Human Intuition vs. Computational Precision: Neurologists, Feature-based Models, and Deep Learning for Stroke Prognosis

The study compared the ability of six neurologists to predict outcomes in patients with large vessel occlusion stroke against the performance of computational models. Neurologists predicted outcomes based on clinical data and CT scans from 40 patients from the MR CLEAN trial, with their performance compared to two models: the validated MR PREDICTS model and a deep learning model. Results showed that standalone models significantly outperformed unaided neurologists in predicting the full range of outcomes (MR PREDICTS with kappa 0.51 versus neurologists with kappa 0.27). Neurologists exhibited systematic overoptimism and had low accuracy in estimating imaging features. For binary outcome prediction (good versus poor), accuracy was comparable between neurologists and models (64.17% versus 67.50% and 63.16%). Model assistance modestly improved neurologists' predictions. The study concluded that deep learning models have potential to serve as an automated second opinion to support decision-making in acute stroke treatment.