Joint detection and clinical score prediction in Parkinson's disease via multi-modal sparse learning
Introduction
Parkinson's disease (PD) (a.k.a, hypokinetic syndrome rigid or primary Parkinsonism) is caused by the deterioration of nervous system (Martínez-Murcia, Górriz, Ramírez, Illán, & Ortiz, 2014), and its main symptoms are derived from the death of dopamine-generating cells in an area of the brain known as the substantia nigra. However, existing studies also found no evidence of dopaminergic deficit in some PD patients, specifically, scans without evidence of dopamine deficit (SWEDDs) (Aerts et al., 2012, Nicastro et al., 2016). This discovery greatly increases the challenges of PD diagnosis. Currently, PD diagnosis is mainly based on the presence of primary symptoms (e.g., shaking, stiffness, slowness of movement) and the subject's response to medication (Prashanth, Roy, Mandal, & Ghosh, 2014). Although it is known that these symptoms are derived from the degeneration of a type of nerve cell called dopamine (Mahfuz et al., 2015), there is still no definitive cure for PD because the cause for death of this cell is still unknown (Feinsod, Kreinin, Chistyakov, & Klein, 1998). Dorsey et al. (2007) predicted that there will be about nine million individuals over 50 with PD in 15 nations by 2030. Hence, it is imperative to take effective measures for diagnosis and treatment of PD.
Machine learning techniques play increasingly important role for PD prognosis and diagnosis and have become a hot topic recently (Long et al., 2012, Prashanth et al., 2014, Salamanca et al., 2015, Singh and Samavedham, 2015; Wang, Long, & Chen, 2013). However, most of existing research mainly focuses on the single-function model (Lee and Lim, 2012, Rana et al., 2015). For example, Yuvaraja et al. proposed a new functional connectivity indexed feature via bispectral analysis to perform the classification of PD patients without dementia (Yuvaraj et al., 2016). Bhalchandra et al. employed shape-based features pre-processed from single photon emission computed tomography (SPECT) to conduct discriminant analysis (Bhalchandra, Prashanth, Roy, & Noronha, 2015). In addition, most existing methods only adopt single modality features for PD clinical score prediction (Hou et al., 2016, Prashanth et al., 2014), and it is argued that different brain's functional or structural imaging modalities preserve different information. For example, diffusion tensor imaging (DTI) reveals microscopic details about tissue architecture either in a normal or diseased state (Bendlin et al., 2012), whereas magnetic resonance imaging (MRI) provides structure information about the brain tissue type.
In recent years, multi-modality has become a hot topic and many studies have shown that the combination of different modalities can produce better performance in disease diagnosis (Liu, Wee, Chen, & Shen, 2014; Zhu, Suk, & Shen, 2014a; Zhu, Suk, Lee, & Shen, 2016a). A method introduced in Lei, Chen, Ni, and Wang (2016) improves classification performance of neurodegenerative brain disease by capturing both individual and shared information via multimodal features and canonical correlation analysis (CCA). Therefore by using multi-modal data, we expect to boost the PD diagnosis and prediction performance of our technique.
Machine learning methods can be used to predict neurological disease progression and assess the stage of pathology (Cheng, Zhang, Chen, Kaufer, & Shen, 2013). Hence, some researchers have applied machine learning methods to estimate clinical scores from brain images (Cheng et al., 2013, Hou et al., 2016, Lei et al., 2015) and found reliable correlation between estimated clinical scores and different PD stages. Appropriate and targeted treatment can then be carried out to treat PD effectively. For example, Hou et al. predicted clinical scores of unified Parkinson's disease rating scale (UPDRS) in PD patients using resting-state functional magnetic resonance imaging (fMRI) (Hou et al., 2016).
Unlike many previous studies concentrated on only one task (i.e., classification or prediction), Zhang and Shen proposed a novel method to jointly address the regression and classification (JRC) problem (Zhu, Suk, & Shen, 2014b) for Alzheimer's disease (AD), which tackled both problems simultaneously in a unified framework. However, there is no prior method and framework to jointly conduct PD disease classification and clinical score prediction. For PD, the clinical scores mainly comprise of the Montreal cognitive assessment (MoCA) (Kandiah et al., 2014), the UPDRS (Hou et al., 2016), the Epworth sleepiness scale (ESS) (Johns, 1991), the mini-mental state examination (MMSE) (Tombaugh & McIntyre, 1992), the geriatric depression scale (GCS) (Yesavage, 1988), and the university of Pennsylvania smell identification test (UPSIT) (Doty, Shaman, Kimmelman, & Dann, 1984). In our study, we adopted the values of ESS as sleep scores and the degrees of UPSIT as olfactory scores.
