Machine learning framework for early MRI-based Alzheimer's conversion prediction in MCI subjects
Introduction
Alzheimer's disease (AD), a common form of dementia, occurs most frequently in aged population. More than 30 million people worldwide suffer from AD and, due to the increasing life expectancy, this number is expected to triple by 2050 (Barnes and Yaffe, 2011). Because of the dramatic increase in the prevalence of AD, the identification of effective biomarkers for the early diagnosis and treatment of AD in individuals at high risk to develop the disease is crucial. Mild cognitive impairment (MCI) is a transitional stage between age-related cognitive decline and AD, and the earliest clinically detectable stage of progression towards dementia or AD (Markesbery, 2010). According to previous studies (Petersen et al., 2009), a significant proportion of MCI patients, approximately 10% to 15% from referral sources such as memory clinics and AD centers, will develop into AD annually. AD is characterized by the formation of intracellular neurofibrillary tangles and extracellular β-amyloid plaques as well as extensive synaptic loss and neuronal death (atrophy) within the brain (Mosconi et al., 2007). The progression of the neuropathology in AD can be observed many years before clinical symptoms of the disease become apparent (Braak and Braak, 1996, Delacourte et al., 1999, Morris et al., 1996, Serrano-Pozo et al., 2011, Mosconi et al., 2007). AD pathology has been therefore hypothesized to be detectable using neuroimaging techniques (Markesbery, 2010). Among different neuroimaging modalities, MRI has attracted a significant interest in AD related studies because of its completely non-invasive nature, high availability, high spatial resolution and good contrast between different soft tissues. Over the past few years, numerous MRI biomarkers have been proposed in classifying AD patients in different disease stages (Fan et al., 2008, Duchesne et al., 2008, Chupin et al., 2009, Querbes et al., 2009, Wolz et al., 2011, Hinrichs et al., 2011, Westman et al., 2011a, Westman et al., 2011b, Westman et al., 2012, Cho et al., 2012, Coupé et al., 2012; Gray et al., 2013; Eskildsen et al., 2013, Guerrero et al., 2014, Wang et al., 2014). Despite of many efforts, identifying efficient AD-specific biomarkers for the early diagnosis and prediction of disease progression is still challenging and requires more research.
In the current study, we present a novel MRI-based technique for the early detection of AD conversion in MCI patients by using advanced machine learning algorithms and combining MRI data with standard neuropsychological test results. In more detail, we aim to predict whether an MCI patient will convert to AD over a 3 year period (this is referred as progressive MCI or pMCI) or not (this is referred as stable MCI or sMCI) using only data at the baseline. The data used in this work is obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (www.loni.usc.edu/ADNI) and it includes MRI scans and neuropsychological test results from normal controls (NC), AD, and MCI subjects with a matched age range. Recently, several computational neuroimaging studies have focused on predicting the conversion to AD in MCI patients by utilizing various types of ADNI data such as MRI (e.g. Ye et al., 2011, Filipovych and Davatzikos, 2011, Batmanghelich et al., 2011), positron emission tomography (PET) (Zhang and Shen, 2011, Zhang and Shen, 2012, Cheng et al., 2012, Shaffer et al., 2013), cerebrospinal fluid (CSF) biomarkers (Zhang and Shen, 2011, Cheng et al., 2012, Davatzikos et al., 2011, Shaffer et al., 2013), and demographic and cognitive information (see Table 1, Table 7). Our method is a multi-step procedure combining several ideas into a coherent framework for AD conversion prediction:
- 1.
Semi-supervised learning, using data from AD and NC subjects to help the sMCI/pMCI classification
- 2.
Novel random forest based data integration scheme
- 3.
Removal of age related confound.
In the experimental sections we will demonstrate that all these provide a significant contribution towards the accuracy of the combined prediction model. Our method differs in the following aspects from earlier studies.
