The problem isn't the simulation. It's never really been about the simulation.
Georgia Witchel learned this the hard way, building physics engines for surgical robots before pivoting her startup toward something less glamorous but perhaps more essential: the messy, unglamorous work of wrangling healthcare data into something a computer model might actually trust. "Roughly 80 percent of clinical trials face delays due to data inaccuracies," she says, citing figures her company, Mantis Biotechnology, uses in pitch decks—numbers that echo frustrations across the sector, though independently verified sources for this specific statistic remain elusive.
The broader market tells a seductive story. Mordor Intelligence estimates the global healthcare digital twins market climbed from around $2.8 billion recently toward a trajectory that could reach $14 billion or more within five years—a compound annual growth rate north of 30 percent. Venture dollars are flowing. Regulatory frameworks are evolving. A handful of companies have crossed the threshold from prototype to commercial product, securing FDA clearances and, crucially, reimbursement codes.
Yet beneath those headlines, a different narrative is taking shape. The real bottleneck, a growing cohort of founders and researchers argue, isn't simulation fidelity or computing horsepower. It's data—fragmented, inconsistent, siloed across incompatible systems, and stubbornly resistant to the kind of standardization that machine learning models and physics-based simulations demand.
When Simulation Meets the Real World
The companies furthest along share an instructive pattern. They didn't try to boil the ocean. HeartFlow, for instance, focused relentlessly on one thing: using CT scans to compute fractional flow reserve for coronary artery disease, sparing patients invasive catheterization when possible. The strategy paid off. With an effective date of January 1, 2024, the company secured a CPT Category I code—the kind of reimbursement milestone that signals payers are willing to foot the bill for computational predictions of coronary physiology. Medicare Administrative Contractors followed suit through late 2024, and commercial insurers including Aetna added coverage in the months that followed.
FEops pursued a similar path with its HEARTguide system, earning FDA De Novo authorization in 2021 for patient-specific left atrial appendage occlusion planning. Philips received 510(k) clearance for SmartHeart, an AI-powered cardiac MR planning tool, on March 6, 2026. These aren't research curiosities anymore. They're commercial products navigating regulatory pathways, demonstrating clinical utility, and—critically—getting paid.
The trial optimization angle is gaining traction too, though the value proposition is trickier to prove. Unlearn.AI has been pursuing regulatory qualification with the European Medicines Agency for its approach to generating digital twins of control-arm patients. The pitch: use historical data and machine learning to create synthetic control cohorts, potentially shrinking the number of real patients randomized to placebo. In March, the company announced support for an ALS clinical study using AI-generated twins to strengthen statistical signals. Collaborations with AbbVie and Johnson & Johnson were presented at the Alzheimer's Association International Conference in 2024.
Whether regulators will embrace synthetic controls at scale remains an open question. But the fact that serious conversations are happening suggests the ground is shifting.
Operational twins occupy a different niche entirely. Siemens Healthineers rolled out ActExcell Operational Twin for radiology departments recently, using AI to optimize workflow and resource allocation—less about individual patient physiology, more about keeping the machinery of a hospital running efficiently.
Regulatory Tailwinds (and Headwinds)
Several forces are converging to make this moment feel different. The FDA's ongoing Model-Informed Drug Development program continues to evolve. A draft ICH M15 guideline on general principles for MIDD appeared in late 2024. The FDA Modernization Act 2.0, enacted in late 2022, explicitly permits nonclinical testing methods—including in silico approaches—as alternatives to mandated animal studies where scientifically justified.
It's not a blank check. But it opens doors.
For medical devices, the FDA updated its computational modeling and simulation guidance in late 2023. The agency's ongoing symposia with the Medical Device Innovation Consortium—most recently in spring 2024—are refining standards for model credibility. The European Medicines Agency held a multi-stakeholder workshop in late 2025 focused on reporting and qualification of mechanistic models for regulatory assessment, signaling Europe's intent to formalize pathways for digital evidence.
Data access is improving, at least on paper. The Office of the National Coordinator for Health Information Technology finalized the HTI-1 rule in late 2023, introducing transparency requirements for predictive algorithms embedded in certified EHR systems. Deadlines extending into early 2026 mandate standards like SMART on FHIR and USCDI v3. Information blocking enforcement has accelerated, nudging providers and vendors toward more open data flows.
Physiologically based pharmacokinetic modeling offers a useful proxy for regulatory acceptance. Recent analyses of FDA submissions indicate PBPK was used in roughly a quarter to a third of new drug and biologics applications over a five-year window, with adoption climbing in the biologics division across gene therapies, vaccines, and cell therapies. These are mechanistic models of drug distribution and metabolism—simpler than whole-organ digital twins, perhaps, but they demonstrate regulators are comfortable with computational predictions when properly validated.
The Infrastructure Layer Nobody Talks About
This is where Witchel's bet gets interesting. Mantis doesn't build organ simulations or patient avatars directly. It's tackling what she describes as a persistent infrastructure gap: clinical and biomedical data live in fragmented silos—electronic data capture systems, clinical trial management systems, lab vendors, omics platforms. Building a digital twin or training a machine learning model typically requires bespoke data wrangling for each study or application.
Mantis proposes a "domain-aware data platform" that encodes biological and clinical meaning into reusable, canonical datasets with lineage back to source systems. Faster queries, cross-study analytics, partner benchmarking, AI-readiness—the usual promises.
It's an infrastructure play in a market dominated by entrenched players. Veeva's Vault Clinical and CDMS platforms have gained momentum, with multiple top-20 pharma migrations reported recently. Medidata, now part of Dassault Systèmes, anchors clinical trial data for a large swath of the industry. Oracle and Databricks are pushing healthcare-specific data lakehouse architectures.
