Presentation (CTI)

OBJECTIVES: Patient performance status (PS) is a key determinant of cancer outcomes, but is often missing in Real World Data due to lack of data capture and structured entry in electronic health record and claims data. The Disability Score (DS), developed using Medicare claims and survey data, was intended to overcome this and support health status ascertainment at time of cancer diagnosis. We explore the external validity of this measure in real world cohorts of patients across multiple cancer types.
METHODS: This study leveraged Tempus AI records with available Komodo claims data. We included patients with metastatic breast, colorectal, and non-small cell lung cancer (NSCLC) and required a six month period of continuous medical and pharmacy insurance enrollment prior to their metastatic diagnosis (index) date. Patients were required to have a documented PS assessment +/-30 days of index date. The features from the DS measure were reconstructed in this dataset, and the original coefficients used to predict the patient’s DS. Predictive performance was analyzed across thresholds from 0.1-0.9.
RESULTS: There were 5,784 records, 1,260 breast cancer, 1,509 colorectal cancer, and 3,015 NSCLC. Most (96.4%) were classified as having good PS, comparable to the 90.7% reported for the cohort used to develop DS. Sensitivity (0.95-1), accuracy (0.93-0.96), and PPV (0.97-0.96) were consistently high across all thresholds, while specificity (0.10-0.01) was consistently low, indicating that this model was overclassifying records with good PS as having poor DS.
CONCLUSIONS: The DS measure was not externally valid across real-world cohorts of metastatic cancer patients. Future PS imputation models should leverage datasets that more closely align with the target population. Additionally, researchers should assess the validity of historical algorithms to their data prior to imputing missing data using these algorithms.