<p>Venous thromboembolism (VTE) is a major global cause of death, with recurrence risk rising after inadequate treatment. Existing VTE recurrence prediction models often overlook the phenomenological aspects of blood clot formation. This study aimed to predict Recurrent VTE (RVTE) by combining an artificial neural network (ANN) with a phenomenological clot formation model. A spatiotemporal model of clot formation was proposed and simulated using hematological data from 235 patients. Principal component analysis (PCA) and partial least squares (PLS) regression were applied to reduce the number of significant variables, and multiple ANN structures were trained for RVTE predictions. The dataset was initially divided into training and validation subsets to avoid data leakage. No data from the validation subset was used to build the model, ensuring the model's integrity. The validation conducted in this study corroborates the model's good generalizability. The spatiotemporal model effectively differentiated patients based on clot formation characteristics. The application of PCA and PLS reduced the original variables by over 99.99%. The best-performing model achieved, for unseen observations, an area under curve of 0.9689, accuracy of 0.9149, sensitivity of 1.0000, specificity of 0.8919, and F1-score of 0.8333, aligning with established literature values. This methodology outperformed nine other machine learning models commonly used for binary classification. A significant advantage of this approach is its reliance on only six clinical variables for prediction. This study demonstrates that mathematical methods can reveal insights into the spatiotemporal evolution of blood clots, presenting a promising avenue for RVTE prediction. The model could serve as a clinical decision-support system for RVTE treatment.</p>