The efficacy of machine learning algorithms in evaluating factors associated with shunt-dependent hydrocephalus after subarachnoid hemorrhage: a systematic review and meta-analysis.

The identification of factors associated with chronic shunt-dependent hydrocephalus (CSDH) following spontaneous subarachnoid hemorrhage (SAH) remains challenging, despite numerous studies. Early recognition of patients at higher risk for requiring shunt placement is important for optimizing management strategies. This systematic review and meta-analysis evaluated the efficacy of machine learning (ML) algorithms in analyzing datasets related to CSDH post-SAH, assessing performance metrics such as sensitivity, accuracy, and specificity. A systematic review was conducted across five databases (PubMed, Scopus, Cochrane Library, Embase, and Web of Science) to identify studies employing ML to analyze factors associated with CSDH following SAH. Data extraction included ML techniques, input features, and performance metrics such as area under the receiver operating characteristic curve (AUC-ROC), accuracy, sensitivity, specificity, precision, and F1 score. Two independent reviewers extracted and organized the data, including details on machine learning models, validation processes, and metrics. Out of 993 reviewed studies, five met the inclusion criteria for analyzing ML models in relation to CSDH post-SAH. The pooled AUC-ROC across these models was 0.79 (95% CI: 0.78-0.81), with moderate heterogeneity (I² = 42.58%, Q (19) = 34.79, p = 0.01). No significant differences in AUC-ROC were observed between linear, tree-based, and deep learning models (Q (2) = 0.99, p = 0.61). Studies utilizing fewer than 10 input features showed a lower pooled AUC-ROC of 0.78, whereas those with more than 10 features achieved a higher AUC-ROC of 0.82, with heterogeneity of 7.52% and 66.53%, respectively. Machine learning algorithms can assist in identifying factors associated with the development of chronic hydrocephalus following spontaneous SAH through analysis of clinical datasets. Incorporating a greater number of relevant risk factors may further improve the performance of these ML models in understanding high-risk patient profiles.