Abstract
Background and aims
Patients with acute medium/distal vessel occlusions (MDVO) are heterogeneous and two trials failed to establish superiority of endovascular treatment (EVT) over best medical treatment (BMT). Because subgroups showing heterogeneity of the treatment effect (HTE) are unknown, we assessed HTE across strata of outcome predictions from an externally developed model.
Methods
A TabPFN model was trained on patients of a local stroke registry meeting DISTAL inclusion criteria. The model predicts excellent early neurological outcome (EENO, 24h-NIHSS 0-2) in patients treated with BMT alone using admission clinical and imaging variables. The model was then used to predict EENO in never-seen patients of the DISTAL trial (n=539). HTE of BMT alone vs BMT+EVT was evaluated across strata of predicted and continuous probabilities using interaction analyses with mRS shift as primary outcome.
Results
In the external development cohort (n=257), 54% of patients showed EENO. The model performed acceptable in the entire DISTAL cohort (AUC 0.72, calibration intercept: −0.13, slope: 0.52, 35% with EENO). There was HTE by predicted probabilities (p-interaction=0.018, Fig1): In patients predicted to have an EENO (n=136), EVT was harmful (cOR 0.50, 95%-CI 0.27-0.90, number needed to harm ≈9 for mRS 0-2, Fig2), while there was no significant effect of EVT in patients predicted not having EENO (n=403). Results of sensitivity analyses were comparable.
Conclusions
Our deep-learning model identified heterogeneous treatment responses, indicating that EVT may be harmful compared to BMT alone in patients predicted to have an EENO. The model will be made available during ESOC to complement clinical decision-making.
Conflict of interest
Nothing to disclose
Figure 1 - belongs to Results
Figure 2 - belongs to Conclusions
