Medical AI has a gap nobody really talks about openly. You can train a model on millions of hospital records, textbooks, and clinical notes, and it will still fall apart the moment you ask it about a rare disease or an unusual condition.
The data simply does not exist, at least not in any usable form. A New York startup called Mantis Biotech thinks it has found a practical way around this, and the answer looks nothing like what most people expect.
Mantis is building digital twins of the human body, virtual, physics-based replicas of real people that can be tested, stressed, broken, and studied without ever touching an actual patient.
What Exactly is a Human Digital Twin, and How Does Mantis Build One?
A digital twin, in simple terms, is a highly detailed virtual model of something real. Engineers have used them for aircraft engines and factory floors for years. Mantis is applying the same idea to human anatomy and physiology.
The platform pulls data from a surprisingly wide range of sources, medical imaging, biometric sensors, motion capture cameras, training logs, even textbooks. An LLM-based system then routes and validates all of that messy, scattered information, synthesising it into a clean dataset.
That dataset gets fed through a physics engine, which produces a high-fidelity model of how a specific human body moves, performs, and responds.
CEO and founder Georgia Witchel describes the physics layer as genuinely essential. Without it, you are essentially generating plausible-looking data that has no grounding in how the body actually works. With it, the synthetic data behaves realistically, because it has to follow the laws of physics.
Why Does Medicine Struggle So Much With Data Availability?
The short answer is access. Patient data is rightfully protected by strict privacy laws and ethical guidelines. Researchers working on rare diseases often cannot build representative datasets because there simply are not enough patients, and even those patients’ records are locked behind regulatory walls.
This leaves AI models stranded. They perform brilliantly on common conditions where data is plentiful and badly on everything else. Rare diseases, edge cases, unusual anatomical variations, these are precisely the areas where good predictive models could save lives, and they are also the areas where training data is hardest to find.
Witchel gave a pointed example. If you wanted to build a hand-pose estimation model for someone missing a finger, you would struggle to find a single publicly labelled dataset covering that situation. With Mantis, you just modify the physics model, remove the finger, regenerate, and you have your dataset almost instantly.
How are Sports Teams Already Using This Technology?
Mantis has found its earliest commercial success in professional sports, which makes a lot of sense. Sports organisations are obsessed with performance data, they have the budget to invest in emerging technology, and they already have motion capture and biometric infrastructure in place.
One of the startup’s current clients is an NBA team. The platform creates detailed digital representations of individual athletes, tracking not just how a player jumps today but how their jump mechanics have shifted across an entire season, cross-referenced against sleep data, training load, and cumulative physical stress. Coaches and medical staff can use that to spot early signs of injury risk before anything actually goes wrong.
Witchel used a more vivid analogy to explain the mindset she wants people to bring to the technology. Think of a three-year-old holding a Barbie doll by one leg and slamming it against a table.
That total freedom to experiment, to test things to destruction, that is exactly what digital twins are supposed to enable. You can simulate the worst-case scenario as many times as you want, because nothing real is at risk.
What is Next for Mantis Biotech and Synthetic Patient Data?
The company recently closed a $7.4 million seed round led by Decibel VC, with Y Combinator and Liquid 2 also participating. The funding goes toward hiring, marketing, and building out go-to-market capacity.
The roadmap beyond sports is notably ambitious. Mantis is working toward releasing a public-facing preventative healthcare product, and is actively developing tools for pharmaceutical labs and researchers running FDA trials.
The aim there is genuinely useful: delivering real-time insight into how individual patients respond to treatments, without requiring those patients to hand over sensitive personal data.
Synthetic patient data generated through digital twins could eventually let researchers run experiments that would otherwise be ethically or logistically impossible. That is a significant shift in how biomedical research could operate, and Mantis is, so far, one of very few companies trying to get there through physics-based modelling rather than pure generative AI.
The data problem in medicine is old and deep. The approach Mantis is taking to solve it is, at least, genuinely new.
