FIG 3.

Feature engineering pipeline. Clinician notes and other unstructured EHR data were transformed into structured tables by curators. Data tables constructed directly from structured EHR variables included demographic, laboratory (eg, albumin levels) and clinical measurement (eg, BMI), condition, drug, procedure, visit, and observation data. Condition codes and drug codes were aligned into common coding systems. Laboratory measurements underwent cleansing procedures, and units were harmonized. Intermediate tables were constructed containing data for each patient at various time points. Features were created by aggregating data using different time windows, example features displayed on the right. The ICD-10 codes used in the features are defined in the Data Supplement (Table S1). AJCC, American Joint Committee on Cancer; BMI, body mass index; dx, diagnosis; EHR, electronic health record; ICD, International Classification of Diseases; ICI, immune checkpoint inhibitor; OH, one-hot encoding; OMOP, Observational Medical Outcomes Partnership; PHEWAS, phenome-wide association studies; rel freq, relative frequency; SNOMED, Systemized Nomenclature of Medicine; TWA, time-weighted average.