preDiabet: Neural Network for Diabetes Risk Prediction, AI Empowers Early Prevention

The preDiabet project is maintained by M7tnj and was released on GitHub on June 3, 2026. It aims to build a neural network-based model to predict diabetes risk by analyzing health indicators and identify high-risk groups to support early prevention and intervention. The project covers the entire machine learning workflow and provides a practical case for the application of medical AI in chronic disease management.

Diabetes is a common chronic disease worldwide. According to WHO statistics, there are about 422 million patients globally, and 1.5 million people die directly from diabetes each year. A large number of people are in the asymptomatic pre-diabetes stage, and if not intervened, the risk of developing diabetes is extremely high. AI technology can identify high-risk groups by analyzing health data to achieve early warning, and the preDiabet project is an attempt in this direction.

The project's tech stack includes data collection and preprocessing, model architecture design, training, and result visualization. Neural networks are chosen because they can capture non-linear relationships, perform automatic feature learning, handle high-dimensional data, and support end-to-end learning. The typical architecture includes an input layer (corresponding to health indicators), hidden layers (fully connected + ReLU activation), regularization (Dropout/L2/early stopping), and an output layer (Sigmoid activation for binary classification). The loss function uses binary cross-entropy, and the optimizer of choice is Adam.

Prediction requires physiological indicators (blood glucose, blood pressure, BMI, age), lifestyle factors (diet, exercise, smoking and alcohol consumption), family medical history (diabetes history, gestational diabetes history), and other medical indicators (blood lipids, insulin, renal function). Common data sources include the Pima Indians Diabetes Dataset, NHANES survey data, and desensitized hospital electronic medical records.

Data preprocessing includes cleaning (missing value/outlier handling), feature engineering (scaling, encoding, selection), and data splitting (70% training /15% validation /15% testing). Training uses batch updates, monitoring training/validation loss, accuracy, and AUC-ROC. Evaluation metrics include confusion matrix, accuracy, precision, recall, etc. In medical scenarios, recall is more focused on to reduce missed diagnoses.

The project can achieve early warning (intervention reduces the risk of onset by 58%), optimize medical resources (focus on high-risk groups), and assist personal health management (risk awareness and behavior change). Studies show that pre-diabetes intervention is effective, and AI tools can promote the transformation of preventive medicine from "treating existing diseases" to "preventing future diseases".

Current limitations include dependence on data quality, the black-box problem of neural networks, and insufficient adaptation to individual differences. Future directions include multi-modal data fusion (genetics + imaging + wearables), time-series modeling (RNN/Transformer), and federated learning (collaborative training under privacy protection).

preDiabet demonstrates the potential of AI in the healthcare field. Although it cannot replace doctors, it can assist in identifying high-risk groups. For developers, the project covers the entire machine learning workflow and is an excellent case for medical AI and deep learning practice, promoting the development of chronic disease management towards a data-driven prevention model.