Deep learning-based model for difficult transfemoral access prediction compared with human assessment in stroke
Pere Canals1,2, Alvaro Garcia-Tornel3, Manuel Requena3,4
1Stroke Unit, Neurology, Vall d'Hebron University Hospital, Barcelona, Spain pere.canals@vhir.org.
Journal of Neurointerventional Surgery
|May 3, 2024
Summary
A new machine learning model accurately predicts difficult transfemoral access for mechanical thrombectomy (MT) in stroke patients. This AI tool improves upon expert assessments, aiding procedural planning and enhancing stroke treatment outcomes.
Area of Science:
- Neurology
- Medical Imaging
- Artificial Intelligence
Background:
- Mechanical thrombectomy (MT) success depends on cervical vessel access.
- Vascular tortuosity is a key factor influencing MT procedure duration and outcomes.
- Current methods lack speed and reliability in identifying anatomical challenges for cervical access.
Purpose of the Study:
- To develop and validate a machine learning model for predicting difficult transfemoral access (DTFA) in MT.
- To assess the model's performance against expert human assessment.
- To provide a tool for optimizing arterial access decisions in MT.
Main Methods:
- A retrospective analysis of 513 patients undergoing transfemoral MT for large vessel occlusion stroke.
- Development of a machine learning model using 29 anatomical features from head-and-neck CTA to predict DTFA.
- Comparison of the model's predictive accuracy against three expert raters on a subset of 116 cases.
Main Results:
- 11.5% of MT procedures experienced DTFA.
- The machine learning model achieved an AUROC of 0.76 for DTFA prediction.
- The model demonstrated superior performance (F1-score 0.70) compared to expert assessments (F1-scores 0.43-0.50).
Conclusions:
- A fully automated machine learning model for predicting DTFA in MT has been successfully developed and validated.
- This AI-driven approach significantly enhances the prediction of difficult arterial access compared to traditional expert evaluation.
- The model's predictions can inform strategic decisions regarding patient selection and procedural planning for MT.
