Performance of a foundation model to generate thousands of predicted outcomes across multiple use cases
Year: 2026
Type: preprint
Abstract: Background Predictive artificial intelligence (AI) models hold great promise for medical care. While many types of models have proliferated across healthcare and the life sciences, systematic, rapid generation of multi-condition predictions of diagnoses, future medical events and future clinical outcomes from a single foundation-model architecture remains rare. PhenOM® is a generative AI foundation model trained and validated on longitudinal medical-event data from more than 400 million patient records in the OM1 Real World Data Cloud (RWDC) and designed to complete predictive tasks for a broad array of medical diagnoses, events and clinical outcomes (collectively ‘outcomes’) at varying time intervals, currently up to five (5) years. This paper describes the implementation of PhenOM to complete a vast number of predictive tasks across clinical domains and outcomes. Methods PhenOM was evaluated across four outcome categories (Future Risk, Screening, Care Opportunity, and Treatment Effect) in multiple domains spanning chronic disease, acute medical events, healthcare utilization, and rare and frequently undiagnosed conditions. Performance was assessed using area under the receiver operating characteristic curve (AUROC). Outcomes were defined within an eligible-population framework using standardized medical code sets (ICD-10, SNOMED, CPT, NDC, LOINC) unless outcomes could not be specified with an existing code. All benchmarks reported were derived from data excluded from training. Results As of this publication, the single PhenOM model has been used to complete and validate thousands of predictive tasks covering more than 1,770 outcomes spanning most organ systems and clinical specialties, with new predictive tasks being validated each day. Overall, 80% of predictive tasks achieved AUROC of at least 0.70 and 30% achieved AUROC of at least 0.80. We present a sample of 177 of these tasks as examples that show the breadth of tasks the model can complete. Performance varied systematically by outcome and clinical domain. Conclusions PhenOM demonstrates the ability to complete thousands of predictive tasks across clinical domains and outcomes at the individual patient level, without additional training. Outcomes include screening for undiagnosed conditions, predicting future conditions, medical events, effects of treatments, and assessing care opportunities. These can be ascertained at the individual patient and population levels. These findings support applications in individual and population health management, risk management, drug development, and real-world evidence generation, and establish a transparent, reproducible performance reference for a large-scale real-world predictive AI platform.
Source: medRxiv
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7224203803