Optimizing temporal windows for wearable-augmented post-discharge risk prediction: a methods study.

Area of Science:

• Digital Health
• Health Informatics
• Predictive Analytics

Background:

• Traditional readmission risk models use static data, limiting predictive accuracy.
• These models fail to capture patient recovery trajectories post-hospitalization.

Purpose of the Study:

• To identify optimal parameters for dynamically predicting readmission risk.
• To leverage post-discharge step-count data from remote monitoring devices.

Main Methods:

• Combined data from adults aged 55+ from two studies with longitudinal activity data.
• Constructed patient-day datasets with static and dynamic activity features over various retrospective windows (3-10 days).
• Trained logistic regression and LightGBM models to predict readmission/death over prospective horizons (3-10 days) using 5-fold cross-validation.

Main Results:

• LightGBM outperformed logistic regression (AUC 0.82 vs 0.76).
• Model performance improved with longer prospective horizons; insensitive to retrospective window length.
• LightGBM demonstrated good calibration, while logistic regression showed miscalibration.

Conclusions:

• Post-discharge step count data significantly enhance dynamic readmission risk prediction.
• Optimizing temporal windows and employing flexible, non-parametric models like LightGBM improves prediction accuracy and calibration.
• This approach supports more effective post-discharge care management.