AutoFocal Loss with Lightweight DNN for Stroke Prediction in Fog Computing

The growing volume of Internet of Things (IoT) healthcare data makes efficient patient monitoring and disease prediction increasingly important. Cloud Computing (CC) enables scalable storage and processing, but struggles with latency-sensitive applications. Fog Computing (FC) addresses this issue by placing resources closer to users. Existing fog-based Deep Learning (DL) models for disease monitoring often perform poorly on imbalanced healthcare datasets, are biased toward the majority class, and are typically designed for large cloud systems. This paper introduces a fog-based framework for real-time stroke prediction using a lightweight, quantization-optimized DNN with automatically weighted focal loss. The proposed model handles severe class imbalance, maintains diagnostic accuracy, and reduces model size for fog deployment. Experiments on a public stroke prediction dataset show that the proposed system achieves higher AUPRC, recall, and F1-score than baseline models, while remaining compact enough for fog computing.