Machine Learning on Microcontroller-Class for Oxygen Dynamics Prediction: A Systematic Review

Machine Deploying machine learning (ML) on microcontroller-class hardware introduces unique challenges due to limitations in memory, computation, and power. This is particularly critical in health monitoring systems, where timely and accurate oxygen dynamics prediction is essential. This systematic review evaluates the use of ML techniques on microcontroller platforms for oxygen dynamics prediction, focusing on model efficiency, task performance, hardware optimization, and real-time responsiveness across health and environmental monitoring applications. A comprehensive review of 62 peer-reviewed articles published between 2015 and 2025 was conducted across Scopus, Web of Science, and Google Scholar. Eligible studies were analyzed based on hardware type, ML models, deployment tools, evaluation metrics, task categorization, and application domains. The findings revealed a steady increase in research output, peaking in 2021 and 2023. Journal articles comprised 66% of sources, with dominant contributions from China (43.55%), Egypt (9.68%), and Bangladesh (8.06%). Widely adopted ML tools included LogNNet (24.19%), deep learning frameworks (11.28%), and Python-based platforms. ESP8266 (19.35%), ESP32 (14.52%), and Arduino-class boards were the most common hardware. Classification tasks led with 59.68%, followed by prediction (30.65%) and regression (6.45%). ANN (20.96%), SVM (14.52%), and Random Forest (11.29%) were the most applied models. Evaluation focused on accuracy (27.37%) and real-time latency (25.76%), though only 6.44% of studies reported inference performance. Application domains were led by wearable health monitoring (28.98%) and water quality analysis (17.71%). Microcontroller-based ML systems for oxygen dynamics prediction are rapidly evolving, with strong adoption of lightweight algorithms and embedded AI frameworks. Despite progress in model quantization, pruning, and on-device inference, key limitations persist in transparency, inference-time metrics, and hardware-specific integration. Future research should target more balanced evaluation strategies, power-aware deployments, and expanded exploration of domain-specific and sensor-fusion techniques.