Most of the earlier studies were based on supervised learning methods, where only labeled data samples are used for learning the model. Semi-supervised learning (SSL) approaches are able to use unlabeled data in conjunction with labeled data in a learning procedure for improving the classification performance. The great interest in SSL techniques over the past few years (Zhu, and Goldberg, 2009) is related to the wide spread of application domains where providing labeled data is hard and expensive compared to providing unlabeled data. The problem studied in this work, predicting the AD-conversion in MCI subjects, is a good example of this scenario since MCI subjects have to be followed for several years after the data acquisition to obtain a sufficiently reliable disease label (pMCI or sMCI). Few recent studies (listed in Table 1) have investigated the use of different semi-supervised approaches for diagnosis of AD in different stages of the disease. In Zhang and Shen (2011), MCI subjects' data were used as unlabeled data to improve the classification performance in discriminating AD versus NC subjects. They achieved a significant improvement, the AUC score increased from 0.946 to 0.985, which is high for discriminating AD vs. NC subjects. Ye et al. (2011), Filipovych and Davatzikos (2011), and Batmanghelich et al. (2011) used AD and NC subjects as labeled data and MCI subjects as unlabeled data and predicted disease-labels for the MCI subjects. In all these studies, the improvement in the predictive performance of the model was significant over supervised learning. The best classification performance in discriminating sMCI versus pMCI using only MRI data was achieved by Ye et al. (2011) with the area under the receiver operating characteristic curve (AUC) equal to 0.73 for prediction of conversion within 0–18 month period. We hypothesize that the classification performance of semi-supervised learning approaches could be improved if MCI subjects who have been followed up for long enough would be used as labeled data. In this work, we develop a semi-supervised classifier for AD conversion prediction in MCI patients based on low density separation (LDS) (Chapelle and Zien, 2005) and by using MRI data of MCI subjects. Our results demonstrate applicability of the proposed semi-supervised method in MRI based AD conversion prediction in MCI patients by achieving a significant improvement compared to a state of the art supervised method (support vector machine (SVM)).
We perform two processing steps in between our voxel based morphometry style preprocessing (Gaser et al., 2013) and the learning of the LDS classifier. First, we remove age-related effects from MRI data before training the classifier to prevent the confounding between AD and age-related effects to brain anatomy. Previously, a similar technique has been used for the classification between AD and NC subjects, but this study has not considered AD-conversion prediction in MCI subjects (Dukart et al., 2011). In addition, the impact of age was studied recently for detecting AD (Coupé et al., 2012) as well as for predicting AD in MCI patients (Eskildsen et al., 2013). Second, we perform feature selection on MRI data independently of the classification procedure using the auxiliary data from AD and NC subjects. Feature selection is an essential part of the combined procedure since the number of features (29,852) available after the image preprocessing significantly exceeds the number of subjects. We assume that AD vs. NC classification is a simplified version of the pMCI vs. sMCI and the same features that are most useful for the simple problem are useful for the complex one. This idea is implemented by applying regularized logistic regression (RLR) (Friedman et al., 2010) on MRI data of AD and NC subjects for finding the image voxels that are best discriminated between AD and NC subjects. Next, we use these selected voxels for predicting conversion to AD within MCI patients. Most of existing studies incorporating feature selection rely only on a dataset of MCI subjects by using it for feature selection and classification task. In particular, previous studies (Ye et al., 2011, Ye et al., 2012, Janoušová et al., 2012) have considered feature selection based on RLR for MCI-to-AD conversion prediction, but the feature selection was performed with the data from MCI subjects not utilizing data from AD and NC subjects. Auxiliary data from AD and NC subjects to aid the classification of MCI subjects have been considered by Cheng et al. (2012) in a domain transfer learning method. Briefly, the method utilizes cross-domain kernel build from target data (MCI subjects) and auxiliary data (AD and NC) subjects to learn a linear support vector machine classifier. As Cheng et al. reduced the number of features to 93 by partitioning each MRI into 93 regions of interest and did not consider feature selection, the approach to use the auxiliary data is different from our approach.
