensemble-based approach, performed slightly better, achieving an AUC of 0.77. Although the improvement over other models was modest, it suggests that gradient boosting can exploit non-linear interactions and complex feature relationships more effectively. In contrast, the DT model lagged behind, with an AUC of 0.63, reflecting weaker discriminative ability and confirming the known tendency of single-tree classifiers to overfit and perform poorly on unseen data.
When evaluating other performance metrics, XGBClassifier again achieved the highest overall performance. It achieved the highest F1 score (0.77) and recall (0.88). This indicates effectiveness in correctly identifying participants with good mental health while maintaining a balance between sensitivity and precision, thus keeping the false positive rate manageable. Such reliability is important in public health and clinical contexts, where accurate stratification supports appropriate allocation of resources and follow-up care. RF also achieved comparable results (F1 score = 0.77 and recall = 0.84) and recorded the highest accuracy (0.71) and precision (0.71). By balancing precision, recall, and F1, it provided a robust alternative to XGBClassifier among the ensemble models. LR and SVM showed nearly identical performance, demonstrating the robustness of linear models, whereas DT showed the weakest accuracy and F1 score, indicating limited generalizability.
Taken together, these findings suggest that ensemble-based methods (RF and XGBClassifier) offer distinct advantages over simpler classifiers by leveraging multiple decision rules and capturing complex, non-linear patterns in the data.
Takeaways From Human Factors Perspectives:
Wearable data combined with ML models offer promising applications in human factors research by enabling continuous, unobtrusive monitoring of mental health in naturalistic settings. Such assessments can support the design of adaptive systems that detect early signs of mental strain and guide interventions in safety-critical domains such as healthcare, transportation, and aviation. In healthcare, daily-life monitoring of activity and energy expenditure may aid in identifying poor mental health symptoms, tracking recovery, and supporting adherence to treatment plans. In transportation, recognizing positive or negative mental health states can help flag risks such as stress, anxiety, or other emotional regulation that impair attention and decision-making, enabling timely interventions to improve road safety.
From a human factors perspective, these findings underscore both opportunities and challenges in integrating wearable-based mental health assessment into practice. Ensemble models such as XGBClassifier and RF demonstrate strong predictive performance, yet their limited interpretability remains a critical consideration for human-centered system design. Transparency and user trust are essential if such models are to meaningfully inform decision-making or guide interventions. Beyond detecting positive mental health states, the capacity to capture true negatives—accurately identifying poor mental health—is equally important to ensure that at-risk individuals are not overlooked. Looking forward, future work could also incorporate additional features such as sleep quality and sedentary behavior, which may capture richer behavioral dynamics, as well as temporal dependencies across weekly cycles or circadian rhythms to model fluctuations more effectively. To advance human-centered systems, further considerations include enhancing interpretability for both clinicians and end users, ensuring representativeness across populations, safeguarding privacy, and designing feedback mechanisms that are supportive rather than stigmatizing.
Event Type
Poster Presentation
TimeMonday, March 234:45pm - 6:15pm EDT
LocationRhinelander Gallery
Digital Health