Optimizing predictive performance without sacrificing explainability: Comparing logistic regression and ensemble decision trees for abdominal aortic aneurysm repair outcomes

<h2>ABSTRACT</h2><h3>Background</h3> An abdominal aortic aneurysm (AAA) is an irreversible dilation of the terminal aorta, posing significant rupture risks. Current guidelines for surgical intervention rely on aneurysm diameter thresholds (5.0 cm for women, 5.5 cm for men). However, ruptures frequently occur below these thresholds, highlighting limitations in diameter-based approaches. Machine learning (ML) techniques that integrate clinical, biomechanical, and morphological data may enhance predictive accuracy, but can suffer from limited interpretability due to their inherent complexity. <h3>Objective</h3> This study aims to create an interpretable surrogate model for understanding a clinician's decision to perform AAA repair surgery. This was done by transforming complex machine learning (ML) outputs into a logistic regression framework using Shapley feature contributions, thereby increasing interpretability in clinical decision-making. <h3>Methods</h3> In a previous study, CT images were used to perform biomechanical and morphological analyses of AAA. The resulting dataset included various clinical variables (demographics, comorbidities, and pharmaceutical use). A total of 512 patients with either stable or repair outcomes were included in this study, with the training set (n = 410) and testing set (n = 102) split using an 80:20 ratio. Custom Python scripts were developed to train and evaluate logistic regression and ensemble boosted tree models on the dataset to classify repair outcomes from stable ones. Odds ratios were computed for the logistic regression, and Shapley Additive exPlanations values were calculated to interpret feature importance in the ensemble boosted tree model. <h3>Results</h3> The logistic regression model demonstrated an average cross-validation accuracy of 91% ± 3%, specificity of 77% ± 11%, sensitivity of 61% ± 11%, and an area under the receiver operating characteristics curve (AUROC) of 79% ± 5%. In comparison, the XGBoost model achieved an average accuracy of 95% ± 2%, specificity of 93% ± 7%, sensitivity of 74% ± 11%, and an AUROC of 86% ± 5%. <h3>Conclusions</h3> The utility of providing a probability for an event to occur can be potentially beneficial to clinicians based on the surrogate logistic function, as they are familiar with the risk score derived from odds ratios. The feature weights can be provided to promote a better understanding of which variables contribute to decision-making. Future work to understand clinicians' attitudes towards decision-support tools may lead to greater adoption of well-validated artificial intelligence tools.