The first wave of diagnostic AI investment was largely modality-specific: a company that applied deep learning to radiology images, a company that processed pathology slides, a company that analyzed ECG waveforms. These are meaningful advances, and several portfolio companies — including our investments in Cleerly and Iterative Health — are building durable value within their respective modalities. But the pattern of investment I'm watching now is different. The next cycle is about synthesis across modalities, and it is going to produce a different kind of winner.
A cardiologist managing a patient with suspected heart failure doesn't think in modalities. She thinks in problems. The imaging findings, the biomarker values, the patient's symptom history, the medication record — all of this is relevant to her clinical judgment, and integrating it is the actual work of diagnosis. Current AI tools generally help with one input at a time. The cardiologist still has to synthesize across inputs herself, which is precisely the step that is most cognitively demanding and most prone to error. Single-modality AI assists one step of a multi-step reasoning chain. Multimodal synthesis could restructure the reasoning chain itself.
That's the commercial opportunity. But it comes with a harder clinical validation challenge. When a single-modality tool claims it can detect a finding in an image with a certain sensitivity and specificity, that claim is testable against a defined endpoint in a prospective study. When a multimodal synthesis tool claims it can improve overall diagnostic accuracy across a patient encounter, the endpoint is harder to define, the study design is more complex, and the regulatory pathway is less established. FDA has 510(k) pathways for single-modality diagnostic AI that are well-worn by now. The De Novo pathway for multimodal systems is being built in real time.
What I find most interesting — and what we are actively looking for in the deal pipeline — is not the synthesis application layer, but the infrastructure that makes multimodal synthesis possible. Specifically: data normalization and interoperability infrastructure. Most health systems have modality-specific data silos with incompatible formats, inconsistent metadata, and no unified patient-level representation that would allow a synthesis algorithm to operate across them. Building that infrastructure is genuinely hard work that requires deep health-systems knowledge, and the companies doing it well are building something closer to a data utility than a product application. The analogy I keep coming back to is Ribbon Health in our portfolio: they solve a data normalization problem — provider directory accuracy — that enables a downstream layer of applications. That infrastructure position tends to be sticky and high-value in ways that application-layer companies are not. I expect to see analogous infrastructure companies emerge at the multimodal synthesis layer over the next three to five years, and we are actively looking for them.
The clinical validation challenge and the infrastructure opportunity are connected. Founders building multimodal synthesis platforms who are also investing in the data normalization layer underneath them — who are treating the data quality problem as a core product problem rather than a business development problem — are the ones I find most credible. That's where the defensible value will be built.