Individualized treatment estimates to inform on personalized cancer care decisions for treatment selection and treatment management.

Abstract

e13571 Background: Clinical trials report average effects of treatments at a population level. However, clinicians must balance treatment benefits against risks to personalize therapy. Here, we present a statistical framework which incorporates real-world data to estimate efficacy and symptom burden for an individual patient. We then assess the potential impact of continuing immunotherapy beyond progression in patients with metastatic solid malignancies on immune-related adverse events, progression, and survival. Methods: We corrected for sampling biases with inverse propensity score weighting (IPTW), and used the IPTW-transformed data to train predictive models of patient outcomes under each option. The models utilize a peer-based similarity based on 222 routinely-collected patient characteristics, such as cancer and treatment specific factors, laboratory tests, demographics, vitals, stage, comorbidities and symptoms, which are extracted from the electronic health record or natural language processing of clinician narratives, and Social Determinants of Health from Census data. We developed our methods on “Dataset A”: 5,919 breast and genitourinary (GU) cancer patients at a tertiary cancer center. We computed individualized treatment effect (ITE) estimates on “Dataset B”: 5,486 breast, GU, and thoracic cancer patients who received immunotherapy. Results: Our approach eliminated biases in Dataset A: the IPTW-transformation reduced the between-group standardized mean difference across identified confounders from an average of 0.225 to an average of 0.022. The similarity model exhibited strong discrimination (ROC-AUC > 0.75), and identified associations between patient characteristics and outcomes that matched clinicians’ expectations. Validations are ongoing to assess the ITE estimates on a held-out subset of Dataset B. These include subgroup analyses for metastatic triple-negative breast cancer and non-small cell lung cancer. Conclusions: We introduce a method to estimate the efficacy and toxicity of interventions, actions, or treatment decisions, which replicates the statistical robustness of a clinical trial while obtaining individualized insights. This framework provides causal estimates of symptom burden and risks of progression and survival to inform clinical decisions, including continuation of immunotherapy beyond progression. After IPTW, confounders in Dataset A were balanced between the treatment groups, with weighted standardized mean differences <0.1. [Table: see text]