Predicting delayed neurological sequelae in patients with carbon monoxide poisoning using machine learning models

Delayed neurological sequelae is a common complication following carbon monoxide poisoning, which significantly affects the quality of life of patients with the condition. We aimed to develop a machine learning-based prediction model to predict the frequency of delayed neurological sequelae in patients with carbon monoxide poisoning.

A single-center retrospective analysis was conducted in an emergency department from January 01, 2018, to December 31, 2023. We analyzed data from patients with carbon monoxide poisoning, which were divided into training and test sets. We developed and evaluated sixteen machine learning models, using accuracy, sensitivity, specificity, and other relevant metrics. Threshold adjustments were performed to determine the most accurate model for predicting patients with carbon monoxide poisoning at risk of delayed neurological sequelae.

A total of 360 patients with carbon monoxide poisoning were investigated in the present study, of whom 103 (28.6%) were diagnosed with delayed neurological sequelae, and two (0.6%) died. After threshold adjustment, the synthetic minority oversampling technique-random forest model demonstrated superior performance with an area under the receiver operating characteristic curve of 0.89 and an accuracy of 0.83. The sensitivity and specificity of the model were 0.9 and 0.8, respectively.

The study developed a machine learning-based synthetic minority oversampling technique-random forest model to predict delayed neurological sequelae in patients with carbon monoxide poisoning, achieving an area under the receiver operating characteristic curve of 0.89. This technique was used to handle class imbalance, and shapley additive explanations analysis helped explain the model predictions, highlighting important factors such as the Glasgow Coma Scale, hyperbaric oxygen therapy, kidney function, immune response, liver function, and blood clotting.

The machine learning-based synthetic minority oversampling technique-random forest model developed in this study effectively identifies patients with carbon monoxide poisoning at high risk for delayed neurological sequelae.