This research addresses the challenge of interpreting machine learning models used to analyze continuous biosensing data. By developing novel interpretability techniques tailored to time-series health data, we enable clinicians to understand why models make specific predictions about patient health.
Machine learning models excel at finding patterns in complex biosensing data, but their black-box nature creates a barrier to clinical adoption. This study develops interpretability methods specifically designed for understanding ML predictions on continuous health monitoring data.
We adapted and developed multiple interpretability techniques including SHAP values for time-series data, temporal attention visualization, and pattern discovery algorithms. These methods allow clinicians to see exactly which patterns in biosensing data led to model predictions.
We validated our interpretability methods with clinicians who specialize in mast cell diseases and autonomic disorders. Clinicians confirmed that the identified patterns matched their clinical intuition and knowledge, suggesting the models learn legitimate medical reasoning.
We developed open-source tools that clinicians can use to inspect model predictions on individual patient data. This enables incorporation of AI into clinical workflows with full transparency and clinician oversight.
Interpretable ML on biosensing data is achievable and clinically valuable. By making models transparent, we enable their responsible use in patient care.