DEM: A Distilled Explanation Model for Interpretable Anomaly Detection in Physiological Sensor Networks

Researchers propose DEM (Distilled Explanation Model), a glass-box framework for anomaly detection in physiological sensor networks that distills gradient boosting expertise into interpretable decision trees while maintaining high accuracy (AUC 0.9964). The model achieves 1235x faster inference than SHAP-based methods, making it practical for real-time medical monitoring with clinically meaningful explanations rather than post-hoc approximations.

DEM addresses a critical gap in medical AI where black-box models excel at prediction but fail at clinical explainability, while transparent models sacrifice accuracy. The framework operates through a three-stage process: establishing a linear baseline, training a gradient boosting expert on residuals, then distilling that non-linear knowledge into a decision tree that generates human-readable rules. This approach is fundamentally different from post-hoc explanation methods like SHAP and LIME, which approximate predictions after-the-fact; instead, the explanation IS the prediction itself.

The research reflects broader momentum toward interpretable AI in healthcare, where regulatory bodies and clinicians demand transparency alongside performance. Medical institutions increasingly face liability concerns when deploying opaque systems, creating genuine demand for models that can justify their decisions through logical rules. The introduction of a distillation fidelity metric provides quantifiable assurance that explanations faithfully represent the underlying model's behavior—a methodological advance absent from previous interpretable approaches.

The performance metrics across multiple clinical datasets (MIMIC-IV, eICU) and wearable stress detection establish credibility for real-world deployment. The 0.17ms inference time per 1000 samples enables genuine real-time monitoring in Wireless Body Area Networks, a critical requirement for continuous patient surveillance. The controllable accuracy-interpretability trade-off represented by adjustable tree depth gives practitioners explicit governance over their systems rather than forcing all-or-nothing choices.

Future development should focus on clinical validation across diverse populations and integration with existing hospital IT infrastructure. Adoption barriers will likely involve clinician training rather than technical limitations, since the if-then rules require minimal domain expertise to understand.

• →DEM distills non-linear gradient boosting into interpretable decision trees, producing explanations that are predictions rather than approximations of predictions.
• →Achieves 0.9964 AUC on clinical anomaly detection while running 1235x faster than SHAP-based explanation methods, enabling real-time physiological monitoring.
• →Introduces novel distillation fidelity metric to quantify how faithfully explanation trees capture expert model behavior, providing trustworthiness measures absent from prior interpretable models.
• →Generates human-readable if-then rules with user-controlled tree depth, allowing explicit accuracy-interpretability trade-offs unique among intrinsically interpretable systems.
• →Evaluated across four physiological datasets including hospital records (MIMIC-IV, eICU) and wearable sensors, demonstrating real-world applicability for medical institutions.