Mantis, with a small team as of its Y Combinator listing, isn't competing head-to-head on scale. The company is betting that general-purpose platforms lack the domain encoding and lineage depth that digital twin developers need. Independent validation is scarce. A LinkedIn post mentioned a seed round led by Decibel Partners, though without formal announcement the details remain unconfirmed. Demo day presentations typically precede official fundraising disclosures anyway.
Whether customers are willing to adopt a new data layer—especially one from a startup—remains to be seen. Incumbents aren't standing still. And the promised reusability and lineage have to deliver measurable time-to-insight, not just architectural elegance.
Where the Money Is (and Isn't)
Cardiology leads the charge, for understandable reasons. The organ is complex but bounded. Imaging modalities are mature. Clinical workflows are well-defined. And cardiovascular disease is a massive market with clear payer incentives to avoid invasive procedures when computational alternatives can deliver comparable diagnostic accuracy.
Dassault Systèmes entered an advanced beta phase of its Living Heart virtual twin platform recently, partnering with NVIDIA to accelerate simulation workflows. Jensen Huang, NVIDIA's CEO, stated in a blog post that "everything will be represented in a virtual twin," underscoring the company's broader push into healthcare modeling. That's enterprise-scale infrastructure—simulation software, high-performance computing, AI-accelerated surrogates—aimed at device makers and researchers developing patient-specific cardiac models.
Oncology is emerging as another sweet spot, though the biology is messier. Twin Health raised a $53 million Series E in mid-2025 for its metabolic health platform, which uses continuous glucose monitors and other sensors to create what the company calls a "Whole Body Digital Twin" for precision nutrition and diabetes management. Clinical outcome claims include A1C reductions, but the model is less about physics-based simulation and more about data-driven personalization. Whether that qualifies as a true digital twin is a definitional debate best left to purists.
On the organ-on-a-chip front, Hesperos published findings in Advanced Science demonstrating what it called the first "true digital twin capability" using microphysiological systems. The work involves coupling physical organ chips with computational models to predict human disease responses—a hybrid approach that bridges in vitro assays and full in silico twins.
Academic work is advancing rapidly. Recent preprints describe frameworks for generating 4D cardiac digital twins directly from ECG data using deep learning surrogates, and graph neural network approaches to accelerate cardiac mechanics simulations. These are research prototypes, not clinical-grade tools. But they signal the direction: hybrid models that blend physics-based equations with machine learning to achieve near-real-time predictions.
The Reality Gap
The National Academies of Sciences, Engineering, and Medicine published a foundational report identifying core research gaps for digital twins: mathematical rigor, computational efficiency, validation methodologies, cross-sector collaboration. Nature Medicine ran an editorial on digital twins for personalized care with cautious optimism—precision medicine potential is real, but complexity and data integration remain formidable.
That's the polite academic version. The blunt version: simulation is only as good as the data feeding it, and healthcare data remains messy despite FHIR mandates and information blocking rules. Practical interoperability is uneven. Data quality varies wildly. And while transparency requirements aim to demystify algorithmic black boxes, they also add compliance burdens that smaller players struggle to shoulder.
Reimbursement will stay case-specific for the foreseeable future. The HeartFlow and FEops examples show that payers will cover digital twin-enabled diagnostics and planning tools when clinical evidence supports them and existing CPT codes or De Novo pathways accommodate them. But there's no blanket reimbursement pathway for "digital twins" writ large. Each application has to prove its value proposition, study by study, indication by indication.
The EU's EDITH consortium released a roadmap charting a "Virtual Human Twin" platform strategy spanning oncology, cardiovascular care, ICU management, osteoporosis, and brain modeling. That suggests European procurements and pilot programs in the pipeline. Whether those materialize into commercial traction is another question entirely.
What Happens Next
The Synopsys-Ansys merger, which closed in mid-2025, could catalyze new simulation workflows when integrated capabilities roll out. Combining electronic design automation with multiphysics modeling might accelerate medical device R&D pipelines and organ-level simulations. Or it might simply consolidate tooling without fundamentally changing adoption curves—mergers often promise more than they deliver.
For now, the market is expanding. Cardiology and oncology are the near-term sweet spots. Trial optimization is gaining traction. Operational twins are proving useful in radiology and potentially other service lines. And a handful of regulatory clearances and reimbursement wins suggest the technology is crossing from research curiosity to commercial viability.
But the path from prototype to ubiquity is rarely smooth, particularly in healthcare. Validation gaps persist. The "reality gap" between simulation and clinical ground truth remains a research challenge, as recent reviews on credibility frameworks note. Data fragmentation is real, even with mandates in place.
The companies that succeed will likely be the ones that pick specific problems, demonstrate clinical utility, navigate regulatory pathways, and—crucially—figure out how to wrangle messy, heterogeneous data into something a model can trust. Witchel's background includes a computer science degree from Harvey Mudd and a stint as a founding engineer in sports tech. Her previous Y Combinator-backed venture, Louiza Labs, worked on a physics engine for digital twins in autonomous robotic surgery and simulated FDA trials. The pivot to data infrastructure suggests she's learned where the leverage points are.
Whether that insight translates into a sustainable business is the open question. Infrastructure bets are notoriously difficult—high upfront investment, long sales cycles, entrenched competition. But if the digital twin market really is headed toward the forecasted growth, someone has to solve the data problem. It might as well be the founders willing to do the unglamorous work that nobody else wants to talk about.
That's the race now. Not just who builds the best digital twin, but who builds the data infrastructure to feed it. Perhaps more importantly, who convinces customers that the infrastructure is worth paying for.
