Predictive study of T2D mellitus based on a multimodal fusion model: an example of electrocardiographic and clinical indicators

The present study constructed a prediction model for type 2 diabetes (T2D) based on deep learning and machine learning techniques. The objective of this construction was to achieve an accurate prediction of disease risk and onset time by fusing unimodal data (e.g. lifestyle habits, health indicators), multimodal data (e.g. ECG images with numerical features) and time series data. In the unimodal task, the Stacking Classifier integrated model achieved 85.28% accuracy and 91.83% AUC; in the multimodal task, the designed ResNet18 and numerical feature fusion neural network (MMNN) performed optimally (AUC 0.93, F1-score 0.88); in the onset time prediction task, the LSTM model effectively captured the temporal dependence with an AUC of 0.93 and stable convergence of MSE. The experimental results demonstrate that the proposed method significantly outperforms the baseline model in terms of accuracy, precision, recall and AUC. This verifies the clinical value of multimodal fusion for T2D prediction and provides reliable technical support for personalised medicine and health management.