AI Could Fill Critical ECOG Gaps in Prostate Cancer Research

In an interview with Pharmacy Times, Carole R. Berini, PhD, research scientist at Ontada, and Saamir Pasha, MPH, senior biostatistician at Ontada, discussed how artificial intelligence and routinely collected clinical data could help address missing Eastern Cooperative Oncology Group (ECOG) performance status in real-world oncology research. ECOG performance status is central to treatment selection and prognosis in advanced prostate cancer, but it is not consistently documented in electronic health records (EHRs), particularly near treatment initiation. Berini explained that excluding patients without documented ECOG scores can reduce sample sizes and introduce selection bias, producing evidence that may not represent the broader community oncology population.

The retrospective study included adults with advanced prostate cancer treated in The US Oncology Network who had a documented ECOG score near treatment initiation. Researchers evaluated structured demographic and clinical EHR data using 2 approaches: automated machine learning and multiple imputation. Older age, lower body mass index, lower hemoglobin, advanced disease stage, and renal disease emerged as clinically coherent predictors of poorer functional status. Pasha explained that the selected gradient boosting machine could identify nonlinear relationships and interactions among clinical variables, while the conventional statistical model provided an interpretable benchmark.

The machine learning model achieved an area under the receiver operating characteristic curve of 0.77 and an F1 score of 0.37 for poor ECOG status. The multiple imputation model achieved a similar training AUROC of 0.74. In a validation cohort of nearly 1200 patients, it demonstrated 89.4% accuracy and 99.1% specificity but only 5.1% sensitivity for poor ECOG status. The investigators emphasized that the model remains at the feasibility stage. External validation across health systems, tumor types, and patient populations will be necessary before proxy-derived ECOG status can confidently support comparative-effectiveness or regulatory research. They also stressed that predicted scores must preserve established clinical and outcome relationships.