When Dassault Systèmes quietly published its ENRICHMENT Playbook in early 2025—five years of FDA collaboration compressed into a practical guide for running medical device trials entirely in silico—few industry watchers seemed to notice. Which is strange, really, because the playbook landed just months after the FDA finalized its guidance on computational modeling credibility. Together, the documents signaled something the industry had been anticipating but hadn't quite believed: regulators were done tiptoeing around digital twins. They wanted scale.
The problem? The infrastructure to deliver it barely exists.
What's emerged instead is a sprawling land grab. Startups pitch unification. Legacy giants claim pieces of the stack—data pipes here, simulation engines there, validation frameworks somewhere else. Market forecasters, never shy about big numbers, project the healthcare digital twins sector will swell from $2.69 billion in 2024 to nearly $60 billion by 2030, a 68% compound annual growth rate according to MarketsandMarkets. Even the more conservative Coherent Market Insights sees $6.80 billion by 2032, growing at 25.7% annually. The in-silico clinical trials market—just the pharma and medtech slice—is expected to nearly double from $3.76 billion in 2023 to $6.39 billion by 2033.
These aren't just aspirational figures. They reflect a dawning recognition among regulators and payers that precision medicine's future hinges on something counterintuitive: predicting what happens to a patient, a device, a drug regimen before anyone takes a real-world risk.
How Regulators Opened the Door
For years, the digital twin conversation in healthcare circled around scientific plausibility. Could you really model a beating heart with enough fidelity to predict how a stent would behave? Increasingly, the answer is yes. But that was never the bottleneck.
The bottleneck was regulatory acceptance. And that dam broke—quietly, bureaucratically—over the past 18 months.
The FDA published its final guidance on "Assessing the Credibility of Computational Modeling and Simulation in Medical Device Submissions" in January 2024, formally endorsing the ASME V&V 40 credibility framework that industry had been refining for years. Then, in December 2024, the agency finalized its Predetermined Change Control Plans guidance for AI-enabled devices, creating a pathway for iterative updates to machine learning-powered twins without re-clearing every version. A draft guidance on AI device lifecycles followed in January 2025. Across the Atlantic, the European Medicines Agency launched a consultation on mechanistic modeling guidelines—PBPK, physiologically-based biopharmaceutics modeling, quantitative systems pharmacology—running through May 2025. The EU AI Act, which entered force in August 2024, established high-risk AI requirements that apply to medical twins, forcing rigor around datasets, risk management, and human oversight.
Regulators aren't just accommodating computational models anymore. They're building guardrails so companies can rely on them. The FDA's Catalog of Regulatory Science Tools now includes VICTRE, an in-silico mammography trial pipeline that mirrored a human trial's outcome back in 2019 and has been updated through 2025. The ICH M15 draft guidance, released in 2024, codifies Model-Informed Drug Development principles. The FDA's MIDD Paired Meeting Program runs through 2027 under PDUFA VII. The UK's MHRA introduced an AI-Airlock sandbox and is aligning with the FDA and Health Canada on PCCP principles.
The message, blunt and clear: if you can prove your computational model is credible, agencies will use it to make decisions.
The Unglamorous Data Problem
Building a credible digital twin of a human—or even just a human organ system—requires longitudinal, multimodal data at a scale and fidelity most companies simply don't possess. Electronic health records are fragmented across systems. Imaging sits locked in DICOM archives. Wearables generate biometric streams with no standard schema. Clinical trial data lives in proprietary databases. Motion-capture systems, genetic panels, lab results: separate pipes, all of them.
Enter TEFCA and FHIR. The Trusted Exchange Framework and Common Agreement, which designated its first Qualified Health Information Networks in December 2023, is expanding steadily through 2025 and 2026. Oracle Health secured QHIN status in November 2025, joining a growing roster of nationwide data exchange hubs. The Sequoia Project's FHIR roadmap, now in its second version, lays out staged QHIN-to-QHIN API exchange. The US Core FHIR Implementation Guide continues mapping to USCDI versions 5 and 6, with FHIR R4 remaining the dominant standard across EHRs in 2025 and 2026.
Cloud providers are building the middleware. Google Cloud's Healthcare API offers FHIR and DICOM stores with built-in de-identification pipelines, letting companies assemble research-grade datasets without directly touching protected health information. Oracle Health's QHIN positioning means it can aggregate, cleanse, and normalize data at scale.
But aggregation is table stakes. The harder problem—perhaps the problem—is validation: proving your twin is accurate enough to guide an actual clinical or regulatory decision.
Simulations That Matter
Once you have the data, you need a model that does something useful with it. The stack splits roughly into mechanistic models—physics-based, grounded in first principles—and data-driven twins that learn patterns from historical outcomes.
Mechanistic models dominate the drug development side. Certara's Simcyp platform, which uses physiologically-based pharmacokinetics to simulate drug absorption, distribution, metabolism, and excretion, became the first and only EMA-qualified PBPK software in August 2025. Dassault Systèmes' Living Heart Project, now entering a beta phase for an AI-powered cardiac twin announced in February 2025, uses multiscale biomechanical finite element analysis to predict how devices interact with heart tissue. NVIDIA's Omniverse, Isaac for Healthcare, and Holoscan platforms—collectively branded as "physical AI" infrastructure—let companies simulate entire anatomies and clinical environments, then deploy models to edge devices. NVIDIA has inked partnerships with GE HealthCare, Barco, and Siemens to embed these capabilities into imaging and surgical workflows.
