OM1 Publishes Validation Results from a Single Healthcare Foundation Model Across More than 1700 Predicted Clinical Outcomes

Preprint reports that 80% of PhenOM^® predictive tasks reach an AUROC of at least 0.70 and that one-year discrimination for six common conditions matched or exceeded widely used published risk scores.

BOSTON--(BUSINESS WIRE)--OM1, Inc., a leader in healthcare AI and real-world data, today announced the posting of a preprint manuscript, Performance of a Foundation Model to Generate Thousands of Predicted Outcomes Across Multiple Use Cases, on the medRxiv preprint server. The paper is the first published performance study of PhenOM^®, OM1's generative AI foundation model trained and validated on longitudinal medical histories from more than 400 million de-identified patient records in the OM1 Real-World Data Cloud™. It describes how a single model, without additional training, has completed and validated thousands of predictive tasks across more than 12 clinical domains, and presents a representative sample of 177 of them.

Most published clinical prediction models address a single outcome in a single disease context. PhenOM^® instead uses one pretrained model to generate a likelihood score for any defined outcome, patient population and time window, with every prediction validated on a holdout set of roughly 200 million patients excluded from training. Key findings include:

• Scale: 1,770 validated predictive tasks were in the PhenOM^® catalog at the time of analysis, with new tasks validated each day. Across all of them, 80% achieved an AUROC of at least 0.70 and 30% at least 0.80.
• Breadth: Tasks fall into four categories — Future Risk, Screening, Care Opportunity and Treatment Effect — spanning cardiovascular, metabolic, renal, neurological, behavioral health, oncology, respiratory, musculoskeletal and surgical, women's health, gastrointestinal, rare and underdiagnosed conditions, and healthcare utilization and mortality, with horizons from one month to five years.
• Strong performers: One-year future risk of Alzheimer's disease (AUROC 0.92), heart failure (0.87) and all-cause mortality (0.87); screening for undiagnosed idiopathic pulmonary fibrosis (0.91) and pulmonary arterial hypertension (0.89); and post-operative delirium after hip or femur fracture surgery (0.85).
• Benchmarking: For six well-studied outcomes — Alzheimer's disease, dementia, heart failure, myocardial infarction, stroke and breast cancer — PhenOM^®'s one-year discrimination was comparable to or higher than published performance of established risk scores such as CAIDE, PCP-HF, the Pooled Cohort Equations, QRISK3, the Framingham Stroke Risk Profile and the Gail model, most of which predict over ten-year horizons. The comparisons are illustrative rather than head-to-head.

The authors acknowledge that all metrics derive from the OM1 Real-World Data Cloud and that external validation to date covers a fraction of available outcomes. Expanded external validation partnerships with health systems and payers are planned.

"Medicine has focused traditionally on population means, but now it's possible to use an individual’s probabilities for nearly every decision in healthcare," said Richard Gliklich, MD, OM1 founder and a co-author of the paper. "This paper shows that a single foundation model, trained on the longitudinal histories of hundreds of millions of patients, can answer thousands of clinically meaningful questions. A practice can find patients with undiagnosed, treatable conditions or start preventive treatment when the risk of future events like heart attack or stroke, or future conditions, like diabetes or dementia, are high. A value-based care organization can see which members are most likely to be hospitalized and intervene early. A surgeon can counsel a patient on their individual likelihood of benefit or complication. A drug developer can model the cohorts and events that matter to a trial. These capabilities exist in PhenOM today, and we believe this is what the shift toward predictive, proactive medicine will look like in practice."

To read the paper, click here.

About OM1

OM1 is a healthcare AI and real-world evidence company founded to close the gap between what healthcare demands and what traditional evidence can deliver. Through the PhenOM^® AI Foundation Model, clinical registries and built-for-purpose datasets, expert consulting services, and the OM1 Real-World Data Cloud built on more than 400 million patients and 2 billion patient-years, OM1 helps the healthcare industry predict what’s next to make more precise, data-driven decisions and ultimately, improve outcomes. Learn more at www.om1.com.