In general, there are three key steps for classification or regression, i.e., feature extraction, feature selection, and classification or regression. In the field of medical image analysis, it is challenging to select informative features to solve the problem of small sample size and high feature dimension. For example, there are only 42 subjects (i.e., 28 PD patients and 14 NC) in Bowman's study (Bowman, Drake, & Huddleston, 2016), and 46 samples (i.e., 19 PD patients and 14 NC) in Long et al. (2012), but the feature dimension of both studies is quite large. It is hard to improve the signal-to-noise ratio (SNR) (Lu et al., 2015) and construct a universal robust model (Zhu, Huang, Tao Shen, Cheng, & Xu, 2012). Due to over-fitting problem caused by the high-dimensional data, feature selection has become a well-researched topic due to its effectiveness to address this problem. Once highly correlated features have been selected, classification, and regression can be straightforwardly implemented by support vector machine (SVM) and multivariate logistic regression (MLR) (Alexopoulos, 2010). In the neuroimaging field, dimensionality reduction is often carried out by subspace learning (Guerrero et al., 2014, Lei et al., 2015; Liu, Tosun, Weiner, & Schuff, 2013) or feature selection (Chu et al., 2012, Young et al., 2012). The idea of subspace learning method is to transform original feature to a low-dimensional space using Fisher's linear discriminant analysis (LDA) (Perronnin, Sánchez, & Mensink, 2010) and locally linear embedding (LLE) (Roweis & Saul, 2000). Meanwhile, the main goal of feature selection such as t-test method (Wang, Zhang, Liu, Lv, & Wang, 2014) and sparse linear regression (Mateos, Bazerque, & Giannakis, 2010) is to obtain the most discriminative and highly noise-resistant subsets from its original feature sets. Existing study shows that it is more effective to use feature selection methods than subspace learning, especially for neurodegenerative diseases since the selected features may be directly related to its essential features (Zhu, Suk, Lee, & Shen, 2016b). In addition, the performance of feature selection can be further improved by manifold learning techniques according to the recent neuroimaging studies (Zhu et al., 2013, Zhu et al., 2013).
To harness multi-modal data motivated by study in Bowman et al. (2016), Salamanca et al. (2015) and Zhu, Suk, and Shen (2014a), we propose a new united novel feature selection framework via multi-modal neuroimaging data for PD diagnosis, which simultaneously performs classification and clinical sores prediction based on a novel loss function. Specifically, we combine three features, gray matter (GM) of MRI images, cerebrospinal fluid (CSF) of MRI images, and fractional anisotropy (FA) coefficient of DTI images to discriminate normal controls (NCs) and PD/SWEDD. In addition, we jointly predict two clinical scores (i.e., sleep and olfactory score). Our proposed method has been evaluated using the public available Parkinson's progression markers initiative (PPMI) database (http://www.ppmi-info.org).
Section snippets
Materials and dataset
The experimental data of this study was acquired from the PPMI database (http://www.ppmi-info.org). PPMI (Marek et al., 2011) is the first all-around, broad-scale, multi-focus, observational, and international study to identify PD progression biomarkers [2]. Since PPMI is a longitudinal study, the state of participants varies depending on the study time. All data used in this work were acquired by a Siemens MAGNETOM Trio 3.0T MRI scanner. The high-resolution 3-dimensional (3D) T1-weighted
Experimental settings
In our experiments, we perform two binary classifications via SVM: NC vs. PD, NC vs. SWEDD. For each set of experiment, the feature selection model is trained using four different feature sets, namely, GM of MRI (T1G for short), CSF of MRI (T1C for short), FA coefficient of DTI (DTI for short), and T1G+T1C+DTI (GCD for short). For each feature set, two regression models are obtained, where the first model is to predict olfactory scores, and the second model is to predict sleep scores, and the
Limitations and future direction
In spite of the appealing performance achieved by the proposed method, there are several limitations of our method. First, the template used to extract features in our study is AAL atlas, which only has 116 ROIs segmented from brain areas. In fact, the extended AAL map (AAL-290) used in Bowman et al. (2016) has excellent classification performance and contains 290 regions. AAL-290 may provide more detailed and robust features by combining the theory of pivot selection (Mao, Zhang, Li, Liu, &
Conclusions
In this study, a united feature selection framework based on a novel objective function was proposed to simultaneously perform classification and clinical sores prediction in PD based on multi-modal neuroimaging data. From our experimental results, we observed that fusing different modalities enhances classification performance. We also observed that the proposed feature selection method based on a novel objective function outperformed its counterpart for single modality and multiple
Disclosure statement
The authors have no financial and personal conflicts of interest.
