Quantify Heterogeneity in Treatment Effects; A Case Study for Patients with Acute Myeloid Leukemia (AML)

Introduction Randomized clinical trials (RCT) have been conducted to evaluate the impact of conditioning regimens (MAC vs RIC/NMA) for patients with AML . While the largest RCT (BMT CTN 0901) showed an overall benefit of patients randomized to MAC, one smaller RCT did not reach the same conclusion. Treatment effects can vary based on patient characteristics such as age, but older patients were excluded from the RCTs . Traditional regression and machine learning methods have been used to estimate patient-specific treatment effects and identify characteristics correlated with the heterogeneity in treatment effects. Objectives Using an observational AML cohort of 1075 patients (PMID: 36881031), we sought to quantify the heterogeneity using the standard deviation (SD) of patient-specific treatment effects in relapse-free survival (RFS) differences and observed that different methods yield qualitatively different conclusions. Our goal is to identify the most reliable method that could aid with treatment decision making . Methods SDs of RFS differences at times of interest post-transplant (3, 6, 12 months) were estimated by four models: Cox proportional hazard regression (Cox), Cox with LASSO shrinkage (CoxLasso), BART fitted to all patients (BART1), and BART fitted separately to two regimen groups (BART2). We simulated realizations from prior distributions of models to decompose the source of heterogeneity. To identify the optimal method for data like ours, we established a simulation framework to construct a neighborhood of the observed data and compared adjusted mean squared errors (MSEs) and coverage probability of the competing methods across the neighborhood. Results The estimated ATEs were similar across methods (RFS benefit = 0.06 at month 12). The SDs varied substantially, leading to different clinical implications of regimen choice ( Figure 1 ). Results from CoxLasso/BART1 suggested MAC would benefit most patients (77%-88%), while results from Cox/BART2 implied MAC would be harmful to 37%-40% of the patients. Simulations from prior distributions indicated that results were highly correlated with model assumptions ( Figure 2 ). Using the created data neighborhood, we found that CoxLasso had the best performance with the smallest adjusted MSEs and the highest coverage probability. It identified covariates interacting with treatment, including measurable residual disease , age, risk category, and donor group. Conclusion We established a simulation framework to translate method assumptions to RFS improvement. This framework helped identify covariates that contributed to treatment effect estimations in an AML cohort with the most reliable model. It will help investigators distinguish the influence of prior assumptions inherent in method and its parameters from the strength of the evidence in the data for estimating the heterogeneity in treatment effects.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.