Abstract. Continuous multi-sensor monitoring (heart rate, SpO₂, respiration rate, blood pressure, and temperature from bedside and wearable sensors) generates extensive data streams. Clinicians will not respond to signals that they are unable to explore, posing a barrier to the clinical adoption of black-box ML early warning models. Place the gap: deep models are sensitive but opaque, whereas existing early warning scores (NEWS2) are interpretable but insensitive. Your contribution will be an explainable machine learning pipeline to help close the trust gap.
To predict clinical deterioration within a 4-6 hour window, an ensemble of gradient-boosted models (XGBoost or LightGBM) trained on multi-parameter physiological time series from a public ICU dataset (MIMIC-III/IV via PhysioNet is the obvious choice; alternatively, the eICU database). The feature engineering is carried out using the five vital-sign channels' sliding windows. Explainability layer: physiological plausibility is evaluated against NEWS2 criteria, with SHAP for global feature attribution and local explanations for each alarm. Logistic regression and NEWS2 are used as baseline comparisons.
Provide a qualitative conclusion based on the SHAP analysis (for example, respiratory rate and SpO2 trend dominating attributions, consistent with clinical literature and reviewers' validation), sensitivity at a clinically acceptable false-alarm rate, and AUROC and AUPRC vs baselines. Explainable models pave the way for dependable implementation in IoT-enabled patient monitoring by achieving equivalent or superior traditional ratings while providing clinically acceptable alarm explanations.