A clot grows like a snowball rolling down a mountain: small at first, it picks up new layers, and predicting when it will turn into an avalanche is tricky. In the blood, a long chain of reactions, the coagulation cascade, unfolds, where one substance activates another, and the clot gains mass. The problem is that at the bedside, doctors see only a few numbers (say, fibrinogen levels or clotting factors) while the main trigger (tissue factor, which kicks off the whole process) often stays in the shadows. Without it, a clot growth model is like a weather forecast that ignores ocean temperatures.
The authors built a "smart observer"—a computational framework based on hidden neural differential equations. It takes the crumbs we have: concentrations of four blood proteins and a couple of early measurements of clot size, and reconstructs the missing tissue factor value. Then, guided by the physics of diffusion (how substances spread) and the biochemistry of the cascade, the model step by step draws how the clot will grow. Among the seven methods tested, two stood out as noticeably more accurate: stochastic neural ordinary differential equations (SNODE) and their "big brother" with memory — neural functional differential equations (SNFDE). They work like a translator who doesn't just memorize words but understands the grammar of the process, so they make fewer mistakes.
The practical takeaway is simple: the more measurements the model gets, the more confident its predictions. Conversely, trying to see far ahead increases uncertainty. This approach, for the first time, combines hidden parameter inference and clot growth prediction in one flexible tool, paving the way for personalized thrombosis modeling — where the computer tells the doctor how a clot will behave in a specific patient, not just the average one.
The blood coagulation cascade is one of the fastest amplification mechanisms in the body: a single molecule of tissue factor can trigger the formation of millions of fibrin molecules within minutes.