Identifying delayed neurological sequelae during the acute phase of carbon monoxide poisoning based on diffusion-weighted imaging and clinical features

Delayed neurological sequelae (DNS) represents a critical and potentially fatal complication. Therefore, the timely recognition of individuals at risk of developing DNS in early phase holds significant clinical value. This study aims to identify diffusion-weighted imaging (DWI) characteristics related to developing DNS, and construct a predictive model encompassing DWI characteristics and clinical variables to early prediction of DNS during the acute phase of carbon monoxide (CO) poisoning. We retrospectively include 120 poisoned patients with newly diagnosed with CO poisoning. The subjects were divided into non-DNS group ( n = 75) and DNS group ( n = 45) after at least 60-day follow-up. The fractional anisotropy (FA) and mean diffusivity (MD) value were measured in the regions of interest placed on the amygdala, caudate, hippocampus, pallidus, putamen, thalamus and ventricle. A support vector machine (SVM) model integrated both DWI and clinical features was developed and evaluated. And mean impact value was used to rank the features that had impacts on classification. A total of 24 clinical features and 28 DWI features were included. 8 clinical features and 23 DWIs were included in the SVM model. Three SVM models were established based solely on clinical features or DWI features, and combined clinical and DWI features, with prediction accuracy of 0.76, 0.94, and 0.97, respectively. The precision, sensitivity, F1 score, macroscopic mean, weighted mean, and AUC of the combined model is 1.00, 0.92, 0.96, 0.98, 0.97, 0.97, respectively, and the result of 10-flod cross validation was 0.98. The mean fractional anisotropy of left caudate had the highest impact on the SVM model. Our comprehensive SVM model integrated clinical features and FA and MD characteristics on DWI had excellent accuracy and clinical practicability in identifying DNS.