We integrate MRI data with age and cognitive measurements, also acquired at the baseline, for improving the predictive performance of MCI-to-AD conversion. As opposed to several other studies combining MRI with other types of data (Davatzikos et al., 2011, Zhang and Shen, 2012, Shaffer et al., 2013, Cheng et al., 2012, Wang et al., 2013), we purposely avoid using CSF or PET based biomarkers, the former because it requires lumbar puncture, which is invasive and potentially painful for the patient, and the latter because of its limited availability compared to MRI, as well as its cost and radiation exposure (Musiek et al., 2012). Previously, the combination of MRI derived information and cognitive measurements has been considered by Ye et al. (2012) who trained an RLR classifier with standard cognitive measurements and volumes of certain regions of interest as features and Casanova et al. (2013) who combined outputs of two classifiers, one trained based on MRI and the other trained based on cognitive measurements, based on a sum-rule for the classifier combination. In order to use more efficiently MRI and basic (age and cognitive) measures, we develop what we call an aggregate biomarker by utilizing two different classifiers, i.e. LDS and random forest (RF), in different stages of the process. We first derive a single real valued biomarker based on MRI data using LDS (our biomarker) and thereafter use this as a feature for the aggregate classifier (RF). We will highlight the importance of using a transductive classifier (e.g., LDS) instead of an inductive one (e.g., a standard SVM) during the first stage of the learning process and provide evidence of the effectiveness of the aggregate biomarker for the AD conversion prediction in MCI patients based on MRI, age and cognitive measures at the baseline.
Section snippets
ADNI data
Data used in this work is obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (http://adni.loni.usc.edu/). The ADNI was launched in 2003 by the National Institute on Aging (NIA), the National Institute of Biomedical Imaging and Bioengineering (NIBIB), the Food and Drug Administration (FDA), private pharmaceutical companies and non-profit organizations, as a $60 million, 5-year public–private partnership. The primary goal of ADNI has been to test whether serial MRI,
MRI biomarker
In this section, we consider the experimental results obtained using the biomarker based on solely MRI data as described in the MRI biomarker section. The feature selection reduced the number of voxels in MRI data from 29,852 to 309 voxels. Fig. 3 shows the locations of the selected 309 voxels overlaid on the standard template. Supplementary Table S1 provides the ranking of the brain regions of the loci of the selected voxels according to the Automatic Anatomical Labeling (AAL) atlas. It can be
Discussion
For the early identification of MCI subjects who are in risk of converting to AD, we developed a new method by applying advanced machine learning algorithms for combining MRI data with standard cognitive test results. First, we presented a new biomarker utilizing only MRI data that was based on a semi-supervised learning approach termed low density separation (LDS). The use of LDS in place of more typical supervised learning approaches based on support vector machines was shown to provide
Acknowledgments
Data collection and sharing for this project was funded by the Alzheimer's Disease Neuroimaging Initiative (ADNI) (National Institutes of Health Grant U01 AG024904) and DOD ADNI (Department of Defense award number W81XWH-12-2-0012). ADNI is funded by the National Institute on Aging, the National Institute of Biomedical Imaging and Bioengineering, and through generous contributions from the following: Alzheimer's Association; Alzheimer's Drug Discovery Foundation; BioClinica, Inc.; Biogen Idec
References (84)
- et al.
How early can we predict Alzheimer's disease
Neurobiol. Aging
(2013) - et al.
Unified segmentation
Neuroimage
(2005) - et al.
The projected effect of risk factor reduction on Alzheimer's disease prevalence
Lancet Neurol.
(2011) The use of the area under the ROC curve in the evaluation of machine learning algorithms
Pattern Recogn.
(1997)- et al.
Individual subject classification for Alzheimer's disease based on incremental learning using a spatial frequency representation of cortical thickness data
Neuroimage
(2012) - et al.
Simultaneous segmentation and grading of anatomical structures for patient's classification: application to Alzheimer's disease
Neuroimage
(2012) - et al.
Automatic classification of patients with Alzheimer's disease from structural MRI: a comparison of ten methods using the ADNI database
Neuroimage
(2011) - et al.
Integration and relative value of biomarkers for prediction of MCI to AD progression: spatial patterns of brain atrophy, cognitive scores, APOE genotype and CSF biomarkers
Neuroimage Clin.
(2014) - et al.
Prediction of MCI to AD conversion, via MRI, CSF biomarkers, and pattern classification
Neurobiol. Aging
(2011) - et al.
