Kernel Density Estimation of Wearable Signals to Predict Preoperative Cancellation Risk

Digital health approaches that leverage consumer wearables and machine learning offer scalable means to detect acute illness pre-symptomatically, enabling earlier intervention and improved resource planning. Prior studies have shown that wearable-derived signals can identify infections-including COVID-19, malaria, H1N1, and rhinovirus-before symptom onset or in asymptomatic cases. Elective surgery pathways remain vulnerable to day-of-surgery cancellations due to unexpected patient illness, leading to underutilized operating rooms, equipment, and staff. To address this, we developed a prototype preoperative alert system that applies a kernel density estimation (KDE)-based illness prediction model to consumer wearable data. Using a publicly available longitudinal dataset, we evaluated the model's ability to flag impending illness in advance. The KDE approach detected most sick users but yielded a high false-positive rate. We investigated potential causes-including inter-individual physiological variability, activity-induced confounding, data sparsity, and threshold calibration-and outlined recommendations for future models: personalization with user-specific baselines and drift handling; integration of multimodal features (e.g., heart rate, heart rate variability, sleep, temperature, and activity context); robust anomaly scoring and calibration; temporal smoothing with alert confirmation windows; and prospective validation in perioperative cohorts. Our findings highlight both the promise and current limitations of presymptomatic illness prediction for surgical scheduling and provide actionable steps to reduce false positives while preserving sensitivity, with the goal of minimizing elective surgery cancellations.