A Deep Learning–Based Multi-Source Sensor Data Fusion Model for Real-Time Disaster Prediction and Healthcare Risk Assessment

Abstract

Simultaneous natural disaster prediction and healthcare risk assessment of affected populations is a complex problem in modern intelligent systems because the data resulting from diverse sensor modalities are often heterogeneous, high-dimensional, and suffer temporal misalignment. Current disaster forecasting and clinical health risk monitoring methods consider them as separate domains, neglecting their inherent spatio-temporal interdependencies during crisis events. We present a novel Multi-Source Deep Fusion Network (MS-DFNet) that jointly learns from heterogeneous sensor streams, including seismic accelerometers, meteorological sensors, environmental IoT nodes and wearable physiological monitors in a single, multi-scale, hierarchical architecture. The latest form of intelligent systems, one such challenge is managing prediction of natural disasters and healthcare risk assessments for populations in parallel due to data from multimodal sensors being heterogeneous and high-dimensional with potential temporal misalignment. Current methods treat disaster predictions and epidemiological risk monitoring as separate processes, without leveraging their inherent correlations in time and space during crisis events. In this paper, we introduce MS-DFNet, a new MS-DFNet that operates on heterogeneous sensor streams such as seismic accelerometers, meteorological sensors and environmental IoT nodes or wearables’ physiological monitors within a unique hierarchical architecture. The proposed framework has three main contributions: (1) a Cross-Modal Temporal Attention (CMTA) mechanism that adaptively aligns and emphasises heterogeneous time-series signals; (2) a Dual-Task Cascade Prediction Head (DTCPH) for simultaneously generating both disaster risk probability maps along with individual healthcare risk scores; and (3) an Adaptive Sensor Dropout Regularisation (ASDR) strategy to improve the model robustness against sensor failure and missing data, which is an actual need in disaster situations. MS-DFNet achieves an average disaster prediction F1-score of 93.7% and healthcare risk AUC-ROC of 0.961, surpassing stateof-the-art baselines by 4.2–8.9% on all major metrics across extensive experiments over four benchmark datasets (i.e., STEAD seismic, CAMELS hydro-meteorological, PhysioNet MIMIC-III and a newly curated multi-hazard IoT corpus). The resulting model inference can be executed end-to-end in less than 38 ms where the required embedded edge hardware is available, indicating real-time practicality. These findings validate that MS-DFNet is an important step toward unified situational-awareness systems for smart cities, emergency response and remote healthcare infrastructure.