Integration of clinical data with scanned ECGs using deep learning methods for stroke risk prediction in Indian patients with atrial fibrillation: evidence from the KERALA-AF study

Figures

• 

  Fig. 1. Overview of the proposed multimodal deep learning architecture. The framework consists of two parallel branches: (1) an Image Encoder (left) utilising a ResNet-50 backbone,16 MLP, and Dropout layers to extract features from ECG images; and (2) a Tabular Encoder (right) processing categorical (e.g., Sex) and continuous (e.g., BMI) clinical features via embeddings, self-attention mechanisms, and feed-forward networks. The resulting feature vectors from both modalities are concatenated and passed through a final classifier to generate the prediction.

• 

  Fig. 2. Receiver operating characteristic (ROC) curves for multimodal deep learning and machine learning models in 1-year stroke prediction. ROC curves compare the discrimination performance of the multimodal deep learning model (orange; integrating paper-based ECG images and clinical data) with traditional machine-learning models trained on tabular clinical variables only: random forest (blue), logistic regression (green), support vector machine (red), decision tree (brown), and k-nearest neighbours (purple). The dashed black line represents a random classifier (AUC = 0.50). The deep learning model achieved the highest AUC (0.816, 95% CI 0.704–0.914), outperforming all other approaches. AUC, area under the receiver-operating characteristic curve; CI, confidence interval; ECG, electrocardiogram; KNN, k-nearest neighbours; SVM, support vector machine. All ROC curves were derived from the same independent validation cohort.

• 

  Fig. 3. Receiver operating characteristic (ROC) curves for CHA2DS2-VASc–based models in 1-year stroke prediction. ROC curves illustrate the discrimination performance of four CHA2DS2-VASc–based classifiers: logistic regression (red), random forest (green), support vector machine (blue), and simple neural network (grey). The dashed black line represents a random classifier (AUC = 0.50). AUC values with 95% confidence intervals are shown in the legend. AUC, area under the curve; CI, confidence interval; SVM, support vector machine; ROC, receiver operating characteristic. All ROC curves were derived from the same independent validation cohort.

• 

  Fig. 4. Top 10 most important clinical tabular variables contributing to stroke prediction in the multimodal deep-learning model. Bars represent permutation-based feature importance, quantified as the mean AUROC decrease when each clinical variable was permuted on the validation set while the ECG-image branch remained unchanged. Higher values indicate stronger predictive contribution. AST/SGOT, diastolic blood pressure, heart rate, systolic blood pressure, age, BMI, and creatinine clearance were the top contributors, followed by left atrial size, TSH, and diabetes.