İnternal karotid arter darlığını tahmin etmede makine öğrenme algoritmalarının kullanımı ve öngörüm başarısının dubleks Doppler ultrasonografi kriterleriyle karşılaştırılması

Purpose: There is a discrepancy between duplex Doppler ultrasonography (DUS) and digital subtraction angiography (DSA) for determining internal carotid artery (ICA) stenosis. We aim to train machine learning algorithms (MLAs) with DUS velocity values for predicting ICA stenosis and comparing their success to DUS criteria. Materials and methods: DUS values (peak systolic velocity (PSV) and end-diastolic velocity of the common carotid artery (CCA) and ICA) and DSA studies of 159 ICA stenoses were reviewed retrospectively. Stenoses were classified as <50%, 50-69%, ≥70% by each modality. Linear regression models with descriptive and predictive analysis and MLAs; LightGBM, XgBoost, KNeighbors, Support Vector Machine (SVM), Decision Tree, Random Forest were trained with DUS values for predicting DSA stenosis. Results: Predicted values of regression models have a linear relationship with DSA stenosis between 0-60%. LightGBM and SVM achieved the highest classification accuracy (69%), while all algorithms failed in the 50-69% interval. DUS criteria outperformed all MLAs in predicting DSA stenosis of ≥70% (sensitivity:0.91). Both MLAs and DUS criteria were unsuccessful in the 50-69% interval where DUS mostly overestimates and MLAs underestimate. MLAs using ICA PSV/CCA PSV ratio had higher accuracy for predicting DSA stenosis <50%. Conclusion: DUS criteria could be considered as the sole diagnostic tool for ICA stenosis over 70%. Improved DUS criteria or wider training datasets for MLAs are warranted to detect 50-69% stenosis accurately.