Individualized Modeling to Distinguish Between High and Low Arousal States Using Physiological Data

Area of Science:

• Digital Health
• Machine Learning in Psychology
• Wearable Sensor Technology

Background:

• Continuous physiological data collection via wearable sensors is increasingly feasible.
• Smartphones enable self-reported psychological state data but increase participant burden.
• Machine learning (ML) can potentially distill physiological data into insights on psychological states.

Purpose of the Study:

• To investigate ML approaches for predicting psychological states from passive physiological data.
• To compare ML classifiers for affective arousal state prediction.
• To evaluate the efficacy of individual-specific versus general predictive models.

Main Methods:

• Features were extracted from physiological data using the tsfresh Python package.
• A k-nearest neighbor classifier with dynamic time warping and a random forests classifier were employed.
• Individual-specific and general random forests models were developed to predict affective arousal states.

Main Results:

• Both k-nearest neighbor and random forests classifiers were effective in predicting affective arousal states.
• Individual-specific predictive models demonstrated superior performance compared to general models.
• The study supports the feasibility of using passively collected wearable data for psychological state prediction.

Conclusions:

• Passively collected wearable sensor data can be utilized to predict psychological states.
• Individualized ML models enhance prediction accuracy for psychological states.
• Integrating passive data collection can reduce or eliminate the need for active self-report assessments in mental health monitoring.