Prédire la transformation hémorragique symptomatique à la suite d'un infarctus cérébral : introduction à une approche clinico-radiologique en machine learning et basée sur l'IRM

Ischemic stroke is a leading cause of death and disability in adult. Reperfusion therapies using intravenous thrombolysis (IVT) and endovascular interventions such as mechanical thrombectomy (MT) improve functional outcomes in patients with acute ischemic stroke (AIS) but unfortunately, these therapeutics increase the risk of intracranial haemorrhage (ICH) of ischemic brain tissue. Several definitions and classifications exist to define this complication, on a clinical or radiological level. However, high evidence indicates that symptomatic ICH (sICH) is the most relevant definition as it is most correlated with poor outcome. Several methods, based on clinical and/or radiological data, failed to predict the risk of haemorrhagic transformation in clinical practice. We attempted a new approach by training a supervised machine learning (ML) algorithm on clinical and MRI data within a 100 subjects cohort of patients with anterior circulation AIS treated by IVT and/or MT who underwent sICH (n=28), non-symptomatic ICH (n=27) and 45 controls with no bleeding, matched on clinical severity and age. We compared ML algorithm accuracy to the performance of the clinical Totaled Health Risks in Vascular Events (THRIVE) score and to the radiological Alberta Stroke Program Early CT Score applied to MR imaging (DWI-ASPECTS). ML algorithm predicted sICH with an Area Under receiver operating characteristic Curve (AUC) of 0.658 (CI 95% [0.534 – 0.783]). Applied in the cohort, estimated AUC of THRIVE score and DWI-ASPECTS were 0.664 (CI 95% [0.548 – 0.781]) and 0.634 (IC 95% [0.508 – 0.761]), respectively. Although it do not outperform current tools, this work showed that this algorithm was able to synthesize all clinical and radiological data provided and integrating the variety of information provided by MR imaging to provide a probability of sICH. Further studies are needed to improve these performances. Enlarging dataset is needed to improve learning phase, reduce overfitting risk and allow a 3D analysis to avoid data loss. More relevant clinical and radiological data could also be integrated to improve performances. Validation of these results on a more heterogeneous external population is also required.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.