Acknowledgments
This work was supported partly by China 863 program (No. 2015AA015305), National Natural Science Foundation of China (Nos. U1301252, 61402296, and 61471243), Guangdong Key Laboratory Project (No. 2012A061400024), Guangdong Medical Grant (No. B2016094), Natural Science Foundation of Shenzhen (No. JCYJ20140418095735561, No. JCYJ20150731160834611), Shenzhen Key Basic Research Project (No. JCYJ20150525092940986), the Open Fund Project of Fujian Provincial Key Laboratory of Information Processing
References (87)
- et al.
Spatio-temporal dynamics in fMRI recordings revealed with complex independent component analysis
Neurocomputing
(2006) A fast diffeomorphic image registration algorithm
Neuroimage
(2007)- et al.
Microstructural and physiological features of tissues elucidated by quantitative-diffusion-tensor MRI
Journal of Magnetic Resonance
(2011) - et al.
An efficient hybrid kernel extreme learning machine approach for early diagnosis of Parkinson׳s disease
Neurocomputing
(2016) - et al.
Does feature selection improve classification accuracy? Impact of sample size and feature selection on classification using anatomical magnetic resonance images
Neuroimage
(2012) - et al.
Choosing ℓ(p) norms in high-dimensional spaces based on hub analysis
Neurocomputing
(2015) - et al.
Multiclass classification of FDG PET scans for the distinction between Parkinson's disease and atypical parkinsonian syndromes
NeuroImage. Clinical
(2013) - et al.
A detailed anatomical analysis of neurotransmitter receptors in the putamen and caudate in Parkinson's disease and Alzheimer's disease
Neuroscience letters
(1994) - et al.
Prediction of individual clinical scores in patients with Parkinson's disease using resting-state functional magnetic resonance imaging
Journal of the Neurological Sciences
(2016) - et al.
Longitudinal clinical score prediction in Alzheimer's disease with soft-split sparse regression based random forest
Neurobiology of Aging
(2016)
Montreal cognitive assessment for the screening and prediction of cognitive decline in early Parkinson's disease
Parkinsonism and Related Disorders
Parkinson's disease classification using gait characteristics and wavelet-based feature extraction
Expert Systems with Applications
Inter-modality relationship constrained multi-modality multi-task feature selection for Alzheimer's Disease and mild cognitive impairment identification
Neuroimage
Locally linear embedding (LLE) for MRI based Alzheimer's disease classification
Neuroimage
A real time displacement estimation algorithm for ultrasound elastography
Computers in Industry
The Parkinson Progression Marker Initiative (PPMI)
Progress in Neurobiology
Automatic detection of Parkinsonism using significance measures and component analysis in DaTSCAN imaging
Neurocomputing
Effect of mild cognitive impairment on the patterns of neural activity in early Parkinson's disease
Neurobiology of Aging
Parkinson's disease and local atrophy in subcortical nuclei: Insight from shape analysis
Neurobiology of Aging
Scan without evidence of dopaminergic deficit: A 10-year retrospective study
Parkinsonism & Related disorders
Diffusion weighted imaging-based maximum density path analysis and classification of Alzheimer's disease
Neurobiology of Aging
Brain connectivity and novel network measures for Alzheimer's disease classification
Neurobiology of Aging
Automatic classification and prediction models for early Parkinson's disease diagnosis from SPECT imaging
Expert Systems with Applications
Regions-of-interest based automated diagnosis of Parkinson's disease using T1-weighted MRI
Expert Systems with Applications
Skull stripping using graph cuts
NeuroImage
Unsupervised learning based feature extraction for differential diagnosis of neurodegenerative diseases: A case study on early-stage diagnosis of Parkinson disease
Journal of Neuroscience Methods
Automated anatomical labeling of activations in SPM using a macroscopic anatomical parcellation of the MNI MRI single-subject brain
NeuroImage
t-Test feature selection approach based on term frequency for text categorization
