Presentation (CTI)

OBJECTIVES: Patient identification in real-world data (RWD) sources often relies on standardized diagnostic codes, which fail to capture patients who are undiagnosed, misdiagnosed, or living with conditions lacking specific diagnosis codes. This study evaluated the performance of PhenOM, a digital phenotyping platform, across three conditions representing distinct identification challenges: Metabolic dysfunction-associated steatohepatitis (MASH), obscured by asymptomatic disease and poor billing code specificity; Fabry disease (FD), a rare disorder with heterogeneous multi-organ presentation; and treatment-resistant depression (TRD), which lacks DSM classification or a dedicated diagnosis code.
METHODS: PhenOM is a digital phenotyping platform that predicts diagnoses, future medical events, conditions, and likely treatment response. It is trained on longitudinal, linked EMR and claims data from the OM1 Real-World Data Cloud, which includes over 370 million de-identified U.S. individuals with an average 10-year patient history. The platform identifies signals (diagnoses, labs, procedures, medications, and comorbidity patterns) to construct individual-level phenotypic profiles. Discrimination was quantified using the area under the receiver operating characteristic curve (AUROC). Clinical experts reviewed model outputs in each disease area to assess concordance with known disease biology.
RESULTS: In MASH, AUROC was 0.86 in out-of-sample testing. In FD, AUROC was 0.82 and was maintained in sex-stratified subgroups (male 0.83; female 0.82). FD prevalence in the highest-risk 1% of patients was 23.9 times greater than background population prevalence. In TRD, AUROC was 0.87, with consistency across subgroups by sex, age, and race. PhenOM identified 4,300 TRD patients not captured by either a regulatory treatment failure definition or a structured data-driven cohort definition.
CONCLUSIONS: PhenOM achieved AUROC of 0.86 (MASH), 0.82 (FD) and 0.87 (TRD), with subgroup performance maintained in each condition. These results support the use of AI-powered digital phenotyping for patient identification and real-world data studies where standard coding methods produce incomplete cohorts.