Stress prediction using micro-EMA and machine learning during COVID-19 social isolation

Area of Science:

• Psychology
• Data Science
• Public Health

Background:

• Accurate perceived stress prediction is crucial for early intervention, especially during the COVID-19 pandemic, to prevent mental and physical health issues.
• Existing systems require extensive data and offer limited short-term (1-2 days) prediction ranges.

Purpose of the Study:

• To develop an advanced perceived stress prediction system capable of identifying risks up to 7 days earlier.
• To reduce user burden by optimizing micro-EMA question delivery.

Main Methods:

• Utilized history data from micro-ecological momentary assessment (micro-EMA) for stress prediction.
• Implemented an Elastic Net regularization model to select optimal time-series features from micro-EMA responses.
• Employed an ensemble prediction model to forecast perceived stress levels for the subsequent 7 days.

Main Results:

• Achieved a mean absolute error of 4.26 (10.65% of the scale) in predicting 7-day perceived stress scores.
• Demonstrated over 81% accuracy in classifying stress labels for the next 7 days.

Conclusions:

• The proposed system effectively predicts perceived stress 7 days in advance, enabling earlier risk identification.
• Optimized micro-EMA data collection and advanced modeling techniques enhance prediction accuracy and range.