Prediction of Alzheimer's disease in subjects with mild cognitive impairment from the ADNI cohort using patterns of cortical thinning
Neuroimage
(2013)
Spatial patterns of brain atrophy in MCI patients, identified via high-dimensional pattern classification, predict subsequent cognitive decline
Neuroimage
Semi-supervised pattern classification of medical images: application to mild cognitive impairment (MCI)
Neuroimage
Estimating the age of healthy subjects from T1-weighted MRI scans using kernel methods: exploring the influence of various parameters
Neuroimage
Partial volume segmentation with Adaptive Maximum a Posteriori (MAP) approach
Neuroimage
Alzheimer's DiseaseNeuroimaging Initiative. Random forest-based similarity measures for multi-modal classification of Alzheimer's disease
Neuroimage
Interpretable whole-brain prediction analysis with GraphNet
Neuroimage
Manifold population modeling as a neuro-imaging biomarker: application to ADNI and ADNI-GO
Neuroimage
Predictive markers for AD in a multi-modality framework: an analysis of MCI progression in the ADNI population
Neuroimage
Early detection of Alzheimer's disease using neuroimaging
Exp. Gerontol.
Direct comparison of fluorodeoxyglucose positron emission tomography and arterial spin labeling magnetic resonance imaging in Alzheimer's disease
Alzheimers Dement.
Sparse logistic regression for whole-brain classification of fMRI data
Neuroimage
Fast and robust parameter estimation for statistical partial volume models in brain MRI
Neuroimage
The Alzheimer's disease neuroimaging initiative: a review of paper published since its inception
Alzheimers Demen.
AddNeuroMed and ADNI: similar patterns of Alzheimer's atrophy and automated MRI classification accuracy in Europe and North America
Neuroimage
Multivariate analysis of MRI data for Alzheimer's disease, mild cognitive impairment and healthy controls
Neuroimage
Combining MRI and CSF measures for classification of Alzheimer's disease and prediction of mild cognitive impairment conversion
Neuroimage
Multimodal classification of Alzheimer's disease and mild cognitive impairment
Neuroimage
Selection bias in gene extraction on the basis of microarray gene-expression data
Proc. Natl. Acad. Sci.
Disease classification and prediction via semi-supervised dimensionality reduction
Development of Alzheimer-related neurofibrillary changes in the neocortex inversely recapitulates cortical myelogenesis
Acta Neuropathol.
Random forests
Mach. Learn.
Semi-supervised discriminant analysis. In Computer Vision, 2007. ICCV 2007
Evaluating the impact of different factors on voxel-based classification methods of ADNI structural MRI brain images
Int. J. Biomed. Datamin.
High dimensional classification of structural MRI Alzheimer's disease data based on large scale regularization
Front. Neuroinform.
Classification of structural MRI images in Alzheimer's disease from the perspective of ill-posed problems
PLoS One
Alzheimer's disease risk assessment using large-scale machine learning methods
PLoS One
Semi-supervised classification by low density separation
Semi-supervised Learning
Domain transfer learning for MCI conversion prediction
MICCAI
Fully automatic hippocampus segmentation and classification in Alzheimer's disease and mild cognitive impairment applied on data from ADNI
Hippocampus
The effect of unlabeled data on generative classifiers, with application to model selection
Comparison and validation of tissue modelization and statistical classification methods in T1-weighted MR brain images
IEEE Trans. Med. Imaging
Cited by (563)
An Alzheimer's disease classification model using transfer learning Densenet with embedded healthcare decision support system
2023, Decision Analytics JournalPredicting long-term progression of Alzheimer’s disease using a multimodal deep learning model incorporating interaction effects
2024, Journal of Translational MedicineMachine learning prediction of future amyloid beta positivity in amyloid-negative individuals
2024, Alzheimer's Research and Therapy
- 1
Data used in preparation of this article were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (http://adni.loni.usc.edu). As such, the investigators within the ADNI contributed to the design and implementation of ADNI and/or provided data but did not participate in analysis or writing of this report. A complete listing of ADNI investigators can be found at http://adni.loni.usc.edu/wp-content/uploads/how_to_apply/ADNI_Acknowledgement_List.pdf.