AI Predicts Blood Clot Growth from Sparse Data

This research introduces a novel computational framework designed to predict blood clot growth, or thrombosis progression, even when clinical data is sparse and key model inputs are difficult to measure. The framework leverages latent neural differential equations to infer unknown parameters of a multiphysics blood-clotting model from limited early observations of clot size and known biochemical inputs. The study compared seven different probabilistic methods, including various forms of neural differential equations and other baselines. Stochastic Neural Ordinary Differential Equations (SNODE) emerged as the top performer, excelling at both inferring the unmeasured tissue-factor parameter and accurately forecasting subsequent clot growth trajectories. Stochastic Neural Functional Differential Equations (SNFDE) also performed strongly. The accuracy of predictions improved with more available observations, though longer forecasting horizons naturally increased uncertainty. This work provides a promising foundation for developing personalized thrombosis models, which could significantly impact clinical applications.