From the Institute of Biomaterials and Biomedical Engineering, University of Toronto, Toronto, Ont. (Bhagwat, Chakravarty); the Cerebral Imaging Centre, Douglas Mental Health University Institute, Verdun, Que. (Bhagwat, Chakravarty); the Kimel Family Translational Imaging-Genetics Research Lab, Research Imaging Centre, Campbell Family Mental Health Research Institute, Centre for Addiction and Mental Health, Toronto, Ont. (Bhagwat, Pipitone, Voineskos); the Department of Psychiatry, University of Toronto, Toronto, Ont. (Voineskos); and the Department of Psychiatry, McGill University, Montreal, Que. (Chakravarty), Canada.
J Psychiatry Neurosci. 2019 Jul 1;44(4):246-260. doi: 10.1503/jpn.180016.
The development of diagnostic and prognostic tools for Alzheimer disease is complicated by substantial clinical heterogeneity in prodromal stages. Many neuroimaging studies have focused on case–control classification and predicting conversion from mild cognitive impairment to Alzheimer disease, but predicting scores from clinical assessments (such as the Alzheimer’s Disease Assessment Scale or the Mini Mental State Examination) using MRI data has received less attention. Predicting clinical scores can be crucial in providing a nuanced prognosis and inferring symptomatic severity.
We predicted clinical scores at the individual level using a novel anatomically partitioned artificial neural network (APANN) model. The model combined input from 2 structural MRI measures relevant to the neurodegenerative patterns observed in Alzheimer disease: hippocampal segmentations and cortical thickness. We evaluated the performance of the APANN model with 10 rounds of 10-fold cross-validation in 3 experiments, using cohorts from the Alzheimer’s Disease Neuroimaging Initiative (ADNI): ADNI1, ADNI2 and ADNI1 + 2.
Pearson correlation and root mean square error between the actual and predicted scores on the Alzheimer’s Disease Assessment Scale (ADNI1: r = 0.60; ADNI2: r = 0.68; ADNI1 + 2: r = 0.63) and Mini Mental State Examination (ADNI1: r = 0.52; ADNI2: r = 0.55; ADNI1 + 2: r = 0.55) showed that APANN can accurately infer clinical severity from MRI data.
To rigorously validate the model, we focused primarily on large cross-sectional baseline data sets with only proof-of-concept longitudinal results.
The APANN provides a highly robust and scalable framework for predicting clinical severity at the individual level using high-dimensional, multimodal neuroimaging data.
阿尔茨海默病的诊断和预后工具的发展受到前驱期临床异质性的影响。许多神经影像学研究集中在病例对照分类和预测从轻度认知障碍到阿尔茨海默病的转化上，但使用 MRI 数据预测来自临床评估（如阿尔茨海默病评估量表或简易精神状态检查）的分数受到的关注较少。预测临床评分对于提供细致的预后和推断症状严重程度至关重要。
我们使用一种新的解剖分区人工神经网络（APANN）模型在个体水平上预测临床评分。该模型结合了与阿尔茨海默病中观察到的神经退行性模式相关的两种结构 MRI 测量值的输入：海马分割和皮质厚度。我们在 3 项实验中使用来自阿尔茨海默病神经影像学倡议（ADNI）的队列进行了 10 轮 10 倍交叉验证，评估了 APANN 模型的性能：ADNI1、ADNI2 和 ADNI1 + 2。
在 ADNI1 上，阿尔茨海默病评估量表（ADNI1：r = 0.60；ADNI2：r = 0.68；ADNI1 + 2：r = 0.63）和简易精神状态检查（ADNI1：r = 0.52；ADNI2：r = 0.55；ADNI1 + 2：r = 0.55）的实际分数与预测分数之间的 Pearson 相关系数和均方根误差表明，APANN 可以从 MRI 数据中准确推断临床严重程度。
为了严格验证该模型，我们主要关注具有初步概念验证的纵向结果的大型横截面基线数据集。
APANN 为使用高维多模态神经影像学数据在个体水平上预测临床严重程度提供了一个高度稳健和可扩展的框架。
