Presentation
DH14 - Leveraging Machine Learning Algorithms to Predict Self-Rated Physical Health: Insights From the All of Us Dataset
SessionPoster Session 1
DescriptionIntroduction:
Aging often brings cognitive and physical declines, such as reduced memory and slower movement, which affect health outcomes. However, the aging process is not homogeneous, as it is influenced by non-chronological factors related to daily activities and behaviors [1]. Lifestyle choices like daily activity levels and sleep habits can impact the aging process. Physiological measurements, such as heart rate, also indicate how aging is progressing. These factors vary widely even among individuals of the same age.
Since each physical measure affects the aging process differently, it is challenging but appealing to quantify the non-chronological aging process using a single unit of measurement beyond just counting years. As an initial step in defining such a unit, we plan to explore self-rated physical health (SRH-P), a widely used subjective measure for perceived health status. SRH-P, a strong predictor of morbidity and premature mortality [2], reflects both medical conditions and individuals' perceptions of their physical abilities and daily function, making it closely linked to non-chronological aging. Establishing the relationship between physical measures and SRH-P is then critical for using SRH-P to develop a unit of measurement for non-chronological aging.
Multiple approaches can be employed to establish the relationship between physical measures and SRH-P, with machine learning (ML) techniques being particularly effective due to their ability to capture complex, non-linear patterns compared to traditional statistical methods. ML techniques such as logistic regression (LR), support vector machine (SVM), and decision tree (DT) have been used for their robustness in health-related predictions.
The goal of this study, therefore, is to lay the groundwork for quantifying non-chronological aging by validating the connection between SRH-P and physical and physiological measures. By identifying key predictors of SRH-P, this research will clarify the connection between objective measurements and subjective health, facilitating more accurate health predictions and enabling personalized interventions.
Approach:
Representative data is essential to ensure unbiased findings between SRH-P levels and physical measures. The All of Us Research Program, led by the National Institutes of Health, with over 838,000 participants, provides such data. The program’s Fitbit data, which captures individuals' physical and physiological measures, includes 15,620 participants, making it the largest known Fitbit dataset accessible to researchers to date. In this study, we explored SRH-P levels and Fitbit-derived features, including activity, heart rate, and sleep metrics from the All of Us Research Program’s Registered Tier Dataset v7.
Fitbit devices track activity metrics, including activity calories, basal metabolic rate (BMR), calories out, elevation, fairly active minutes, floors climbed, lightly active minutes, marginal calories, sedentary minutes, steps, very active minutes, and total steps, reflecting daily physical activity patterns and energy expenditure. Fitbit also tracks heart rate metrics, segmented into four heart rate zones: cardio, fat burn, peak, and out of range, capturing features like minimum and maximum heart rate, time spent in the zone, and calories burned. Sleep data from Fitbit is also analyzed, including features such as total minutes in bed, minutes asleep, minutes awake, minutes restless, and time spent in each sleep stage: deep sleep, light sleep, REM sleep, and wakefulness. For each participant, we calculated the yearly average of the Fitbit-derived features, matching the metrics with the participant’s SRH-P levels. This approach helps us to analyze long-term trends and their association with perceived health status.
To predict physical health levels, we used five machine-learning algorithms: LR, SVM, DT, random forest (RF), and gradient boosting decision tree (GBDT). We applied grid-search to optimize hyperparameters for each model. Cross-validation was used to train and evaluate the models, providing robust performance and reducing the risk of overfitting.
In the All of Us dataset, participants rate their SRH-P levels on a five-point scale: poor, fair, good, very good, excellent. Following the methodology outlined in [3], we created binary classes: positive (very good and excellent) and negative (poor, fair, good). The machine-learning algorithms were evaluated based on the area under the receiver operating characteristic curve (AUC), providing a robust metric for binary classification performance.
Findings:
We present our findings in three parts: individual features, combined features, and model comparison.
Individual feature sets: The activity features yielded the highest predictive performance compared to heart rate and sleep data alone. Across all machine learning models, SVM achieved the best AUC of 0.7484, closely followed by LR (0.7471) and GBDT (0.7468). This suggests that activity-related metrics are the most significant predictors of SRH-P. The heart rate features performed moderately, with GBDT producing the highest AUC (0.6779). This indicates that while heart rate metrics are useful, they are less predictive of SRH-P on their own compared to activity metrics. Sleep data alone had the lowest predictive power, with SVM yielding the highest AUC (0.6116). This indicates that sleep metrics may be less directly correlated with self-reported physical health compared to activity and heart rate data.
