Predictors of In-Hospital Mortality After Thrombectomy in Anterior Circulation Large Vessel Occlusion: A Retrospective, Machine Learning Study

Background: Despite the increased use of mechanical thrombectomy (MT) in recent years, there remains a lack of research on in-hospital mortality rates following the procedure, the primary factors influencing these rates, and the potential for predicting them. This study aimed to utilize interpretable machine learning (ML) to help clarify these uncertainties.

Methods: This retrospective study involved patients with anterior circulation large vessel occlusion (LVO)-related ischemic stroke who underwent MT. The patient division was made into two groups: (I) the in-hospital death group, referred to as miserable outcome, and (II) the in-hospital survival group, or favorable outcome. Python 3.10.9 was utilized to develop the machine learning models, which consisted of two types based on input features: (I) the Pre-MT model, incorporating baseline features, and (II) the Post-MT model, which included both baseline and MT-related features. After a feature selection process, the models were trained, internally evaluated, and tested, after which interpretation frameworks were employed to clarify the decision-making processes.

Results: This study included 602 patients with a median age of 76 years (interquartile range (IQR) 65-83), out of which 54% ( = 328) were female, and 22% ( = 133) had miserable outcomes. Selected baseline features were age, baseline National Institutes of Health Stroke Scale (NIHSS) value, neutrophil-to-lymphocyte ratio (NLR), international normalized ratio (INR), the type of the affected vessel ('Vessel type'), peripheral arterial disease (PAD), baseline glycemia, and premorbid modified Rankin scale (pre-mRS). The highest odds ratio of 4.504 was observed with the presence of peripheral arterial disease (95% confidence interval (CI), 2.120-9.569). The Pre-MT model achieved an area under the curve (AUC) value of around 79% utilizing these features, and the interpretable framework discovered the baseline NIHSS value as the most influential factor. In the second data set, selected features were the same, excluding pre-mRS and including puncture-to-procedure-end time (PET) and onset-to-puncture time (OPT). The AUC value of the Post-MT model was around 84% with age being the highest-ranked feature.

Conclusions: This study demonstrates the moderate to strong effectiveness of interpretable machine learning models in predicting in-hospital mortality following mechanical thrombectomy for ischemic stroke, with AUCs of 0.792 for the Pre-MT model and 0.837 for the Post-MT model. Key predictors included patient age, baseline NIHSS, NLR, INR, occluded vessel type, PAD, baseline glycemia, pre-mRS, PET, and OPT. These findings provide valuable insights into risk factors and could improve post-procedural patient management.