Inspiration
Medicine still learns many things by trying them on people. Drugs fail late in trials, and surgeons rarely get a second attempt at an operation. We wanted to know what it would take to test those decisions on a computational copy of a patient first. The idea of a whole-body “digital twin” is easy to state and hard to build, because the pieces exist at different scales and nobody has joined them into something that can be trusted. We wanted to work out how to join them and how to know whether the result is any good.
What it does
The project has two parts.
A research paper. It reviews the field, from the Physiome Project and whole-cell models to PBPK models, cardiac digital twins and regulatory in silico trials, and proposes a five-layer framework: cellular, tissue, organ, systemic and individualization. The cell is the building block, but the framework chooses model detail by explicit criteria rather than assuming it is always needed. It states four hypotheses that data could reject, and it includes:
a validation plan; a bound on how errors compound across layers; a subgroup fairness audit; a governance model for updating a twin after deployment; an example message format for the interface between layers.
A working prototype called Strand. It chains a cardiac cell model, a strip of heart tissue and a statistical personalization step. On synthetic patients it estimates a hidden parameter from one noisy reading and predicts how much a drug will lengthen the QT interval, with an uncertainty interval. It also audits accuracy across two patient groups and tests whether the layers agree when they exchange information.
How we built it
We started from the literature and checked the key sources against publisher and library records before relying on them. We then restructured the paper to a research-proposal format and revised it against outside reviews, which pushed us to add a novelty comparison, resolution criteria and a concrete first use case.
The prototype is a single self-contained web page in JavaScript. The cell layer is a published two-current cardiac model with the drug acting on the outward current. The tissue layer is a one-dimensional strand with a pseudo-ECG. Personalization is a grid-based Bayesian estimate. We ran the logic headlessly before publishing to catch errors. The interface is drawn on ECG graph paper, with the real time grid, so it looks like the subject and not a generic dashboard.
Challenges we ran into
The first prototype was the wrong thing. Our early virtual-cell simulator was tuned to look plausible, not fitted to data, so we dropped it from the paper rather than leave it as unsupported evidence. Scope. A “whole body” is not testable, so we narrowed the first milestone to one organ and one question. Model limits. The cardiac cell model stopped repolarizing when the outward current was too weak, which forced us to restrict the range of the sliders to where it works. The error bound did not always hold. At long exchange intervals the cell–tissue coupling itself became unstable, and the bound failed. That is consistent with the paper’s assumptions, but we had to say so instead of hiding it. Fairness. The simple textbook fairness criterion is the wrong target for physiological predictions, so we replaced it with comparable accuracy and calibration across groups. Honesty about evidence. Some references and several claims came from memory or summaries, so we marked what was verified and what still needs checking.
Accomplishments that we're proud of
A paper that states plainly that it is a framework and reports no experimental results, with hypotheses that could be proven wrong. A conditional error-propagation bound that turned a vague claim about composability into something testable, and a prototype that shows where it fails. A runnable demonstration where the chain from cell to tissue to personalized prediction works end to end on synthetic patients, with baselines and a group audit. A design that avoids the generic look: a figure drawn on real ECG paper, plain labels and a dark mode.
What we learned
The hard part of a digital twin is not any single model but the interfaces, the data and the evidence that it is reliable. Narrowing the scope made the project stronger. A credible single-organ twin is a better step than a vague whole-body claim. When the same model generates and fits the data, success proves little. The real test needs real, held-out data. Failures such as the unstable coupling and the weak group coverage taught us more than the clean results did.
What's next for Toward a Whole-Body DigitalTwin
Replace synthetic patients with retrospective clinical ECG data and published drug-potency data, and test the cardiac use case against real held-out outcomes and the defined baselines. Build an organ-only competitor so the hypothesis about cell-level detail can finally be tested. Replace the single-point sensitivity estimate in the bound test with a proper estimate with uncertainty. Write the full per-layer interface specification and a realistic cardiac cell model. Add the prototype as a clearly labeled synthetic demonstration in the paper, then extend to a second organ and the systemic layer.
Log in or sign up for Devpost to join the conversation.