Identifying novel risk factors for aneurysmal subarachnoid haemorrhage using machine learning

Abstract

Aneurysmal subarachnoid haemorrhage (aSAH) is a type of stroke with high mortality and morbidity. This study aimed to identify novel aSAH risk factors by combining machine learning (ML) and traditional statistical methods. Using the UK Biobank, we identified aSAH cases via hospital-based ICD codes and analysed 618 baseline variables covering demographics, lifestyle, medical history, and physical measurements. The CatBoost ML algorithm and Shapley Additive Explanations (SHAP) identified the top 25 variables most influential in predicting aSAH. Logistic regression further described these variables while adjusting for established aSAH risk factors. Among 501,847 participants, 893 aSAH cases were identified. ML identified 214 variables with non-zero SHAP values. Logistic regression of the top 25 variables revealed four potential novel aSAH risk factors. Increased aSAH risk was associated with mean sphered cell volume (OR 1.02, 95% CI 1.00-1.03) and tea intake (OR 1.03, 95% CI 1.01-1.05). Decreased aSAH risk was associated with peak expiratory flow (OR 0.80, 95% CI 0.66-0.96), and haematocrit percentage (OR 0.97, 95% CI 0.95-1.00). Future research should validate these findings and explore the potential non-linear relationships and interactions indicated by the ML models.

Fig. 1

The 25 most important potential risk factors identified by the CatBoost algorithm. SHAP = SHapley Additive exPlanations, SHBG = Sex hormone binding globulin, IGF-1 = Insulin-like growth factor 1.

Fig. 2

Flowchart of study design. Numbers in parentheses indicate the number of rows (i.e. participants) and columns (i.e. variables) respectively. aSAH = aneurysmal subarachnoid haemorrhage, SHAP = SHapley Additive exPlanations.