Machine learning models for outcome prediction of patients with ischaemic stroke undergoing reperfusion therapy: a...

Reperfusion therapy, including thrombolysis and thrombectomy, is crucial for ischaemic stroke treatment. However, patient outcomes often remain suboptimal. Conventional regression models show limited accuracy in predicting outcomes after reperfusion therapy. Machine learning predictive models offer potential by integrating multidimensional data. However, their relative advantages over conventional regression models in this context remain uncertain.

We aim to compare the performance of conventional regression models, machine learning models in predicting the prognosis of patients undergoing reperfusion therapy.

We identified studies using regression or machine learning models to predict outcomes in patients with ischaemic stroke undergoing thrombolysis or thrombectomy. Model performance was summarised as the area under the receiver operating characteristic curve (AUC), with 95% CIs for prediction of modified Rankin Scale (mRS), symptomatic intracranial haemorrhage (sICH) and mortality. Heterogeneity was assessed using Cochran's Q test. Pooled AUCs were calculated. Risk of bias was assessed using Prediction model Risk Of Bias Assessment Tool (PROBAST) and reporting quality was assessed using Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD).

In total, 53 studies were included, of which 37 reported AUCs with 95% CI on validation datasets. Pooled analyses were conducted for mRS (n=37), sICH (n=14) and mortality (n=2). Specifically, 24 studies used conventional regression models (pooled AUC 0.80 (95% CI 0.77 to 0.82)), while 13 used machine learning models (0.86 (0.81 to 0.90)). Pooled machine learning model performance showed significant improvement over conventional regression models (p=0.004). Significant differences in model performance were also observed in thrombolysis subgroup (pooled AUC 0.79 (95% CI 0.77 to 0.82) for conventional regression models vs 0.88 (0.80 to 0.96) for machine learning models, p for interaction=0.009).

Machine learning models generally outperformed conventional regression models in predicting outcomes after reperfusion therapy, highlighting their potential for prognostic prediction of patients with ischaemic stroke undergoing reperfusion therapy. However, the high risk of bias across studies and limited availability of external validation warrant cautious interpretation of the predictive performance.