Hybrid early warning model for autonomous physiological monitoring with a minimal sensor set

This paper presents a validation of a hybrid model for remote physiological monitoring and patient deterioration prediction using the MIMIC-IV (51,981 patients) with external validation on MIMIC-III (30,528 patients, zero overlap). The objective is to develop and substantiate a methodology for the quantitative assessment of prediction quality using four vital signs (HR, SpO₂, RR, Temp) and to evaluate a hybrid model combining a rule-based composite index with machine learning results: A(t)+ML. The methods employed comprise the recalibration of the A(t); analysis of 4 conditions (Full, NEWS2-set, CAS-4 and CAS-4 hybrid); supervised machine learning for classification and regression (XGBoost), the high-performance gradient boosting framework (LightGBM), logistic regression (LR), and recurrent neural network architecture (LSTM). Statistical tools include FDR Benjamini-Hochberg correction, SHAP, regression analysis for determining the VIF (Variance Inflation Factor); and a methodology for determining the Net Reclassification Improvement index (NRI). Results. Using quantitative assessments, an adequate level of prediction quality for the early detection of physiological deterioration was demonstrated when using only four vital sign parameters available from low-cost wearable sensors. It was also shown that the hybrid combination of an analytical risk index with machine learning can compensate for the absence of laboratory data and comprehensive electronic health records. Conclusions. The use of only four basic parameters combined with machine learning procedures for determining a patient's current physiological state of a patient provides clinically meaningful predictions for a broad class of remote monitoring systems, including those currently deployed.