Combined feature sets: Combining activity and heart rate features improved model performance across the board, with GBDT achieving the best overall AUC (0.7809), followed by SVM (0.7677) and Logistic Regression (0.7674). The improvement in AUC suggests that combining activity and heart rate features provides the highest performance. Adding sleep features to activity metrics also enhanced model performance, with GBDT producing an AUC of 0.7741, showing a slight decrease compared to the combination of activity and heart rate. The combination of heart rate and sleep data offered a moderate improvement in AUC compared to using heart rate or sleep data alone. GBDT outperformed other models with an AUC of 0.7109, indicating that while useful, these combined features are still less predictive than activity-based combinations. The combination of all three feature sets yielded strong performance across models, with GBDT achieving an AUC of 0.7668, followed by SVM (0.7604). Interestingly, the inclusion of sleep features did not significantly improve the performance beyond the activity and heart rate features combination, suggesting that sleep features may contribute less predictive value in the presence of activity and heart rate data.
Model Comparison: Among the machine-learning algorithms tested, GBDT consistently outperformed other models in most feature combinations, particularly when multiple feature sets were combined (e.g., activity and heart rate features with an AUC of 0.7809). SVM also performed well, particularly with single feature sets like activity and sleep data.
Takeaways:
Most ML models can identify the two SRH modes from physical activities, sleeping habits, and heart rate (AUC up to 78%). Physical activities are important when identifying the two SRH modes (74 ± 3%). Sleeping habits (2%) and heart rate (2%) can slightly improve the performance of ML models. SRH-P can serve as a measure to quantify physical and physiological changes linked to the aging process among older adults. This two-scale binary division of SRH-P (representing "less than good" or "greater than good" health conditions) lays the groundwork for establishing a unit of measurement for non-chronological aging. By categorizing individuals into these two distinct health conditions, we can begin to assess how various lifestyle factors, physical activities, and physiological changes contribute to aging beyond just chronological years, ultimately leading to more precise and personalized methods for evaluating and tracking the aging process.
However, given the limitation of having only a few SRH-P levels as outcomes and a limited range of physical and physiological measures as predictors in the ML models, future research could expand by including more SRH-P levels, integrating additional physical and physiological factors, and evaluating the models' ability to predict changes in SRH-P over time. Incorporating cognitive and social factors, such as memory, attention, and social support, could also provide a more holistic view of self-rated health and non-chronological aging, improving personalized health monitoring and interventions.
Aging often brings cognitive and physical declines, such as reduced memory and slower movement, which affect health outcomes. However, the aging process is not homogeneous, as it is influenced by non-chronological factors related to daily activities and behaviors [1]. Lifestyle choices like daily activity levels and sleep habits can impact the aging process. Physiological measurements, such as heart rate, also indicate how aging is progressing. These factors vary widely even among individuals of the same age.
Since each physical measure affects the aging process differently, it is challenging but appealing to quantify the non-chronological aging process using a single unit of measurement beyond just counting years. As an initial step in defining such a unit, we plan to explore self-rated physical health (SRH-P), a widely used subjective measure for perceived health status. SRH-P, a strong predictor of morbidity and premature mortality [2], reflects both medical conditions and individuals' perceptions of their physical abilities and daily function, making it closely linked to non-chronological aging. Establishing the relationship between physical measures and SRH-P is then critical for using SRH-P to develop a unit of measurement for non-chronological aging.
Multiple approaches can be employed to establish the relationship between physical measures and SRH-P, with machine learning (ML) techniques being particularly effective due to their ability to capture complex, non-linear patterns compared to traditional statistical methods. ML techniques such as logistic regression (LR), support vector machine (SVM), and decision tree (DT) have been used for their robustness in health-related predictions.
The goal of this study, therefore, is to lay the groundwork for quantifying non-chronological aging by validating the connection between SRH-P and physical and physiological measures. By identifying key predictors of SRH-P, this research will clarify the connection between objective measurements and subjective health, facilitating more accurate health predictions and enabling personalized interventions.
Approach:
Representative data is essential to ensure unbiased findings between SRH-P levels and physical measures. The All of Us Research Program, led by the National Institutes of Health, with over 838,000 participants, provides such data. The program’s Fitbit data, which captures individuals' physical and physiological measures, includes 15,620 participants, making it the largest known Fitbit dataset accessible to researchers to date. In this study, we explored SRH-P levels and Fitbit-derived features, including activity, heart rate, and sleep metrics from the All of Us Research Program’s Registered Tier Dataset v7.
