AI-Based Cardiac Early Disease Detection Using Multimodal Data

Abstract

Accurate and timely diagnosis of cardiac abnormalities is crucial in order to prevent delays in diagnosis and to assist clinical decision making. In this study, the authors compare the performance of a deep learning model trained using demographic and engineered clinical features against an ECG-only model and test if there is some added value to using this extra information. We employed the PTB-XL dataset to create two binary classification models: a CNN that takes solely the 12-lead ECG signals and a late-fusion model that utilizes a combination of ECG signals and engineered tabular and demographic features. The official stratified train–test split was used to assess a total of 21,388 ECG recordings. Model accuracy, ROC-AUC (Receiver Operating Characteristic - Area Under the Curve), precision, recall, false-positive rate, precision–recall analysis, brier score, calibration, and threshold constrained evaluation at a false-positive rate of 5% were used to assess model performance. The accuracy of the ECG-only model was 0.855, with an ROC-AUC of 0.932, and 0.843 with an ROC-AUC of 0.929 for the fusion model. The ECG-only model had a recall of 0.744 with a False - positive rate of ≤5%, and the fusion model achieved a score of 0.733. These findings suggest that a simple demographic and engineered feature set is not always beneficial for cardiac pathology detection when the underlying ECG algorithm is well developed. The results highlight the need for evaluation to be threshold-aware and calibration based for medical multimodal AI systems.