Fig. 4: Study arms: machine learning model, ordering clinician, and independent clinician.

When a Heme-STAMP order is placed by the participating hematologists, a silent EHR alert is triggered, initiating the three study arms to obtain estimates of pathogenicity. Arm A: Order information is used to generate an email (Supplementary Fig. 1) that includes the order date, patient medical record number, name, latest white blood cell count, and latest hemoglobin levels - values that remind the hematologists of relevant patient information. Orders placed over one to two week periods are batched into a consolidated query email to reduce email burden. This batching still protects the integrity of prospective predictions since the turnaround time for Heme-STAMP is around three weeks. These emails are sent to the respective ordering hematologists to obtain their estimates of a pathogenic result in buckets of <10%, 10–30%, 30–50%, 50–70%, 70–90%, and >90%. Arm B: Order information is saved in a secure database server. These cases are later reviewed by an independent hematologist who uses chart review up to the date of the order to provide their estimate. Arm C: The automated ML model integrated into the live EHR uses data directly from Epic Chronicles, the EHR’s transactional database that stores up-to-date patient data, to query structured patient information (e.g., demographics, diagnosis codes, prior lab results, and medications) needed to calculate the probability of a pathogenic test result. We developed a DEPLOYR¹⁴ framework that uses fast healthcare interoperability resources (FHIR)²⁸-based Epic application programming interfaces (API) calls to pull live feature data, with the appropriate mappings to training feature data, in order to enable this real-time prediction.