Bridging Machine Learning and Clinical Endpoints: A METABRIC-Informed Simulation Study of Missing Data Imputation for...

Area of Science:

• Oncology
• Biostatistics
• Machine Learning

Background:

• Missing data, particularly progression-driven dropout, biases longitudinal oncology studies and RECIST criteria response classification.
• Current machine learning imputation methods lack evaluation within clinically interpretable frameworks focused on patient-level endpoints like Best Overall Response (BOR).

Purpose of the Study:

• To propose and evaluate a clinically grounded framework for assessing imputation methods in oncology studies.
• To compare the performance of imputation strategies for Best Overall Response (BOR) classification using RECIST 1.1 criteria.

Main Methods:

• Simulated longitudinal tumor trajectories for 270 patients using Gompertz and Stein-Fojo growth models.
• Introduced realistic missingness, including progression-driven dropout, across nine follow-up visits.
• Evaluated three machine learning imputation models (LSTM, MissForest, MI) using direct and non-responder strategies, assessing performance via BOR classification metrics (accuracy, Cohen's kappa).

Main Results:

• Imputation substantially improved Best Overall Response (BOR) classification accuracy and Cohen's kappa across both simulation models.
• Non-responder imputation strategies consistently outperformed direct imputation, improving accuracy by up to 10 percentage points and kappa by up to 17 percentage points.
• Findings demonstrated robustness across tumor growth models and missingness scenarios, highlighting the impact of imputation.

Conclusions:

• Effective handling of missing data in oncology requires careful consideration of both imputation methods and clinically meaningful endpoints.
• Clinically informed strategies for missing data, particularly progression-related missingness, significantly enhance RECIST-based Best Overall Response (BOR) classification.
• Endpoint selection and estimand strategy for missing data handling may be more influential than the specific imputation model chosen.