Machine Learning-Based Prediction of Delayed Neurologic Sequelae in Carbon Monoxide Poisoning Using Automatically...

Area of Science:

• Neurology
• Radiology
• Artificial Intelligence

Background:

• Delayed neurological sequelae (DNS) are severe complications of carbon monoxide (CO) poisoning.
• Currently, no reliable tools exist for assessing DNS risk.
• Early identification of at-risk patients is crucial for timely intervention.

Purpose of the Study:

• To evaluate the efficacy of machine learning (ML) models in predicting DNS risk.
• To assess the predictive value of imaging features automatically extracted from brain MRI.
• To identify patients with acute CO poisoning who are at risk for developing DNS.

Main Methods:

• Retrospective analysis of 373 acute CO poisoning patients.
• Extraction of 1618 radiomics and 4 lesion-segmentation features from brain MRI (DWI b1000 and ADC images), alongside 62 clinical variables.
• Training and evaluation of four ML models (linear regression, support vector machine, random forest, extreme gradient boosting) and an ensemble model using a five-fold cross-validation approach.

Main Results:

• The best performing ML model achieved an area under the receiver operating characteristic curve (AUROC) of 0.88 [0.86-0.9], with 82% accuracy, 81% sensitivity, and 82% specificity.
• The presence, size, and number of acute brain lesions on MRI were more predictive of DNS risk than advanced radiomics features.
• Machine learning models effectively distinguished patients with and without DNS.

Conclusions:

• ML models utilizing automatically extracted brain MRI features and clinical data can effectively predict DNS risk in CO poisoning patients.
• These models facilitate early risk stratification and inform treatment planning for DNS prevention.
• The developed models offer a promising tool for improving patient outcomes after acute CO poisoning.