Pattern Recognition Letters
Early diagnosis of Parkinson's disease patients using RVM-based classification with multi-characteristics
Brain functional connectivity patterns for emotional state classification in Parkinson's disease patients without dementia
Behavioural Brain Research
Regional homogeneity approach to fMRI data analysis
Neuroimage
Multi-modal multi-task learning for joint prediction of multiple regression and classification variables in Alzheimer's disease
Neuroimage
Self-taught dimensionality reduction on the high-dimensional small-sized data
Pattern Recognition
Dimensionality reduction by mixed kernel canonical correlation analysis
Pattern Recognition
A novel matrix-similarity based loss function for joint regression and classification in AD diagnosis
Neuroimage
Improving the diagnostic accuracy in parkinsonism: A three-pronged approach
Practical Neurology
Introduction to multivariate regression analysis
Hippokratia
Laplacian eigenmaps for dimensionality reduction and data representation
Neural Computation
CSF T-Tau/Abeta42 predicts white matter microstructure in healthy adults at risk for Alzheimer's disease
PLoS One
Early detection of Parkinson's disease through shape based features from 123I-Ioflupane SPECT imaging
Multimodal imaging signatures of Parkinson's disease
Frontiers in Neuroscience
Altered brain activation and connectivity in early Parkinson disease tactile perception
American Journal of Neuroradiology
LIBSVM: A library for support vector machines
ACM Transactions on Intelligent Systems and Technology
Cited by (41)
Gene-related Parkinson's disease diagnosis via feature-based multi-branch octave convolution network
2022, Computers in Biology and MedicineCitation Excerpt :As we know, the effective use of neuroimaging data could be the key solution for PD diagnosis [19–25], and magnetic resonance imaging (MRI) is widely employed for the differential diagnosis of PD [26–28]. At present, many researchers extracted features of the region of interest (ROI) from neuroimaging data for PD diagnosis [29,30]. However, data imbalance and feature redundancy in neuroimaging data often degrade the performance of classifiers.
Multi-center sparse learning and decision fusion for automatic COVID-19 diagnosis
2022, Applied Soft ComputingClassification, Prediction, and Monitoring of Parkinson's disease using Computer Assisted Technologies: A Comparative Analysis
2020, Engineering Applications of Artificial IntelligenceCitation Excerpt :Processing brain images include template matching, find ROI’s de-noising, filtering surface reconstruction rendering, and others. Some of the tools to process brain images include MRICon software (Peng et al., 2017), BrainLab software (Peng et al., 2017), Insight toolkit (Oliveira et al., 2018), CImg library (Oliveira et al., 2018), Visualization toolkit (Oliveira et al., 2018), VBM (Voxel-Based Morphometry) toolkit (Lei et al., 2018), AAL atlas (Lei et al., 2018), FSL toolbox (Lei et al., 2018; Abós et al., 2017), brainnetome atlas (Abós et al., 2017), SPM, FLIRT, brainnet viewer (Lei et al., 2017), brain analyzer software (Betrouni et al., 2019). Feature extraction is also very important to extract discriminative features for robust classification.
Adaptive sparse learning using multi-template for neurodegenerative disease diagnosis
2020, Medical Image AnalysisCitation Excerpt :To improve it, the relationships between modalities are considered in a feature selection model via multimodal multi-task learning (Zhang and Shen, 2012). Also, more data relationships are explored in a joint sparse learning model (Lei et al., 2017b). The multiple subspace learning method can take more relationship information into consideration when dealing with informative neuroimaging data by exploring relationships from many perspectives.
Parkinson's Disease Progression Prediction Using Longitudinal Imaging Data and Grey Wolf Optimizer-Based Feature Selection
2024, International Journal of Intelligent Systems and Applications in Engineering