Abstract
Continuous health monitoring in Internet of Things (IoT) settings increasingly requires inference to be executed on-device rather than in the cloud, especially when latency, intermittent connectivity, privacy, and battery life are primary design constraints. This article presents a resource-aware TinyML deployment benchmark for a partial National Early Warning Score 2 (NEWS2)-aligned health-status classification workload on a resource-constrained ESP32 microcontroller. The task maps four routinely available vital signs; heart rate, respiratory rate, peripheral oxygen saturation, and body temperature, to three derived severity categories: normal/stable, mild, and critical. Because the labels are rule-derived rather than prospectively adjudicated clinical outcomes, the task is treated as a clinically interpretable embedded-inference workload rather than as a validated clinical deterioration prediction model. Five model families were evaluated under a common training, conversion, firmware, and measurement protocol: decision tree (DT), random forest (RF), support vector machine (SVM), multilayer perceptron (MLP), and deep neural network (DNN). The main contribution of this study is an edge-deployment evaluation that exposes trade-offs among predictive accuracy, end-to-end latency, memory footprint, throughput, and current draw. On the held-out software test set, RF obtained the highest accuracy, followed closely by DT, while all models achieved broadly comparable aggregate performance. On-device measurements showed that operational behavior differed much more strongly than predictive performance. RF achieved the highest selected-subset device accuracy but approached the ESP32 memory envelope, whereas DT provided the best deployment profile, with the lowest total latency, highest throughput under the total-latency definition, lowest current draw, and minimal static random-access memory (SRAM) and flash usage. Neural models remained compact after quantization but consumed substantially higher current during continuous inference, while the radial-basis-function SVM imposed high SRAM pressure. The results show that, for low-dimensional health-monitoring workloads on microcontrollers, the best deployable model is not necessarily the most accurate standalone model but the one that preserves latency, memory, and energy headroom.