Development of an autonomous system for identification and predictive modeling of the physiological state of individuals at risk

Abstract

The article presents the development of a local autonomous system for monitoring and predicting critical physiological states in at-risk-individuals. The system is intended for persons with special needs who live independently or reside in medical care facilities. The objective of this work is to develop and substantiate a hybrid model for the real-time assessment and short-term prediction of a person’s critical physiological condition. To ensure such an operational mode, the system functions without reliance on cloud services or external infrastructure. The main tasks addressed include the design of a multisensor architecture based on wearable sensors (MAX30102, MPU6050, GY-906), stationary mmWave radars, and a local Home Assistant server. Mathematical models for risk assessment were developed and investigated, including the instantaneous index A(t), the predictive index P(t), and a recurrent LSTM-based model. The methods employed comprise analytical and computational approaches (normalized nonlinear aggregation of deviations, DTW-based comparison with crisis prototypes, Isolation Forest, and k-NN regression for time-to-event estimation), computational–experimental methods (training and validation of an LSTM model on the PhysioNet BOLD dataset), and hardware–software implementation (ESP32 + Home Assistant + InfluxDB/MySQL). The following results were obtained. An instantaneous alarm index A(t) with clinically justified weights and a nonlinear sensitivity function was proposed. An integral predictive index P(t) was developed, incorporating trajectory similarity to crisis episodes, anomaly detection, and the estimated time to event. A dual-channel LSTM model was implemented and tested (ROC–AUC = 0.6956 on the BOLD dataset), enabling the detection of slow degradation trends. Conclusions. The scientific novelty of the obtained results lies in the proposed autonomous hybrid platform for immediate response and short-term prediction over a 10–30 min horizon, which ensures privacy preservation through the use of mmWave radars instead of video cameras. The models A(t), P(t), and the LSTM approach were further developed and shown to complement each other, enhancing sensitivity to pre-crisis states. This enables the creation of a stable autonomous system, which is particularly important for regions with unreliable infrastructure.