Choi Dong Hyeok, Ahn So Hyun, Chung Yujin, Kim Jin Sung, Jeong Jee Hyang, Yoon Hai-Jeon
Department of Medicine, Yonsei University College of Medicine, Seoul, Republic of Korea.
Department of Radiation Oncology, Yonsei Cancer Center, Yonsei University College of Medicine, Seoul, Republic of Korea.
Sci Rep. 2025 Jul 1;15(1):21987. doi: 10.1038/s41598-025-00743-7.
This study developed machine learning models to predict Aβ positivity in Alzheimer's disease by integrating early-phase F-Florbetaben PET and clinical data to improve diagnostic accuracy. Furthermore, the study explored machine learning models to predict cognitive status from early-phase PET, maximizing the clinical utility of dual-phase imaging protocols. 176 subjects who completed dual-phase F-FBB PET scanning including 38 with normal cognition, 94 with mild cognitive impairment, and 44 with dementia were enrolled. Aβ status was determined from delayed-phase F-FBB PET scans (90-110 min post-injection). To develop a machine learning model for predicting Aβ positivity, we utilized early-phase PET and clinical features. From early-phase F-FBB PET scans (0-10 min post-injection), we extracted brain region-specific standardized uptake value ratios (SUVR) as imaging features. Various classifiers, including Random Forest, Gradient Boosting, and XGBoost, were trained and evaluated using accuracy, ROC AUC, recall, and F1 scores. Feature importance was assessed to identify key predictors, and the importance of features that most significantly influenced each model's results was calculated. The early-phase PET alone showed moderate performance (80.56% accuracy with Random Forest), with hippocampus (importance: 0.086), isthmus of cingulate (0.051), and entorhinal (0.038) SUVR values as top predictors. The combined PET and clinical data model achieved the highest accuracy (88.89%) using Gradient Boosting, with key predictors including APOE genotype (importance: 0.2485), Medial Orbitofrontal SUVR (0.0996), and hippocampal SUVR (0.0663). In predicting cognitive status using early-phase PET, most classifiers achieved high accuracy (> 80%) and F1 scores (0.82-0.90), with Decision Tree showing the highest accuracy of 83.33%. Machine learning models combining PET and clinical data demonstrated superior predictive accuracy for Aβ positivity prediction, while early-phase PET alone showed robust performance in predicting cognitive status, highlighting the synergistic potential of multimodal data and versatile utility of early-phase PET imaging.
本研究通过整合早期F-氟代贝他宾PET和临床数据，开发了机器学习模型来预测阿尔茨海默病中的Aβ阳性，以提高诊断准确性。此外，该研究还探索了机器学习模型，以从早期PET预测认知状态，从而最大限度地提高双相成像方案的临床效用。招募了176名完成双相F-FBB PET扫描的受试者，其中包括38名认知正常者、94名轻度认知障碍者和44名痴呆患者。Aβ状态由延迟期F-FBB PET扫描（注射后90-110分钟）确定。为了开发预测Aβ阳性的机器学习模型，我们利用了早期PET和临床特征。从早期F-FBB PET扫描（注射后0-10分钟）中，我们提取了脑区特异性标准化摄取值比率（SUVR）作为成像特征。使用随机森林、梯度提升和XGBoost等各种分类器，并通过准确率、ROC曲线下面积、召回率和F1分数进行训练和评估。评估特征重要性以识别关键预测因子，并计算对每个模型结果影响最显著的特征的重要性。仅早期PET表现出中等性能（随机森林准确率为80.56%），海马体（重要性：0.086）、扣带回峡部（0.051）和内嗅区（0.038）的SUVR值为主要预测因子。使用梯度提升的PET和临床数据组合模型实现了最高准确率（88.89%），关键预测因子包括APOE基因型（重要性：0.2485）、内侧眶额皮质SUVR（0.0996）和海马体SUVR（0.0663）。在使用早期PET预测认知状态时，大多数分类器实现了高准确率（>80%）和F1分数（0.82-0.90），决策树显示出最高准确率为83.33%。结合PET和临床数据的机器学习模型在预测Aβ阳性方面表现出卓越的预测准确性，而仅早期PET在预测认知状态方面表现出强大性能，突出了多模态数据的协同潜力和早期PET成像的多功能效用。
