Interpretable Independent Recurrent Networks for Forecasting Stroke in Atrial Fibrillation

Figures

• 

  Figure 1. Study Design of Prediction Framework This figure illustrates the definition of our model input and prediction targets. Our model uses a patient’s electronic health records 2 years before the first atrial fibrillation as the input sequence. It predicts whether the patient will develop the first stroke within 1 year after the first atrial fibrillation, formulated as a binary classification. The input sequence is further summarized over 1 month to help the training process, which will be detailed in the Data Preprocessing section.

• 

  Figure 2. Model Architectures for Time-Series Prediction The schematic assumes a sequential data set with T time steps and m features per time step. (A) Standard gated recurrent unit (GRU), which uses the shared GRU and fully connected layer for all features to make predictions. (B) Forecasting Strokes via Interpretable Independent Networks (ForeSIIN) with a shared fully connected layer, which allows independent GRU layers for every feature. (C) The fully-independent structure of ForeSIIN with independent GRU and fully connected layers. Each model branch consisting of a distinct GRU and fully connected layer for a specific feature can be regarded as a sub-classifier and contribute to the final prediction independently.

• 

  Figure 3. Model Performance Metrics This figure shows the curves comparing predictive performance across models. The tables on the lower right list the area under curves with the lower/upper bounds of the 95% CI. (A) Receiver-operating characteristic (ROC) curves. (B) Precision-recall curves (PRC). LR = logistic regression; RF = random Forest; SVM = support vector machine; XGB = extreme gradient boosting; other abbreviations as in Figure 2.

• 

  Figure 4. Feature Impact Analysis Feature contributions to stroke prediction were quantified using their influence on the model’s logit output in the ForeSIIN framework. The 20 most impactful features are presented, ranked by the average difference between patients who did and did not experience ischemic stroke within 1 year of atrial fibrillation (AF) onset (“t1–t0”). Positive values reflect stronger predictive separation. ALT = alanine aminotransferase; AST = aspartate transferase; BMI = body mass index; eGFR = estimated glomerular filtration rate; GLUAC = fasting glucose; HbA1c = glycosylated hemoglobin; Hct = hematocrit; hsCRP = high-sensitivity C-reactive protein; LDL = low-density lipoprotein; LVEF = left ventricular ejection fraction; TG = triglyceride; TR max PG = peak tricuspid regurgitation velocity.

• 

  Figure 5. Kaplan-Meier Analysis and Confusion Matrix (A) Kaplan-Meier curve stratified by predicted stroke risk using the decision threshold maximizing the Youden index on the validation set. (B) Confusion matrix at the same threshold. Parenthetical values indicate true negative rate, false positive rate, false negative rate, and true positive rate. Recall, precision, and geometric mean (G-mean) were 0.63, 0.09, and 0.69, respectively.

• 

  Central Illustration. Deep Learning for 1-Year Stroke Risk Analysis A schematic of Forecasting Strokes via Interpretable Independent Networks (ForeSIIN), a novel deep learning framework leveraging gated recurrent units (GRUs) to predict 1-year ischemic stroke risk in patients with atrial fibrillation (AF). The model captures temporal feature importance and provides interpretable predictions through feature-specific sequential modeling.