A Hybrid Mechanistic–Machine Learning Framework for Patient-Specific Thrombin Generation and Interpretable Prediction of Recurrent Venous Thromboembolism

Predicting recurrence after a first venous thromboembolism episode remains difficult: existing risk scores use few variables and generalize poorly, whereas machine learning often lacks mechanistic interpretability. We developed a hybrid framework combining an artificial neural network with a mechanistic model of the coagulation cascade expressed as a system of ordinary differential equations, optimized using metaheuristic optimization algorithms, to infer patient-specific kinetic parameters from 39 clinical and hematological variables in 235 patients with a first venous thromboembolism (49 recurrences, 20.9%), divided into a single 70% training and 30% independent test partition, with 15 recurrent events in the test partition. The estimated kinetics were propagated into a spatiotemporal model of clot growth described by partial differential equations, and the resulting clot-size variable, together with hematological features, was used to train thirteen classifiers with interpretation based on Shapley additive explanations. Among the clinically compliant configurations, particle swarm optimization produced the most balanced hybrid model (area under the curve of 0.851 and F1-score of 0.800 in the test partition). Patient-specific parameterization significantly altered simulated clot-size predictions relative to population-averaged kinetics and was associated with improved discrimination of recurrent venous thromboembolism in the full cohort. A histogram-based gradient boosting classifier achieved the highest point-estimate predictive performance (area under the curve of 0.872 and F-measure of 0.750 in the test partition), with clot size emerging as the most influential feature in model interpretation. These findings support the potential of patient-specific mechanistic modeling to generate physiologically informed features for recurrent venous thromboembolism prediction while preserving interpretability, warranting external validation before clinical application.