Kernel Density Estimation of Wearable Signals to Predict Preoperative Cancellation Risk.

Area of Science:

• Digital Health
• Machine Learning
• Wearable Technology

Background:

• Consumer wearables and machine learning offer scalable methods for pre-symptomatic acute illness detection.
• Previous studies demonstrate wearable data can identify infections before symptom onset or in asymptomatic individuals.
• Elective surgery cancellations due to unexpected patient illness cause significant resource underutilization.

Purpose of the Study:

• To develop and evaluate a prototype preoperative alert system using a kernel density estimation (KDE)-based illness prediction model applied to consumer wearable data.
• To assess the model's capability in flagging impending illness in advance using a publicly available longitudinal dataset.
• To investigate the causes of false positives and propose recommendations for future model development.

Main Methods:

• Developed a prototype preoperative alert system utilizing a kernel density estimation (KDE) illness prediction model.
• Applied the KDE model to consumer wearable data from a publicly available longitudinal dataset.
• Evaluated the model's performance in detecting pre-symptomatic illness and analyzed false-positive rates.

Main Results:

• The KDE-based illness prediction model successfully detected most sick users.
• A high false-positive rate was observed, indicating challenges with the current approach.
• Potential causes for false positives identified include physiological variability, activity confounding, data sparsity, and threshold calibration.

Conclusions:

• Presymptomatic illness prediction using consumer wearables shows promise for improving surgical scheduling by minimizing cancellations.
• Recommendations for future models include personalization, multimodal feature integration, robust anomaly scoring, temporal smoothing, and prospective validation.
• Actionable steps are outlined to reduce false positives while maintaining sensitivity in illness detection for perioperative settings.