Application Research Of A Machine Learning Model Integrating Quantitative Parameters Of Energy Spectrum CT In Predicting Symptomatic Carotid Atherosclerotic Plaque

Objective:To evaluate the predictive ability of machine learning（ML）models integrating quantitative parameters of energy spectrum CT,traditional plaque features and clinical risk factors to symptomatic carotid plaque.Methods:A retrospective analysis was performed on 171 patients who underwent head and neck CTA examination in the First Hospital of Shanxi Medical University from 2018 to 2021and found plaque from the beginning of the common carotid artery to the skull of the internal carotid artery.Two observers independently observed,measured and calculated the quantitative parameters of energy spectrum CT（fat fraction,iodine density,effective atomic order value）and traditional plaque characteristics（plaque thickness,CT value,marginal sign,plaque location,plaque length,stenosis degree,minimum lumen area,surface ulcer）.The consistency between the two observers was analyzed using intra group correlation coefficient（ICC）.In addition,The clinical data of patients were also collected.The subjects were divided into symptomatic group（n=104）and asymptomatic group（n=67）,and the inter group difference analysis were conducted.Based on quantitative parameters of energy spectrum CT,traditional plaque characteristics and clinical risk factors,a predictive model for symptomatic plaque was constructed using the XGboost algorithm.The importance of features was ranked using the Shapley additive explanations（SHAP）method.The XGboost integrated model was constructed using the top ten features.In addition,energy spectrum feature models,traditional plaque feature models,and clinical risk factor models were constructed respectively.Draw the ROC curve,evaluate the model differentiation according to the area under the curve（AUC）,compare the differences of AUC between different models using the Delong test,calculate the accuracy,accuracy,recall,and F1 score of the model to verify and compare the prediction effect of the model.Results:Compared with the asymptomatic group,patients in the symptomatic group had higher admission systolic blood pressure and homocysteine levels（150mm Hg vs140mm Hg,P=0.006;16.7mmol/L vs 13.3mmol/L,P<0.001）,and patients in the symptomatic group had larger FF（29.0%vs 22.2%）,NID（0.041 vs 0.022）,stenosis rate（53.9±14.9 vs 42.7±14.3）,maximum plaque thickness（3.7mm vs 3.1mm）,and smaller CT values（23HU vs 29HU）,Minimum lumen area（10.2mm²vs 17.7mm²）,The incidence of ulcers（25.0%vs 12.0%）and marginal signs（36.0%vs 18.0%）was higher in the symptomatic group,and the differences were statistically significant（P<0.05）.However,there was no significant difference in plaque length and Z_effbetween the two groups（20.5cm vs 21.6cm,P=0.481;7.37 vs 7.39,P=0.070）.The consistency of the data measured by the two observers is good,with ICC>0.75.The XGboost integrated model,which includes the top ten important features（AUC 0.946[95%CI 0.890-1.000]）,has significantly higher AUC values than the spectral feature model（AUC 0.778[95%CI0.638-0.918],P=0.016）,traditional plaque feature model（AUC 0.702[95%CI0.548-0.856],P=0.001）,and clinical risk factor model（AUC 0.754[95%CI 0.608-0.900],P=0.009）,and the XGboost integrated model has the highest accuracy rate,precision rate,Recall rate and F1 score.Conclusion:ML model integrating quantitative parameters of energy spectrum CT,traditional plaque features and clinical risk factors has better predictive ability of symptomatic carotid plaque.