A comparative analysis in a clinical cohort: multiple imputation by chained equations and a novel super learner-based...

Missing data is a challenge in clinical research, especially in real-world data (RWD), where complete case analysis can bias results and reduce power. Ensemble learning approaches like Super Learner (SL) show strong numerical performance for prediction problems, but their use for missing value imputation (MVI) in oncology datasets is unexplored. We sought to develop and evaluate a novel SL-based imputation function that can impute multiple variables and quantify observation-specific uncertainty.

We analyzed two independent cohorts of acute myeloid leukemia patients (n = 1641), 546 patients from the University of Colorado and 1095 from an external real-world cohort. The SL-based MVI function includes data processing, predictor selection, binary and continuous variable pipelines, and automatic performance measurement. Ensembles for both binary and continuous variables integrate diverse base learners, such as generalized linear models, random forests, and neural networks, via a meta learner that optimizes predictive accuracy. The binary variable pipeline's SL ensemble was optimized using area under the curve (AUC), while the continuous variable pipeline's SL ensemble was optimized via non-negative least squares. Performance was compared to multiple imputation by chained equations (MICE) using balanced accuracy, F₁-score, root mean square error (RMSE), and visualizations. Observation-specific uncertainty was quantified for all imputations of both binary and continuous variables, with both additionally having lower and upper resampling-based potential imputation values. The SL cross-validation loop, SL ensemble trained for imputation, and resampling all supported parallelization. Clinically significant features of the cohorts were selected a priori based on prior literature.

In a numerical experiment with 9 clinically important binary features, the proposed MVI function imputed and achieved higher balanced accuracy than MICE for 7/9 variables (mean balanced accuracy 89.04% vs. 80.75%) with comparable performance for the other 2 variables. The continuous variable SL ensemble, across 4 variables, showed an average 24.45% lower RMSE than MICE. On average, the SL ensemble trained for prediction took 145.02 s to process for binary targets.

This study demonstrates that the SL-based imputation function has improved performance over MICE in high-dimensional RWD while providing novel, observation-level uncertainty quantification.