Some of the biggest breakthroughs in medicine don’t come from a single dataset or a single discipline, they come from the moment genomics, imaging, clinical records, and machine learning start collaborating. That’s the space Machine Learning for Health (ML4H) at the Broad Institute works in every day, turning multi-omic and clinical data into earlier diagnoses, sharper risk prediction, and new biological insight. None of this happens in isolation: it takes hospitals willing to share data and expertise, institutions willing to validate each other’s work, and teams willing to sit in the data together until the science holds up. Broad Clinical Labs (BCL), the Broad’s CLIA-certified, CAP-accredited subsidiary, generates and processes the large-scale multi-omic and sequencing data that fuels work like this — and ML4H’s progress is a good example of where that kind of collaboration can lead.
What is multi-omics, and why does scale matter?
Multi-omics refers to combining multiple layers of molecular data – genomics, transcriptomics, proteomics, and more, to build a fuller picture of human biology than any single layer could provide on its own. At BCL, that scale is generated every day: large volumes of multi-omic and sequencing data for research and clinical customers. But generating data at scale is only half the story. Turning it into insight — earlier diagnosis, better risk prediction, a clearer understanding of disease — takes machine learning expertise, and BCL works alongside a range of computational and clinical AI groups to help make that happen.
We spoke with members of the ML4H team about where machine learning is already adding value on top of multi-omic data, and where the field is headed next.
Where machine learning adds value on top of multi-omic data
ML4H’s mission is using AI to learn from longitudinal patient data and to connect what’s known about biology and genetics with what happens to patients over time. The group currently works across cardiovascular disease, brain health, women’s health, psychiatric disease, genetics and precision medicine, and wearable data, with a common thread of earlier prediction, better characterization of disease heterogeneity, and understanding how disease unfolds over time.
The team points to one of the biggest untapped opportunities in large-scale genomic and multi-omic data: linking molecular measurements to longitudinal clinical phenotype. The field has become very good at generating genomic and multi-omic data at enormous scale but knowing that someone carries a particular variant or molecular signature doesn’t tell you whether disease will manifest, when, or in what form. Connecting molecular data to years of clinical history (IE: notes, imaging, electrocardiograms (ECG), labs, medications, outcomes) opens the door to studying penetrance, disease progression, and phenotypic heterogeneity, rather than simple variant-to-disease association.
That shift is a big part of what draws the group to consortium work like National Institute for Health (NIH) MAGen: the scientific question moves from “Is this variant associated with disease?” to “Which carriers actually develop disease, how does it manifest, and can we recognize that trajectory earlier?” A related, concrete application is phenotyping for genome-wide association studies (GWAS) — using clinical data to adjudicate diagnoses and outcomes at scale and folding genetic data into multimodal risk models alongside other data types.
Multimodal by design: combining genomics, proteomics, EHR, and more
ML4H’s projects consistently show that combining data types surfaces things no single modality reveals alone.
One example: cardiac MRI and ECG. The MRI captures the heart’s structure and function — chamber size, wall thickness, vessel abnormalities, how much blood the heart pumps per beat — while the 12-lead ECG captures the heart’s electrical activity, itself a rich signal for current and future disease. The team has been developing methods to combine the two for better risk prediction, to predict one modality from the other when one is missing, and to build representations that separate what’s shared between the modalities from what’s unique to each.
Complementary work in the NIH’s All of Us Research Program combines wearable data — continuous activity, sleep, and heart rate — with EHR data, survey responses, and polygenic risk scores to test whether multimodal inputs improve prediction of disease trajectories beyond any single source.
From research result to clinical insight: biomarkers, diagnosis, risk prediction
Getting a model from a promising research result to something that can inform clinical practice is a long path.
Models are typically first validated within the population they were trained on, then externally in an independent population — a step that depends heavily on data availability and inter-institutional agreements. The fuller sequence looks like: build the model, establish rigorous internal validation, test temporal and subgroup robustness, validate in a genuinely independent population or health system, and determine whether performance holds in the intended clinical workflow — sometimes via a “silent” prospective study that runs alongside care without influencing it, before asking whether the model changes decisions or outcomes.
The ML algorithm itself is rarely the bottleneck. External validation, for example: getting a model to hold up outside the Boston-area hospitals where most ML4H data originates, is where projects most often stall. That’s part of what makes consortia like MAGen, which spans MGB, Boston Children’s, Cedars-Sinai, Vanderbilt, Mount Sinai, and Columbia, valuable: member institutions have agreed to externally validate each other’s models, a rare and difficult arrangement to establish. Further downstream, actual implementation within a healthcare system is its own hurdle, demanding operational, technical, and clinical resources that go well beyond the modeling work itself.
Why data quality, scale, and provenance decide the results
Data quality matters throughout, but there’s a useful distinction between training and evaluation: training can often tolerate some noise, as long as it isn’t systematically biased in the model’s inputs or outcomes, but validation data deserves particular scrutiny, since that’s what determines how confidently the team can characterize a model’s real-world performance.
What is an example of a multimodal AI study?
The wearable-data work within All of Us is a useful illustration of multimodal analysis in practice: combining continuous physiological signals (activity, heart rate, sleep) with EHR data, survey responses, and polygenic risk scores to ask whether disease trajectories can be predicted earlier and more accurately than any single data source allows.
Looking ahead: from single predictions to patient trajectories
The field is moving away from single-purpose models — one model for heart failure, another for dementia — toward models that integrate years of clinical history with genetics, imaging, physiology, and eventually molecular data to characterize a patient’s overall trajectory. One vision of the near future has trusted models running after each patient encounter to update risk assessments or flag missed diagnoses, raising open questions around clinical interpretation, fairness, unnecessary follow-up testing, and accountability. Put another way, the central question is shifting from “can we predict this diagnosis?” to “what trajectory is this patient on, what biology may be driving it, and where is there an opportunity to intervene?”
Underlying this shift is a shared concern that ML models developed primarily on academic-hospital data can leave behind patients without the same access to care — part of why MAGen’s focus on equity, alongside accuracy, resonates with the team.
How to prepare for a collaboration with Broad’s ML and genomics capabilities
For a health system, biotech, or research group considering a collaboration, the starting point is consistent: a strong, clearly defined scientific or clinical question, a defined target population, a sense of the outcome or phenotype of interest, access to relevant data, and critically a domain expert willing to stay engaged through cohort definition, labeling, interpretation, and validation.
It also helps to know what success looks like up front: biological discovery, a clinically meaningful finding, patient stratification, or a model built toward eventual prospective testing. Common failure modes include underestimating the time and staffing needed to label data, data access constraints, and questions that sound answerable but require information that’s too difficult to extract from the clinical record at the available budget.
This post draws on interviews with members of Machine Learning for Health (ML4H) at the Broad Institute. To learn more about ML4H’s work, visit the group’s website, or reach out directly at ml4h@broadinstitute.org with questions. For those who want to keep learning from ML4H’s experts, join the group’s monthly clinical AI seminar series mailing list and attend the Zoom talks at ml4h-talks.
©2026 Broad Clinical Labs | broadclinicallabs.org