Fitbit devices track activity metrics, including activity calories, basal metabolic rate (BMR), calories out, elevation, fairly active minutes, floors climbed, lightly active minutes, marginal calories, sedentary minutes, steps, very active minutes, and total steps, reflecting daily physical activity patterns and energy expenditure. Fitbit also tracks heart rate metrics, segmented into four heart rate zones: cardio, fat burn, peak, and out of range, capturing features like minimum and maximum heart rate, time spent in the zone, and calories burned. Sleep data from Fitbit is also analyzed, including features such as total minutes in bed, minutes asleep, minutes awake, minutes restless, and time spent in each sleep stage: deep sleep, light sleep, REM sleep, and wakefulness. For each participant, we calculated the yearly average of the Fitbit-derived features, matching the metrics with the participant’s SRH-P levels. This approach helps us to analyze long-term trends and their association with perceived health status.
To predict physical health levels, we used five machine-learning algorithms: LR, SVM, DT, random forest (RF), and gradient boosting decision tree (GBDT). We applied grid-search to optimize hyperparameters for each model. Cross-validation was used to train and evaluate the models, providing robust performance and reducing the risk of overfitting.
In the All of Us dataset, participants rate their SRH-P levels on a five-point scale: poor, fair, good, very good, excellent. Following the methodology outlined in [3], we created binary classes: positive (very good and excellent) and negative (poor, fair, good). The machine-learning algorithms were evaluated based on the area under the receiver operating characteristic curve (AUC), providing a robust metric for binary classification performance.
Findings:
We present our findings in three parts: individual features, combined features, and model comparison.
Individual feature sets: The activity features yielded the highest predictive performance compared to heart rate and sleep data alone. Across all machine learning models, SVM achieved the best AUC of 0.7484, closely followed by LR (0.7471) and GBDT (0.7468). This suggests that activity-related metrics are the most significant predictors of SRH-P. The heart rate features performed moderately, with GBDT producing the highest AUC (0.6779). This indicates that while heart rate metrics are useful, they are less predictive of SRH-P on their own compared to activity metrics. Sleep data alone had the lowest predictive power, with SVM yielding the highest AUC (0.6116). This indicates that sleep metrics may be less directly correlated with self-reported physical health compared to activity and heart rate data.
Combined feature sets: Combining activity and heart rate features improved model performance across the board, with GBDT achieving the best overall AUC (0.7809), followed by SVM (0.7677) and Logistic Regression (0.7674). The improvement in AUC suggests that combining activity and heart rate features provides the highest performance. Adding sleep features to activity metrics also enhanced model performance, with GBDT producing an AUC of 0.7741, showing a slight decrease compared to the combination of activity and heart rate. The combination of heart rate and sleep data offered a moderate improvement in AUC compared to using heart rate or sleep data alone. GBDT outperformed other models with an AUC of 0.7109, indicating that while useful, these combined features are still less predictive than activity-based combinations. The combination of all three feature sets yielded strong performance across models, with GBDT achieving an AUC of 0.7668, followed by SVM (0.7604). Interestingly, the inclusion of sleep features did not significantly improve the performance beyond the activity and heart rate features combination, suggesting that sleep features may contribute less predictive value in the presence of activity and heart rate data.
Model Comparison: Among the machine-learning algorithms tested, GBDT consistently outperformed other models in most feature combinations, particularly when multiple feature sets were combined (e.g., activity and heart rate features with an AUC of 0.7809). SVM also performed well, particularly with single feature sets like activity and sleep data.
Takeaways:
Most ML models can identify the two SRH modes from physical activities, sleeping habits, and heart rate (AUC up to 78%). Physical activities are important when identifying the two SRH modes (74 ± 3%). Sleeping habits (2%) and heart rate (2%) can slightly improve the performance of ML models. SRH-P can serve as a measure to quantify physical and physiological changes linked to the aging process among older adults. This two-scale binary division of SRH-P (representing "less than good" or "greater than good" health conditions) lays the groundwork for establishing a unit of measurement for non-chronological aging. By categorizing individuals into these two distinct health conditions, we can begin to assess how various lifestyle factors, physical activities, and physiological changes contribute to aging beyond just chronological years, ultimately leading to more precise and personalized methods for evaluating and tracking the aging process.
However, given the limitation of having only a few SRH-P levels as outcomes and a limited range of physical and physiological measures as predictors in the ML models, future research could expand by including more SRH-P levels, integrating additional physical and physiological factors, and evaluating the models' ability to predict changes in SRH-P over time. Incorporating cognitive and social factors, such as memory, attention, and social support, could also provide a more holistic view of self-rated health and non-chronological aging, improving personalized health monitoring and interventions.
Event Type
Poster Presentation
TimeMonday, March 314:45pm - 6:15pm EDT
LocationFrontenac Foyer
Digital Health (DH)
Simulation and Education (SE)
Hospital Environments (HE)
Medical and Drug Delivery Devices (MDD)
Patient Safety and Research Initiatives (PS)