Comparing Lightweight Models for Physician-Consensus Pharyngitis Classification from Primary-Care Smartphone Images

Background/Objectives: Smartphone-acquired throat images could support accessible pharyngitis assessment, but the performance and computational efficiency of lightweight models on such images require further evaluation. This study compared five lightweight models for predicting physician-consensus bacterial versus nonbacterial labels from smartphone throat images. Methods: We analyzed PGUPharyngitis, a public dataset of smartphone throat images collected in general practice. The final cohort comprised 635 patients, of whom 105 (16.5%) had bacterial physician-consensus labels. The reference outcome was physician consensus based on images and clinical information rather than microbiological testing. Five ImageNet-pretrained lightweight classifiers—four convolutional neural networks (CNNs; MobileNetV3-Small, EfficientNet-B0, ShuffleNetV2, and YOLO26n-cls) and one CNN–Transformer hybrid (MobileViT-XS)—were fine-tuned using patient-level five-fold cross-validation. Performance was assessed using area under the precision–recall curve (PR-AUC) and area under the receiver operating characteristic curve (AUROC), alongside computational efficiency. Device-only and clinical-variable logistic-regression models provided non-image comparisons. Results: Mean PR-AUC across folds ranged from 0.317 to 0.372 and AUROC from 0.663 to 0.708, with no clearly superior architecture. Bacterial-label prevalence differed by acquisition device (Samsung 9.2% versus Xiaomi 26.9%); the device-only and clinical-variable comparators achieved AUROCs of 0.650 and 0.598, respectively. At validation-selected high-sensitivity thresholds, the models flagged 65–78% of patients, with positive predictive values of only 19–22%. All models contained fewer than five million parameters, making them compact candidates for future evaluation on resource-constrained devices. Conclusions: Lightweight models showed moderate performance against physician-consensus labels. However, acquisition device was associated with the reference labels, image-model AUROCs were not clearly distinguishable from that of the device-only comparator, and device-related confounding cannot be excluded. Because performance was measured against these labels and positive predictive values were low, the findings do not establish diagnostic accuracy or clinical applicability, which require validation against microbiological reference standards in independent clinical cohorts.