Diabetes Prediction Based on Limited Medical Indication

Using machine learning methods based on relevant physiological indicators for disease prediction has become a relatively mature and widely applied technique. Diabetes is a common and prevalent disease, posing a significant global public health problem that seriously threatens human health. Currently, various machine learning and deep learning techniques, such as neural networks, have been employed for diabetes prediction. Most methods have achieved significant results by utilizing complex models on large datasets. However, many discussions on these methods lack classification analysis of experimental datasets or unilaterally pursue high precision. In the case of potential patients, a model with high recall rate becomes more important, as it reduces the probability of potential patients being misclassified as normal individuals. In this study, detailed analysis of various attributes of the dataset was conducted through a hybrid approach based on statistics and mathematics. Based on the conclusions drawn from this stage and considering factors of diabetes in society, the composition of potential cases in a small dataset was analysed. The performance of common machine learning models was tested in experiments. Ultimately, a linear model was selected and optimized, and model performance was evaluated using the confusion matrix and ROC curve, demonstrating balanced and satisfactory precision and recall scores.