Comparing training window selection methods for prediction in non-stationary time series.

Area of Science:

• Digital Health
• Computational Psychiatry
• Behavioral Data Science

Background:

• Smartphones offer passive monitoring of daily behavior via sensors.
• Sensor data correlates with psychological symptoms and mood, potentially reducing measurement burden.
• Predicting psychopathology peaks could enable timely interventions in mental healthcare.

Purpose of the Study:

• To investigate optimal window sizes for training models using sensor data to predict psychological symptoms.
• To compare various methodologies for selecting appropriate training window sizes.
• To evaluate the impact of varying rates of change in the sensor-symptom relationship on prediction accuracy.

Main Methods:

• A simulation study was conducted, varying the rate of change in the underlying relationship between sensor data and psychological symptoms.
• Different window size selection methodologies were compared, including heuristic and super learning approaches.
• The predictive performance of averaging predictions across multiple windows versus selecting a single best window was assessed.

Main Results:

• Selecting a single optimal training window can be detrimental to prediction accuracy, especially with time-varying relationships.
• Averaging predictions across different window sizes consistently reduced prediction error.
• The proposed averaging approach demonstrated effectiveness on both simulated and real-world smartphone sensor data.

Conclusions:

• The assumption of a constant or fixed-rate relationship between sensor data and psychological symptoms is often invalid.
• Averaging predictions across multiple time windows is a robust strategy to improve the accuracy of mental health monitoring using sensor data.
• This approach enhances the potential for digital mental healthcare by providing more reliable predictions of psychological states.