An optimized machine learning methodology leveraging computational fluid dynamics-derived hemodynamic features for enhancing prediction accuracy of thin-walled regions in intracranial aneurysms

초록

Thin-walled regions (TIWRs) in intracranial aneurysms (IAs) are closely associated with high-risk spontaneous rupture and can increase intraoperative rupture likelihood due to incomplete wall visualization. While computational fluid dynamics (CFD)–derived hemodynamic features (HFs) are valuable biomedical signals for predicting TIWRs, reliance on any single HF yielded unsatisfactory accuracy, and multiple HF combinations were difficult to further enhance accuracy due to the subjectively determined weighting coefficients. Machine learning (ML) algorithms can coordinate multiple HFs to facilitate prediction, yet a hasty application without hyperparameter tuning cannot maximize accuracy. Here, we develop an optimized ML methodology leveraging CFD-derived HFs to enhance the prediction accuracy of TIWRs in IAs. Given the structures of IA mesh-surface datasets, a nested leave-one-patient-out cross-validation framework was followed to conduct feature selection and improved grey wolf optimizer (IGWO) hyperparameter tuning for ten candidate ML algorithms, thereby determining the ML algorithm with the superior performance and enabling external generalization testing. The results showed that hyperparameter tuning by IGWO—which integrated chaotic mapping, nonlinear convergence factor, and adaptive position update—universally boosted the prediction accuracy. Among ten ML algorithms, Light Gradient Boosting Machine (LightGBM) exhibited the superior predictive performance: on the external testing case, it achieved an accuracy of 91.9 %, a precision of 92.9 %, a recall of 93.9 %, and areas under the relevant curves all exceeding 0.94. Therefore, IGWO-tuned LightGBM leveraging CFD-derived HFs has the potential to enhance preoperative TIWRs prediction accuracy in IAs, thereby reinforcing pretherapy guidance, reducing intraoperative rupture, and providing a methodological foundation for large-scale clinical translation. © 2025 Elsevier Ltd

키워드