Driving after stroke: A trichotomous logistic regression model to support decision making in uncertain cases

Le résumé fourni par la source

BACKGROUND: Assessing fitness to drive after stroke is a complex clinical task, as even mild cognitive deficits can undermine safety. Due to the substantial overlap in cognitive test results between safe and unsafe drivers, binary classification models inevitably carry a risk of misclassification. This study aimed to develop and validate a logistic regression model that introduces a third, indeterminate category - leaving room for clinicians to withhold judgment in uncertain cases and thereby support more cautious, evidence-based decisions. METHODS: A total of 115 stroke survivors underwent a standardized neuropsychological evaluation, including assessments of attention, executive function and visuospatial planning. Novel dynamic response time measures were included. Driving fitness was evaluated through a standardized on-road test, which served as the primary outcome. Logistic regression modeling was combined with leave-one-out cross-validation and trichotomous classification to minimize overfitting and manage diagnostic uncertainty. RESULTS: Based on the on-road evaluation, regarded as the gold standard in the field, 70 % of participants were judged to be safe drivers. Our model demonstrated a ROC-AUC value of 0.95 after validation, while 15 % of the cases were classified as indeterminate. The Trail Making Test, Stroop test, Hungarian version of Road Law and Road Craft Knowledge test and the Starry Night Test all contributed to the model's accuracy. CONCLUSION: Our logistic regression model allows clinicians to refrain from making unfounded decisions in cases where cognitive test results are inconclusive. In a small proportion of uncertain cases, further assessment is recommended, ideally an on-road test. The model supports more targeted use of on-road evaluations by identifying cases where cognitive test results alone are insufficient.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.