计算机科学
生物标志物
相似性(几何)
人工智能
成像生物标志物
磁共振成像
正电子发射断层摄影术
模式识别(心理学)
相似性度量
支持向量机
机器学习
生物标志物发现
多任务学习
神经影像学
任务(项目管理)
医学
心理学
神经科学
放射科
生物
基因
图像(数学)
生物化学
经济
管理
蛋白质组学
作者
Yu Zhang,Po Yang,Vitaveska Lanfranchi
出处
期刊:International Conference on Industrial Informatics
日期:2021-07-21
被引量:8
标识
DOI:10.1109/indin45523.2021.9557584
摘要
Alzheimer's disease (AD) is a typical progressive neurodegenerative disease with insidious onset. Utilising various biomarkers to track and predict AD progression for supporting clinic decisions has recently received wide attentions. Accurate prediction of disease progression will help clinicians and patients make the best decisions on disease prevention and treatment. Typical prediction models focus on extracting biomarker morphological information of different regions of interest (ROIs) from magnetic resonance imaging (MRI) or positron emission tomography (PET), such as the average regional cortical thickness and regional volume. They are effective in modeling AD progression and understanding AD biomarkers, but cannot make full utilise of the internal temporal and spatial relationships between these biomarkers to improve the accuracy and stability of AD prediction. In this paper, we propose a new multi-task learning (MTL) method based on the tensor composed of the spatio-temporal similarity measure between brain biomarkers, using MRI data and cognitive scores of AD patients in different stages can effectively predict the progression of AD. Specifically, we define a temporal and spatial feature similarity measure to calculate the rate of change and velocity of each biomarker in MRI to form a vector, which represents the morphological changing trend of the biomarker, then we calculate the similarity of the changing trend between two biomarkers and encode the data to the third-order tensor, and extract interpretable biomarker latent factors from the original data. The prediction of each patient sample in the tensor is a task and all prediction tasks share a set of latent factors obtained from tensor decomposition to train the AD progression prediction model, which learns task correlation from the spatiotemporal tensor itself. We conducted extensive experiments utilising the data from the Alzheimer's Disease Neuroimaging Initiative (ADNI). Experimental results show that compared with ROI-based traditional single feature regression methods, our proposed method has better accuracy and stability in disease progression prediction in terms of root mean square error exhibiting an average of 4.10 decrease compared to Ridge regression, 0.19 decrease compared to Lasso regression and 0.18 decrease compared to Temporal Group Lasso (TGL) in the Mini Mental State Examination (MMSE) questionnaire.
科研通智能强力驱动
Strongly Powered by AbleSci AI