Edge-Aware Disease Prediction in IoT Healthcare: Cost-Constrained Learning, Calibration, and On-Device Benchmarks

Abstract

Edge inference is increasingly favored in Internet-of-Things (IoT) healthcare to reduce latency, maintain availability under poor connectivity, and keep sensitive biosignals on device. Yet models that excel in the lab often exceed embedded budgets and provide poorly calibrated probabilities for clinical use. This paper presents a deployment-oriented framework—Edge-Aware Disease Prediction—that unifies teacher–student learning, post-training INT8 quantization, probability calibration, and an exportable on-device benchmarking path. Within a leakage-safe nested cross-validation protocol, a higher-capacity teacher guides a compact student via distillation; the student is then quantized to INT8 and calibrated with temperature scaling. We evaluate on the UCI Heart Disease (Cleveland) dataset and compare against strong classical baselines. Across 15 outer folds, the quantized student matches the teacher in discrimination (mean ROC AUC about 0.879) with no statistically significant difference in paired tests, while calibration analysis (Brier score and reliability diagrams) confirms the value of post-hoc scaling. An ONNX export enables device-aware Pareto analysis over accuracy, latency, and size to select an "Edge-Knee" operating point for deployment. The results indicate that edge-constrained models can retain clinical utility while producing better-behaved probabilities, offering a reproducible pathway from development to embedded decision support.

Ähnliche Arbeiten