Why Synthetic Humans Could Be Medicine’s Most Important Controversy

1. Headline & intro

Healthcare has a data problem that no amount of hospital digitisation seems able to fix: the most interesting cases are exactly the ones you’re not allowed – or able – to learn from at scale. Into that gap steps Mantis Biotech, a New York startup building high‑fidelity “digital twins” of humans by fusing large language models with physics engines. According to TechCrunch, its first traction is in elite sports, but the ambition clearly points at medicine. In this piece we’ll look past the funding headline and ask a harder question: are synthetic humans the way out of healthcare’s data deadlock, or just the next beautifully rendered illusion?

2. The news in brief (what happened)

As reported by TechCrunch, Mantis Biotech is developing a platform that creates physics‑based “digital twins” of the human body to generate synthetic biomedical data. The system ingests heterogeneous sources such as textbooks, motion‑capture feeds, biometric sensors, training logs and medical imaging. An LLM‑driven layer routes and validates these inputs, which are then processed through a physics engine to build detailed models of anatomy and movement.

These models can in turn be used to simulate edge cases that are poorly represented in real‑world datasets, such as rare conditions or unusual anatomies. According to TechCrunch, the company has initial commercial traction in professional sports, including at least one NBA team, where digital twins are used to analyse performance and injury risk. Mantis recently raised a $7.4 million seed round led by Decibel, with participation from Y Combinator, Liquid 2 and angel investors, and ultimately wants to expand into preventative healthcare and pharmaceutical research.

3. Why this matters

The core promise here is not better animation; it’s a potential escape hatch from healthcare’s toxic mix of data scarcity, fragmentation and regulation.

Who stands to benefit first?

• Elite sports and performance labs get highly customised models of individual athletes without waiting years to accumulate enough comparable injuries or outcomes. That’s a huge edge in contracts, training and rehab.
• Pharma and medtech R&D could use digital twins to explore dosing strategies, device designs or surgical procedures in silico before touching a real patient, shrinking early‑stage trial costs.
• Hospitals and researchers gain a way to train models on synthetic cohorts that emulate real populations without directly exposing patient records.

But there are potential losers:

• Traditional data brokers and some EHR vendors could see their leverage eroded if realistic synthetic data becomes good enough that buyers no longer need raw patient data at scale.
• Patients and clinicians risk being affected by decisions based on virtual evidence that may not fully capture biological or social complexity.

The immediate implication is strategic: if platforms like Mantis work, the bottleneck in medical AI shifts from data access to model validity. The limiting question becomes less “Can I get the data?” and more “Can I prove my synthetic humans behave like real ones when it matters?”

This also subtly changes the competitive landscape. The most powerful healthcare AI companies may be those that master multi‑modal simulation (language, images, physiology, behaviour) rather than those that merely hoard hospital data. For incumbents built on owning large datasets, that’s an uncomfortable prospect.

4. The bigger picture

Mantis is not the first to talk about “digital twins” in health, but its approach is emblematic of a broader convergence: generative AI, biomechanics and synthetic data.

In recent years, synthetic health‑data companies have emerged to tackle privacy and access problems by statistically mimicking patient populations. Separately, industrial players such as Siemens and Dassault Systèmes have promoted organ‑ and system‑level twins for cardiology or orthopaedics. What’s different now is the LLM‑centred glue: using language models to orchestrate, label and reconcile messy inputs, and then pushing that structure into a physics engine.

Two important industry trends intersect here:

1. Regulators warming to in silico evidence. Drug and device regulators in the US and Europe have begun accepting simulation data as part of submissions in narrow areas, such as cardiac devices or dosing optimisation. If digital twins get good enough, they could progressively replace parts of animal studies or control arms in human trials.
2. The shift from population averages to individuals. Consumer wearables, continuous glucose monitors and camera‑based motion analysis are generating torrents of personalised signals. Digital twins offer a conceptual way to tie those signals to actionable predictions about this body, not just a statistical average.

Historically, every wave of “virtual humans” – from early anatomical CD‑ROMs to VR surgery trainers – has overpromised. The computational models were brittle, the inputs crude and, crucially, they weren’t easily kept in sync with the real person. Mantis and its peers are effectively arguing: with modern sensors plus AI, that sync problem is finally tractable.

Whether that’s true will determine if digital twins become critical infrastructure for medicine or another metaverse‑style buzzword.

5. The European / regional angle

For Europe, digital twins land at the intersection of opportunity and regulation.

On the opportunity side:

• Privacy‑preserving innovation. Under GDPR and national health‑data laws, accessing and linking patient data is slow and politically sensitive. Credible synthetic cohorts could let European hospitals and startups experiment without constantly renegotiating consent and data‑sharing agreements.
• Public health and ageing. Ageing populations and constrained budgets in countries like Germany, Italy and Spain make virtual experimentation attractive. Health systems could stress‑test screening strategies or care pathways on twin populations before spending billions.
• Industrial strengths. Europe already has deep expertise in biomechanics, medical devices and simulation (think university hospitals in Scandinavia, German engineering, French imaging). Digital twins are a natural extension.

On the constraint side:

• The EU AI Act will treat high‑risk medical AI as heavily regulated, regardless of whether it’s trained on real or synthetic data. Providers of twin‑based tools will face stringent documentation, risk management and transparency requirements.
• Under GDPR, even the process of generating twins may count as processing sensitive health data if it starts from identifiable records. Pseudonymisation is not a silver bullet; true anonymisation is hard.

European players – from university hospitals in Leuven or Charité to startups in Barcelona or Ljubljana – could actually be ideal testbeds: they understand regulatory nuance and have strong ethics governance. But they will demand much more evidence and auditability than a sports team signing up to shave a millisecond off a sprint.

6. Looking ahead

Over the next three to five years, expect digital twins to move through three overlapping phases.

1. Performance and training niche. Sports, rehabilitation clinics, defence and space agencies will continue to be early adopters. The value proposition is clearest where physical performance, injury risk and biomechanics dominate.
2. R&D simulation for pharma and medtech. As companies like Mantis mature their physics models and validation stories, pharma labs will increasingly use twins to explore trial designs, virtual patient stratification and device optimisation. Regulators may first accept twin data as supportive evidence, not a replacement for trials.
3. Clinical decision support – cautiously. The hardest and slowest step will be integrating twins into routine care: predicting complications, tailoring surgery or guiding rehab. Here, liability, explainability and bias questions become existential.

Key signals to watch:

• Do any top‑tier medical journals publish prospective studies where twin‑based predictions materially change outcomes?
• How do regulators like the FDA or EMA write guidance around simulation in submissions? Are “virtual control arms” explicitly encouraged or merely tolerated?
• Do insurers or public health agencies start paying for twin‑based risk assessments or planning tools?

Risks are non‑trivial: overreliance on models that under‑represent minorities; misuse by employers or insurers to profile individuals; and the possibility that physics‑based twins created from biased data simply encode a different flavour of bias. On the upside, done well, they could decouple biomedical progress from constant large‑scale surveillance of real patients.

7. The bottom line

Digital twins like those pursued by Mantis Biotech represent a genuine shift: from scraping ever more hospital records to simulating humans with growing fidelity. The idea is powerful and, in data‑constrained healthcare systems like Europe’s, deeply attractive. But the centre of gravity moves from access to accountability: can we trust what synthetic humans tell us, and who is responsible when they are wrong? As these tools move from basketball courts into hospitals, regulators, clinicians and patients will have to decide how real “virtual humans” are allowed to become.