A Lightweight Tiny ML-Based Resource Management for Health Monitoring Using LoRaWAN

Abstract

The rapid proliferation of the Internet of Things (IoT) has transformed healthcare by enabling continuous monitoring through wearable devices, yet challenges remain in achieving reliable, low-power communication persist, particularly in mobile scenarios. Long Range Wide Area Network (LoRaWAN), a prominent Low-Power Wide-Area Network (LPWAN) technology, offers extended range and energy efficiency, but traditional Adaptive Data Rate (ADR) mechanisms exhibit inefficiencies, leading to packet loss and high energy consumption. This thesis introduces a lightweight Tiny Machine Learning (TinyML) framework that enhances resource allocation in LoRaWAN for health monitoring applications. Building upon the AI-ERA approach, a previously developed AI-based resource allocation method, the framework employs an ensemble model integrating Support Vector Classifier (SVC), XGBoost, and Deep Neural Network (DNN), with Out-of-Fold (OOF) predictions (predictions for each data point made by a model that was not trained on that point) and a meta-classifier (Logistic Regression), achieving 86% accuracy in optimal Spreading Factor (SF) prediction. Using NS-3 simulations, the framework demonstrates up to 32% improvement in Packet Success Ratio (PSR) for static scenarios and 28% for mobile ones compared to standard ADR, alongside reduced energy consumption and reducing convergence time (to optimal SF allocation) to 6 hours. Deployable on constrained devices, it enables real-time adaptability in wearables while addressing privacy, latency, and power constraints in health IoT. This work advances scalable health monitoring systems through proactive, edge-based intelligence, with potential extensions to Federated Learning for enhanced privacy. Keywords: TinyML, LoRaWAN, Ensemble Learning, Resource Allocation, Health Monitoring, IoT.