Anesthesia Drift Detector

Problem

Machine learning models deployed in anesthesia monitoring are vulnerable to silent performance degradation as patient populations, devices, and clinical practices evolve. Traditional accuracy metrics fail to capture these shifts in real time, creating safety risks.

Approach

Designed a drift monitoring pipeline that tracks changes in statistical properties of time-series physiological signals. The system compares incoming data against baseline distributions and flags significant deviations using interpretable metrics.

• Time-series feature extraction
• Distributional comparison (statistical tests & thresholds)
• Alerting logic for sustained drift

Validation Strategy

Evaluated drift sensitivity using simulated distribution shifts and retrospective signal segments. Emphasis was placed on minimizing false alarms while ensuring early detection of clinically meaningful changes.

Clinical Impact

This framework supports proactive model governance by surfacing reliability risks before downstream clinical decisions are affected. It provides a practical layer of safety for AI-assisted anesthesia monitoring.

Technology Stack

Python · Statistical analysis · Time-series processing · Visualization