摘要
目的采用Transformer模型,融合结构磁共振成像(sMRI)数据与人口统计学资料,以实现对阿尔茨海默病(AD)病程阶段的分类识别。方法数据来源于阿尔茨海默病神经影像学倡议数据库(ADNI),随机选取543例研究对象,其中包括139例认知功能正常者(NC)、220例早期轻度认知障碍(EMCI)、108例晚期轻度认知障碍(LMCI)以及76例AD患者。采用基于动态阈值调整的L1正则化(L1-DTFS)及基于动态阈值调整的梯度提升决策树(GBDT-DTFS)算法,分别对这些研究对象的272项sMRI数据进行特征选择,筛选出最优特征子集。将筛选后的sMRI特征与3项人口统计学指标(年龄、性别、受教育程度)及简易精神状态检查(MMSE)评分共同输入Transformer模型和逻辑回归(LR)模型,观察其在区分AD连续病程中所有两两组合[共分为NC-EMCI(表示NC组与EMCI组分类,下同)、NC-LMCI、NC-AD、EMCI-LMCI、EMCI-AD以及LMCI-AD 6个分类组]时的分类效果,并通过受试者工作特征曲线下面积(AUC)评价模型的判别性能。结果L1-DTFS和GBDT-DTFS两种特征选择方法均筛选出了6组分类任务中最有贡献的优势特征,且L1-DTFS特征选择下的Transformer模型对NC与LMCI组的分类预测准确率、精确度、敏感度均达100%,AUC值为1.00。结论Transformer模型在AD病程分类中有较好且稳定的表现,其中在NC与LMCI病程分类组表现更佳。
Objective To leverage a Transformer model for integrating structural magnetic resonance imaging(sMRI)data with demo⁃graphic information to classify Alzheimer's disease(AD)progression stages.Methods The data were sourced from the Alzheimer's Disease Neuroimaging Initiative database(ADNI).A total of 543 subjects were randomly selected,including 139 with normal cognition(NC),220 with early mild cognitive impairment(EMCI),108 with late mild cognitive impairment(LMCI),and 76 with Alzheimer's disease(AD).The dynamic threshold adjustment-based L1 regularization(L1-DTFS)and gradient boosting decision tree(GBDTDTFS)algorithms were applied to the 272 sMRI features of these subjects to perform feature selection and identify the optimal feature subsets.The selected sMRI features,along with three demographic indicators(age,gender,and education level)and the Mini-Mental State Examination(MMSE)score,were input into the Transformer model and the logistic regression(LR)model.Their performance was assessed across all six pairwise classification tasks along the Alzheimer's disease continuum(NC vs EMCI,NC vs LMCI,NC vs AD,EMCI vs LMCI,EMCI vs AD,and LMCI vs AD),with the discriminative power quantified by the area under the receiver ope⁃rating characteristic curve(AUC).Results Both feature selection methods,L1-DTFS and GBDT-DTFS,successfully identified the most contributive dominant features across all six classification groups.Notably,the Transformer model incorporating L1-DTFS achieved perfect performance(100%accuracy,precision,and sensitivity)with an AUC of 1.00 in classifying the NC versus LMCI groups.Conclusion The Transformer model demonstrates robust and stable performance in classifying the stages of Alzheimer's disease progression,with particularly superior results in distinguishing between the NC and LMCI stages.
作者
施转芳
范炤
SHI Zhuanfang;FAN Zhao(Department of Physiology,School of Basic Medical Sciences,Shanxi Medical University,Taiyuan 030001,China;Translational Medicine Center of Shanxi Medical University)
出处
《山西医科大学学报》
2026年第2期215-222,共8页
Journal of Shanxi Medical University
基金
山西省留学回国人员科技活动择优资助项目(619017)
山西省重点研发计划国际合作项目(201803D421068)。