Data-driven twins shine when you need counterfactuals—answers to the question, "What would have happened if we hadn't intervened?" Unlearn's digital twin generators, described in a 2024 paper, create synthetic control patients by modeling what would have occurred had someone not received treatment. The company's PROCOVA methodology earned EMA qualification and is being used in multi-year collaborations with Merck KGaA to shrink control arms by 30% or more. QuantHealth claims its patient-level trial simulations—spanning more than 350 trials with up to 90% predictive accuracy—can save top-10 pharma companies $31.4 million per avoided trial. Sanofi Ventures invested in QuantHealth in October 2025.
The validation burden, though, falls squarely on companies to prove their models meet ASME V&V 40 credibility standards. That means documenting how well the simulation matches real-world outcomes, quantifying uncertainty, and showing the model's limitations.
It's not glamorous work. But it's the difference between a research toy and regulatory-grade evidence.
The Companies Crossing the Chasm
A handful of companies have made the leap from promising technology to actual revenue. HeartFlow, which uses computational fluid dynamics to calculate fractional flow reserve from CT scans, has now been used in more than 500,000 patients. The company secured a new 510(k) clearance for Plaque Analysis in September 2025 and expanded payer coverage. It's the closest thing the field has to a durable precedent: an organ-level twin that's reimbursed, guideline-recommended, and integrated into standard care.
PrediSurge is running prospective studies on digital twins for EVAR and TAVI planning, backed by EU funding. Virtonomy's "v-Patients" for structural heart device design claim FDA and EU acceptance as digital evidence, though regulatory pathways for purely in-silico benchtop studies remain murky. Mesh Bio secured HSA approval in Singapore for its HealthVector Diabetes software in October 2023 and is piloting the twin in hospitals to predict chronic kidney disease risk in Type 2 diabetes patients. Twin Health published outcomes in NEJM Catalyst in August 2025, showing its metabolic digital twin reduced GLP-1 use in claims data.
The clinical trial augmentation side is accelerating. A diabetes study published in npj Digital Medicine in early 2025 used a digital twin-driven co-adaptation routine for automated insulin delivery, improving time-in-range and HbA1c after receiving an FDA-approved investigational device exemption. Unlearn used its twins to help interpret a Phase 2a Alzheimer's signal in 2025. An npj Digital Medicine review in 2024 found digital twins effective across 80% of 45 outcomes in various disease areas. A year-long real-world Type 2 diabetes intervention published in Scientific Reports in 2024 showed improved glycemic control and reduced medications.
These aren't proofs of concept anymore. They're revenue-generating products with regulatory clearances, payer contracts, and clinical adoption curves.
The Infrastructure Bet
Which brings us back to the infrastructure race. Dassault, Siemens Healthineers, and NVIDIA have staked claims to different layers: Dassault on cardiac simulation and virtual human ambitions, Siemens on hospital workflow twins, NVIDIA on the compute and deployment stack. But these are horizontal platforms trying to bolt healthcare onto industrial digital twin frameworks. The fit isn't always natural.
Startups see an opening. Mantis Biotechnology, a Y Combinator Winter 2026 company founded in 2025 by Georgia Witchel, describes itself as "infrastructure powering human-in-computer models." The New York-based team raised roughly $6.3 million in seed funding led by Decibel Partners, according to trade press, and positions its platform as unifying fragmented multimodal data—motion capture, biometrics, imaging, training logs—into validated digital twins. A Wellfound job page frames it as a unified biomedical testing and regulatory platform linking CAD to simulation to verification to FDA Q-submissions: "Palantir for biomedical." Initial focus is sports and human performance, with expansion into medical device workflows planned.
The pitch, essentially: someone needs to own the connective tissue between data ingestion, simulation, credibility assessment, and regulatory submission. Incumbents own pieces. Startups like Mantis are betting on unification.
Whether that bet pays off is another matter.
What Happens Next
The next two years will clarify which strategies win. Wider adoption of PCCPs will normalize iterative AI twin updates. The EMA's mechanistic modeling guidance and ICH M15 will codify model-informed drug development. The FDA's Regulatory Science Tools catalog will expand with more organ and device phantoms, building credibility precedents for in-silico evidence. TEFCA's staged QHIN-to-QHIN FHIR exchange will ease data unification for longitudinal twins.
Technically, the convergence of physics-based models and machine learning—fast finite element emulators, physics-informed self-supervised learning—will push patient-specific twins into time-constrained clinical decisions: the cath lab, the operating room. UK Biobank-derived tools are releasing open meshes and models from more than 55,000 MRIs. Automated calcification meshing and generative shape modeling are accelerating virtual cohort generation for in-silico trials.
Organ and tissue twins—cardiac, vascular, orthopedic—will scale before whole-body twins, probably. Hospital workflow twins will mature alongside them, optimizing OR scheduling and device allocation rather than predicting individual outcomes. RCT augmentation and synthetic control approaches will expand in Phase 2 and Phase 3 trials where EMA and FDA precedents already exist.
The upside projections assume regulatory clarity and payer recognition arrive in tandem, which is never guaranteed. Markets have a way of inflating expectations and then deflating them just as fast. But the more revealing signal might be this: startups, pharma giants, device makers, and tech platforms are all building toward the same future—one where the default question isn't "Should we simulate this?" but "Whose simulation engine should we use?"
That's not hype. That's infrastructure taking shape in real time.
