Predicting blood clot growth from sparse post-onset measurements with latent neural differential equations

2026-08-08

AI Summary

This study addresses the challenge of clinical thrombus modeling, which is hindered by sparse patient-specific data and the inaccessibility of key biochemical factors. It introduces, for the first time, latent-variable neural differential equations to model thrombus dynamics, leveraging sparse early-stage thrombus size measurements and partially known factors to jointly infer unknown tissue factor parameters and predict subsequent growth trajectories. The authors systematically evaluate seven probabilistic methods—including SNODE and SNFDE—on multiphysics coagulation simulation data. Results demonstrate that SNODE achieves the best performance in both parameter inference and dynamic prediction, with SNFDE ranking second and significantly outperforming non-differential models. Prediction accuracy improves with more observations yet degrades over longer forecasting horizons, underscoring the framework’s potential to overcome traditional models’ reliance on dense data.

Abstract

Computational models of blood clotting improve understanding of thrombus formation, but their clinical application remains limited because many model inputs are difficult to measure and patient-specific data are often sparse. We present a computational framework based on latent neural differential equations that infers unknown model parameters from sparse measurements and forecasts thrombosis progression. We demonstrate the framework using data generated from a multiphysics blood-clotting model in which clot growth is governed by the coagulation cascade and diffusion. Four known biochemical inputs (fibrinogen and factors IX, VIII, and V), together with sparse early clot-size observations, are used to infer the tissue-factor parameter and predict subsequent clot growth. We compare seven probabilistic methods: stochastic neural ordinary differential equations (SNODE), stochastic neural functional differential equations (SNFDE), a latent neural-process baseline, a monotone probabilistic deep ensemble, empirical trajectory retrieval, PCA-ridge Gaussian posterior, and Gompertz-curve retrieval. SNODE achieved the best performance in inferring the unknown input and forecasting future clot-growth trajectories. SNFDE performed similarly and consistently outperformed the other non-differential models. Prediction accuracy improved as more observations became available, whereas longer forecasting horizons increased uncertainty and decreased accuracy. Latent neural differential equations thus effectively combine parameter inference and clot-growth forecasting from sparse measurements, providing a promising foundation for personalized thrombosis modeling.

Problem

• thrombosis prediction
• sparse measurements
• parameter inference
• clot growth
• patient-specific modeling

Innovation

• latent neural differential equations
• thrombosis prediction
• sparse data inference
• stochastic neural ODE
• personalized modeling