Comparative Evaluation of Lightweight Machine Learning Classifiers for Rapid Infectious Disease Identification on Resource-Constrained Point-of-Care Devices: Accuracy, Latency, and Subgroup Performance

Abstract

Portable point-of-care (POC) diagnostic devices hold considerable promise for improving the detection of infectious diseases in resource-limited healthcare settings, yet the performance of lightweight machine learning (ML) classifiers under realistic hardware constraints remains insufficiently characterized. This study presents a comparative evaluation of eight lightweight classifiers---five traditional ML algorithms (Random Forest, XGBoost, Support Vector Machine, k-Nearest Neighbors, Logistic Regression) and three compact deep learning architectures (quantized MobileNetV2, EfficientNet-B0, SqueezeNet) ---across two clinically relevant tasks: malaria parasite detection from thin blood smear images and tuberculosis identification from chest radiographs. Using the NIH Malaria Cell Images dataset (27,558 pre-segmented cell images from 200 patients) and the NLM tuberculosis chest X-ray datasets (800 radiographs), we assess classification accuracy, deployment-relevant latency, and age-stratified sensitivity patterns. EfficientNet-B0 achieved the highest accuracy on both tasks, while XGBoost provided the strongest baseline among classifiers operating on frozen ResNet-18 embeddings. On Raspberry Pi 4, XGBoost required 38 ms for the classification step alone, but its estimated end-to-end latency increased to approximately 143 ms once the shared ResNet-18 feature-extraction stage was included, compared with 210 ms for quantized MobileNetV2. Age-stratified analysis of the Shenzhen subset showed a consistent tendency toward lower sensitivity among patients over 60, with smaller observed declines for the lightweight deep learning models; however, these subgroup patterns should be interpreted with caution because the sample sizes were modest and no formal significance test was applied. Overall, the study provides a deployment-oriented comparison of lightweight diagnostic classifiers while highlighting the importance of fair latency accounting and cautious interpretation of subgroup differences.