Machine Learning on Microcontrollers for Biological Sensing: A Systematic Review

Microcontroller-class devices, when integrated with machine learning (ML) models, offer transformative potential for biological sensing in resource-constrained environments. However, the deployment of such systems demands a careful balance between computational limitations, sensor integration, and ecological relevance. This systematic review evaluates trends, architectures, constraints, and applications of ML deployed on microcontroller-class hardware for biological sensing between 2015 and 2025. A systematic search across Google Scholar (n = 142), Web of Science (n = 22), and Scopus (n = 4,266) yielded 4,430 records. After screening and eligibility assessment using PRISMA guidelines, 60 studies were included. The review focused on temporal trends, research types, ML toolchains, hardware platforms, task types, model architectures, dataset sources, system constraints, performance metrics, and domain-specific applications. Publication activity surged after 2019, peaking again in 2024. Most studies employed empirical and applied research methods (Fig. 8), with a majority using embedded platforms like Arduino and TinyML (32.61%) and lightweight frameworks such as TensorFlow Lite. ARM-based processors (34%) and AI-focused SoCs (22%) were the most common hardware platforms. Classification tasks dominated (56.36%), followed by monitoring (25.45%) and regression (18.18%). Deep learning architectures (CNNs, LSTMs, VAEs) accounted for 55.56% of models used. Most studies utilized custom, real-world datasets (67.27%) (Fig. 13) and emphasized performance constraints such as low latency (< 500 ms, 52%) and memory optimization (36%). Hardware limitations were primarily memory-based (44%) or unspecified (32%) (Fig. 15). Real-time inference (38.18%) and edge-device suitability (16.36%) were the most reported performance goals. Application areas were led by healthcare monitoring (25.45%) and water quality analysis (23.64%). Dominant toolchains included Arduino (29.09%), TensorFlow Lite (18.18%), and Edge Impulse (12.73%). Machine learning on microcontroller-class hardware is gaining traction in biological sensing, particularly in health and environmental monitoring. Despite progress, challenges persist in standardized benchmarking, performance reporting, and balancing system constraints. This review offers a detailed synthesis of implementation trends and practical bottlenecks, guiding future development of robust, low-power, and domain-specific ML sensing platforms.