导航菜单
首页 >  paddlocrPaddle2ONNX模型转化与预测  > 融合影像、神经认知评价和生物标志等多模态数据预测阿尔兹海默症进展阶段及转化

融合影像、神经认知评价和生物标志等多模态数据预测阿尔兹海默症进展阶段及转化

摘要

目的:研究如何整合并优化影像、神经认知评价和生物标志测量等多来源多模态数据以提高阿尔兹海默症 (Alzheimer disease,AD)发展阶段和转化的分类预测准确率。方法:基于阿尔兹海默症影像计划(Alzheimer’s disease neuroim⁃ aging initiative,ADNI)2004—2018年4个阶段的样本数据,包括从核磁共振成像(magnetic resonance imaging,MRI)影像数据提取的脑图像特征数据、神经认知量表(简易精神状态测量量表和 ADAS⁃Cog13 量表)数据、生物标志测量数据(Abeta、Tau 和 p⁃Tau蛋白及ApoE4基因型)。基于783个样本的基线数据建立二分类和多分类Logistic回归模型用于疾病发展阶段的两两和同时分类预测。基于具有轻度认知障碍(mild cognitive impairment,MCI)状态的352个样本的纵向数据建立二分类Logistic回归模型并用于转化状态的分类预测。将脑图像特征变量、量表数据和生物标志加入到基准模型中,通过交叉验证方法随机划分数据集,并计算准确率、查准率、召回率、F1得分和ROC曲线下面积(area under curve,AUC)等指标进行综合比较,得到最优多模态组合的分类预测模型。结果:对于AD发展阶段的分类,结合脑图像特征数据、量表数据和生物标志数据建立二分类 Logistic 模型表现最佳,区分 AD 组和正常组、MCI 组和正常组以及 AD 组和 MCI 组的准确率分别达到了 100.00%、77.18%和 89.58%;AUC值分别为100.00%、85.52%和96.39%,比仅用脑图像数据进行进展阶段的分类预测有显著提高。对于MCI是否转化的分类预测,脑图像特征数据结合量表数据和生物标志能最大限度地提高准确率,从86.69%提高到90%以上;相应的 AUC值从89.21%提高到94.06%。结论:结合多来源数据能提高AD疾病进展阶段和转化的分类预测准确率,为临床诊断AD 所处的发展阶段和转化提供理论上的支持。

Abstract

Objective:To determine whether the combination of neurocognitive assessment and key biomarker data can improve the accuracy of using MRI image data to predict Alzheimer disease(AD)status and conversion. Methods:Data were collected from the Alzheimer’s Disease Neuroimaging Initiative(ADNI) during 2004—2018. Samples with complete MRI image data,cognitive assessment data and biological measures were screened from the raw data. Seven brain volumetric features including ventricular, hippocampus,whole brain,entorhinal cortex,fusiform gyrus,middle temporal gyrus,and intracerebral volumes were extracted from MRI by toolbox FreeSurfer. Cognitive assessment scale included the MMSE and ADAS ⁃ Cog13 scale. Biological measurement data included four biomarkers,i.e.,Abeta peptide,Tau protein,p ⁃ Tau protein and ApoE4 genotype. Based on the baseline data of 783 samples,logistic regression model was established for classification of disease development stages. Based on the longitudinal data of 352 samples with MCI status,a binary logistic regression was established for distinguishing converted patients from non ⁃ converted patients. We integrated cognitive data,and biomarkers with the brain image data,randomly divided the data set through a cross ⁃ validation method,and calculated accuracy,recall,precision,F1 score,and the area under the ROC curve. Results:For the classification of AD development stage,combining brain image data,cognitive data and biological measures achieved accuracy rates of 100.00%(AD vs. Normal),77.18%(MCI vs. Normal)and 89.58%(AD vs. MCI);the areas under the ROC curves are 100.00%(AD vs. Normal),85.52%(MCI vs. Normal)and 96.39%(AD vs. MCI)respectively;the AUCs for distinguishing Normal,MCI,and AD from the other two categories are 88.30% ,81.00% and 97.26% respectively,which are significantly higher than the classification performance using only brain image data. For classification of MCI conversion,the brain image data combined with the cognitive data can maximize the accuracy rate,from 86.69% to more than 90%;the corresponding AUC increased from 89.21%,which only use the brain image data to 94.06%. Conclusion:Combining data from multiple sources can improve the classification and prediction accuracy of AD status and conversion,thus provide theoretical support for clinical practice in early diagnosis of the AD.

关键词

阿尔兹海默症;MRI影像;神经认知量表;生物标志

Keywords

Alzheimer disease;MRI brain image;cognitive assessment scale;biomarker

阿尔兹海默症(Alzheimer disease,AD)是一种起病隐匿的进行性神经系统疾病,主要在老年群体中发病,俗称老年痴呆症。轻度认知障碍(mild cognitive impairment,MCI)是由正常衰老状态发展为AD的一种早期状态,其症状接近于正常衰老过程,通常被误以为是衰老的表现。研究表明约44%的MCI患者在3年内会转化为AD[1]。全球每年至少会投入1 000亿美元经费用于AD的研究、诊断和治疗,然而迄今为止尚无有效药物和治疗手段能够完全治愈AD。对处于MCI阶段的患者通过加强照顾并及时进行药物和精神治疗,能有效减缓病情的恶化[2-3],因此AD的早期诊断非常重要。尽早诊断MCI可以有效控制延缓疾病的发展,提高患者的生活质量并减轻社会家庭的负担。

AD的临床诊断主要基于认知量表评价、神经影像及重要生物标志物[4-6]。神经影像技术是目前诊断AD最直观可靠的手段。然而神经影像解读容易受个人主观影响,且脑室扩大和脑沟增宽也可出现在正常老年人中,并不是痴呆的唯一征象[7]。因此MRI、CT等检查无法检测出早期无特异性影像学改变的患者[8]。文献表明基于脑图像数据对AD的疾病状态或者对MCI转化进行分类仅能达到80%左右的预测准确率[9-11]。基于脑图像数据、神经认知评价及重要的生物标志等多模态数据有望提高AD早期诊断和MCI进展的准确率。Alam等[12] 提出将脑图像数据及简易精神状态检查量表相结合对AD的3个进展阶段进行两两分类,结果表明使用脑图像数据的分段体积特征对正常组和AD组进行二分类时可达到93.85%的分类准确率,使用皮层下分割后的体积特征区分MCI组与正常组和AD组时,检出率分别达到86.54%和75.12%。Gupta等[13] 提出了一种基于机器学习的分类框架,利用氟脱氧葡萄糖正电子发射断层扫描、结构磁共振成像、脑脊液蛋白水平和ApoE基因来区分AD或MCI,结果表明相比于单一模态分类方法,多模态数据联合分类方法能较大幅度地提高分类性能。

相关推荐: