Machine Learning for Early Detection of Diabetes: A Systematic Review of Methods, Datasets, and Challenges

Diabetes mellitus is a worldwide health issue that is increasing in prevalence across different populations. Early detection is necessary to prevent complications such as cardiovascular diseases, nephropathy, neuropathy and retinopathy. Traditional diagnostic approaches tend to rely on invasive blood tests, are clinically validated but do not often predict disease progression beyond a clinical metabolic diagnosis. Machine learning (ML) is a relatively new method which has the ability to identify emergent risk pathways that could be representative of diabetic influences on a patient that are not evident but share common patterns in a heterogeneous dataset of clinical, demographic and lifestyle variables, therefore representing new pathways to help manage and monitor risk. This systematic review consolidates the studies to identify early detection of diabetes using ML techniques, including traditional algorithms (logistic regression, decision trees, support vector machines), emerging methods (ensemble methods, deep learning), and hybrid systems that involve Internet of Things (IoT) devices and federated learning. It prioritises datasets (such as the Pima Indian Diabetes Dataset, hospital datasets and local databases), metrics of evaluation (accuracy, sensitivity, specificity, and ROC-AUC). Overall, it has found that ML techniques have better predictive accuracy across studies compared to traditional diagnostic methods. Overall, there are many issues facing ML with concerns over data imbalance, interpretability and generalizability across populations. The systematic review presents an indication of mapping trends in methodologies, a discussion about limitations and suggestions for the future including knowledgeable explainable AI and future EHR integration and a privacy-preserving federated learning environment. In summary, the review highlights the potential of ML to improve early identification and support personalized approaches of diabetes management.