Every year brings predictions about healthcare data science that fail to materialize. AI was going to replace radiologists by 2020. Blockchain was going to transform health data exchange. Precision medicine was going to be routine by now. I am wary of trend predictions that confuse what is technically possible with what will actually be adopted.
That said, 2026 is different. Several trends that have been building for years are finally reaching the inflection point where they move from pilot projects to operational reality. Here is what I am watching closely.
Federated Learning Becomes Practical
For years, federated learning has been a theoretically elegant solution to a practical problem: how do you train machine learning models across multiple institutions without sharing raw patient data? The promise is compelling -- larger effective training sets, broader generalizability, preserved privacy.
The reality has been frustrating. Technical complexity, institutional inertia, and misaligned incentives have kept federated learning in the experimental phase at most organizations. But I am seeing this change.
Infrastructure maturity is accelerating adoption. Cloud platforms and specialized federated learning frameworks have reduced the technical barrier dramatically. What required custom engineering three years ago can now be implemented with standard tools.
Regulatory clarity is helping. Privacy offices that were skeptical of federated learning are becoming more comfortable as regulatory guidance clarifies how it fits within existing frameworks. The key insight is that federated learning does not eliminate privacy risk -- it changes the risk profile in ways that are often more manageable.
Competitive pressure is creating urgency. Health systems are recognizing that organizations with access to larger, more diverse patient populations will build better models. Federated consortia are emerging as a way for smaller systems to compete with larger ones.
I expect 2026 to be the year that federated learning moves from "interesting research project" to "something my organization is actually doing." The early applications will likely be in imaging and clinical NLP, where the model architectures are well-established and the value proposition is clearest.
Synthetic Data for Real-World Use
Synthetic patient data -- algorithmically generated records that preserve statistical properties while not representing real individuals -- has obvious appeal. You can share synthetic data externally without privacy risk. You can use it to train machine learning models. You can develop and test analytics pipelines without accessing real patient data.
The challenge has always been fidelity. If the synthetic data does not preserve the complex relationships present in real clinical data, models trained on it will not generalize. Simple approaches that treat variables as independent fail spectacularly when those variables are biologically or clinically linked.
What is changing is the sophistication of synthesis methods. Generative models trained on real patient data can now produce synthetic records that preserve subtle correlational structures, rare event rates, and longitudinal patterns that simpler methods miss.
Development and testing is the first practical use case. Most analytics teams spend significant time waiting for data access approvals before they can start building. Synthetic data environments let developers write and test code against realistic data while the governance process proceeds in parallel.
External collaboration is accelerating. Pharmaceutical companies, academic researchers, and technology vendors all want access to health system data. Sharing high-fidelity synthetic data is emerging as a way to enable these collaborations without the legal and ethical complexity of sharing real patient information.
Regulatory acceptance is growing. The FDA has signaled openness to synthetic data for certain use cases, particularly in device development and testing. This regulatory endorsement is building confidence across the industry.
I will note a caution: synthetic data is not a universal solution. For some applications, particularly those involving rare subpopulations or complex causal relationships, synthetic data may introduce biases that are difficult to characterize. The field needs better methods for validating that synthetic data is appropriate for a given use case.
AI Enters Health Technology Assessment
Health technology assessment -- the structured evaluation of clinical and economic evidence that informs coverage and reimbursement decisions -- has traditionally been a slow, manual, document-heavy process. HTA bodies review clinical trials, synthesize evidence, model cost-effectiveness, and produce recommendations that shape market access.
AI is beginning to change this in two ways.
AI-assisted evidence synthesis is streamlining HTA workflows. The manual work of screening studies, extracting data, and assessing quality is being augmented by AI systems that can process literature at scale. This is not replacing human judgment -- the final assessment still requires expert review -- but it is dramatically reducing the time from evidence generation to evidence synthesis.
AI-generated evidence is beginning to be accepted. More provocatively, HTA agencies are starting to grapple with how to evaluate evidence generated using AI methods. If a machine learning model identifies patient subgroups who benefit from treatment, how should that be weighted against traditional trial evidence? If a foundation model extracts outcomes from clinical notes, what validation is required before those outcomes can inform HTA?
This is contested territory. Some HTA bodies are enthusiastic about AI-generated evidence as a way to accelerate access to effective treatments. Others are skeptical about validation frameworks that are still maturing. The conversation is happening, though, in a way that seemed distant just two years ago.
For pharmaceutical and device companies, this creates both opportunity and risk. AI-generated evidence may accelerate market access for companies that invest in rigorous validation. But it also raises the bar for evidence quality -- HTA bodies will develop expertise in evaluating AI methods and will not accept poorly validated AI outputs.
The Capability-Building Imperative
Underlying all of these trends is a common thread: organizations need to build or acquire data science capabilities that did not exist five years ago. Federated learning requires distributed systems expertise. Synthetic data generation requires generative modeling expertise. AI-assisted HTA requires understanding of both AI methods and regulatory requirements.
The organizations that will capture value from these trends are the ones investing now in talent, infrastructure, and processes. This is not about buying a tool -- it is about building organizational muscle.
I see three patterns in organizations that are successfully building these capabilities:
They start with specific use cases, not technology investments. They identify a concrete problem -- accelerating clinical trial recruitment, improving claims adjudication, generating synthetic data for external partnerships -- and build the capability to solve that problem. The capability then generalizes to other applications.
They invest in partnerships. No single organization has all the expertise needed to implement these emerging methods. The most successful organizations are forming partnerships with academic centers, technology companies, and other health systems that complement their internal strengths.
They plan for iteration. The methods are evolving rapidly. Organizations that try to build a perfect solution upfront will be frustrated. The ones that are succeeding are building capabilities that can adapt as the field matures.
What I Am Skeptical About
Not everything that gets called a trend deserves the label. A few things I am skeptical will have major impact in 2026:
Blockchain for health data. Still waiting for the use case that justifies the complexity. Centralized solutions continue to work fine for most data exchange problems.
Fully automated clinical decision support. AI will continue to augment clinical decisions, but fully autonomous systems that make diagnostic or treatment decisions without human oversight are not ready for widespread deployment and will not be in 2026.
Universal interoperability. FHIR adoption is progressing, but true interoperability -- the ability to seamlessly exchange and use health data across systems -- remains a multi-year effort.
The Bottom Line
2026 is a year of maturation rather than revolution. The technologies and methods that have been building for years are finally becoming practical for real-world deployment. The organizations that have invested in building capabilities will capture disproportionate value. The ones that have waited for the technology to mature may find themselves behind.
Federated learning, synthetic data, AI in HTA -- these are worth watching not because they are novel but because they are reaching the point where watching becomes insufficient. It is time to build.