ARCHIVES

Machine learning now supports diagnosis in many clinical settings, from risk scores built on patient records to deep networks that read scans. The evidence, however, is split by data type. Studies on tabular records, biosignals and medical images use different models, different metrics and different validation practices, so it is hard to judge which approach suits a new diagnostic task. This paper reviews diagnostic learning across three data modalities: structured clinical and screening records, physiological signals collected by sensors and IoT devices, and medical images such as CT, MRI, X-ray, echocardiography and dermoscopy. It links each modality to the models that are reported to work best on it and to the practical problems that come with it. The review finds that feature selection combined with tree ensembles or hybrid classifiers performs well on structured records, that random forests and support vector machines remain strong on extracted signal features, and that convolutional networks and their ensembles lead on images. Across all three modalities, class imbalance, small single-centre datasets and limited explanation are recurring problems, and privacy rules restrict the pooling of data. Future work should report recall and calibration next to accuracy, test models on external sites, and adopt privacy-preserving training such as federated continual learning.