Forty-three consecutive TSC patients with DRE were enrolled, and 235 cortical tubers were selected as the training set. Quantitative indices of cortical tubers on 18F-FDG PET were extracted, and logistic regression analysis was performed to select those with the most important predictive capacity. Machine learning models, including logistic regression (LR), linear discriminant analysis (LDA), and artificial neural network (ANN) models, were established based on the selected predictive indices to identify epileptogenic tubers from multiple cortical tubers. A discriminating nomogram was constructed and found to be clinically practical according to decision curve analysis (DCA) and clinical impact curve (CIC). Furthermore, testing sets were created based on new PET images of 32 tubers from 7 patients, and follow-up outcome data from the cortical tubers were collected 1, 3, and 5 years after the operation to verify the reliability of the predictive model. The predictive performance was determined by using receiver operating characteristic (ROC) analysis.
PET quantitative indices including SUVmean, SUVmax, volume, total lesion glycolysis (TLG), third quartile, upper adjacent and standard added metabolism activity (SAM) were associated with the epileptogenic tubers. The SUVmean, SUVmax, volume and TLG values were different between epileptogenic and non-epileptogenic tubers and were associated with the clinical characteristics of epileptogenic tubers. The LR model achieved the better performance in predicting epileptogenic tubers (AUC = 0.7706; 95% CI 0.70-0.83) than the LDA (AUC = 0.7506; 95% CI 0.68-0.82) and ANN models (AUC = 0.7425; 95% CI 0.67-0.82) and also demonstrated good calibration (Hosmer‒Lemeshow goodness-of-fit p value = 0.7). In addition, DCA and CIC confirmed the clinical utility of the nomogram constructed to predict epileptogenic tubers based on quantitative indices. Intriguingly, the LR model exhibited good performance in predicting epileptogenic tubers in the testing set (AUC = 0.8502; 95% CI 0.71-0.99) and the long-term outcomes of cortical tubers (1-year outcomes: AUC = 0.7805, 95% CI 0.71-0.85; 3-year outcomes: AUC = 0.8066, 95% CI 0.74-0.87; 5-year outcomes: AUC = 0.8172, 95% CI 0.75-0.87).
The 18F-FDG PET image-based LR model can be used to noninvasively identify epileptogenic tubers and predict the long-term outcomes of cortical tubers in TSC patients.
方法:纳入43例连续的TSC患者,选择235个皮质块茎作为训练集。提取了18F-FDGPET上皮质块茎的定量指标,并进行逻辑回归分析以选择具有最重要预测能力的那些。机器学习模型,包括逻辑回归(LR),线性判别分析(LDA),和人工神经网络(ANN)模型,根据选定的预测指标建立,以从多个皮质块茎中识别癫痫性块茎。根据决策曲线分析(DCA)和临床影响曲线(CIC),构建了判别列线图,并发现其在临床上具有实用性。此外,基于来自7名患者的32个块茎的新PET图像创建测试集,术后1年、3年和5年收集皮质块茎的随访结果数据,以验证预测模型的可靠性。通过使用接收器工作特性(ROC)分析来确定预测性能。
结果:PET定量指标,包括SUVmean,SUVmax,volume,总病变糖酵解(TLG),第三个四分位数,上邻近和标准增加的代谢活性(SAM)与致癫痫块茎有关。Suvmean,SUVmax,致癫痫和非致癫痫块茎的体积和TLG值不同,并且与致癫痫块茎的临床特征相关。与LDA(AUC=0.7506;95%CI0.68-0.82)和ANN模型(AUC=0.7425;95%CI0.67-0.82)相比,LR模型在预测癫痫性块茎方面取得了更好的性能(AUC=0.7706;95%CI0.70-0.83),并且还显示出良好的校准(Hosmer-Lemeshow拟合优度p值=0.7)。此外,DCA和CIC证实了根据定量指标构建的用于预测癫痫发生块茎的列线图的临床实用性。有趣的是,LR模型在预测测试集中的癫痫性块茎(AUC=0.8502;95%CI0.71-0.99)和皮质块茎的长期结局(1年结局:AUC=0.7805,95%CI0.71-0.85;3年结局:AUC=0.8066,95%CI0.74-0.87;5年结局:AUC=0.8172,95%CI)方面表现良好.
结论:基于18F-FDGPET图像的LR模型可用于非侵入性识别癫痫性块茎,并预测TSC患者皮质块茎的长期预后。