Automated detection of amnestic mild cognitive impairment in community-dwelling elderly adults: A combined spatial atrophy and white matter alteration approach
Highlights
► Amnestic mild cognitive impairment is identified using neuroimaging. ► A combination of gray and white matter features is used. ► The method is evaluated in community-dwelling elderly individuals. ► Excellent classification performance characteristics are achieved. ► Socio-demographic, lifestyle and health factors help explain misclassifications.
Introduction
Mild cognitive impairment (MCI) is considered to be a transitional stage between normal aging and dementia. The amnestic form of MCI (aMCI) is characterized by memory loss and particularly associated with an increased risk of developing Alzheimer's disease (AD). Indeed, with a 10–15% annual incidence of conversion to AD (Petersen et al., 2001), aMCI is now receiving considerable attention. Identifying predictive markers for aMCI will facilitate the development of treatments for preventing AD or slowing its progression.
By detecting subtle brain changes, neuroimaging is a promising tool for identifying individuals in the early stages of cognitive disorders, particularly those exhibiting normal performance on neuropsychological tests because of cognitive reserve (Mueller et al., 2005). Magnetic resonance imaging (MRI) has been widely used to investigate how aMCI differs from normal aging, typically in terms of volumetric assessments (Chetelat and Baron, 2003, Fennema-Notestine et al., 2009, Ridha et al., 2008). Structural atrophy has been found in individuals with aMCI, and is believed to develop initially and be most severe in the entorhinal cortex and hippocampus (Braak and Braak, 1991, Chetelat and Baron, 2003, Fennema-Notestine et al., 2009). Enlargement of the lateral ventricles has also been reported (Fennema-Notestine et al., 2009, Nestor et al., 2008). These structural changes are potential markers of aMCI and progression towards AD.
While changes in brain structure are important contributors to cognitive dysfunction, the capacity for information flow within and between particular structures must not be overlooked (Wen et al., 2011). AD is reported to be a progressive impairment of fiber-track connectivity, characterized by a loss of afferent and efferent connections between regional neocortical areas associated with pyramidal neuron death (Morrison and Hof, 2002). In seeking to establish the causes of cognitive dysfunction, it is therefore important to examine white matter (WM) integrity. WM integrity can be assessed using fractional anisotropy (FA) values obtained with diffusion tensor imaging (DTI). DTI studies comparing individuals with aMCI and normal controls (NCs) have reported finding that aMCI is associated with significant reductions in FA in the splenium of the corpus callosum (Chua et al., 2009, Parente et al., 2008, Zhuang et al., 2010), crus of the fornix (Zhuang et al., 2010) and posterior cingulum (Chua et al., 2009). These WM changes have the potential to serve as predictive markers of aMCI.
There has been a recent interest in using pattern classification techniques to automatically discriminate individuals with AD or aMCI from NCs. Pattern recognition techniques can capture multivariate relationships among various anatomical regions, which have facilitated their ability to successfully discriminate AD from normal aging. However, the development of algorithms to discriminate aMCI from normal aging is more challenging because the structural differences are more subtle and uncertain. There are several approaches in the use of pattern classification techniques to automatically discriminate individuals with MCI from NCs, which can be grouped into 3 categories depending on the data modalities used. The first category uses features extracted from T1-weighted images. Studies within this category include Davatzikos et al. (2008), who reported 90% accuracy in discriminating between MCI and normal cognition among 30 individuals using voxel-based MRI features, and Gerardin et al. (2009), who used spherical harmonics (SPHARM) coefficients to model hippocampal shape and correctly classify 83% (sensitivity, 83%; specificity, 84%) of 48 individuals as either having or not having MCI. The second category involves features extracted with DTI, and is exemplified by Wee et al. (2011), who used 6 WM connectivity parameters (fiber count, FA, mean diffusivity and the principle diffusivities λ1, λ2 and λ3) to make accurate MCI classifications in 88.9% of 27 individuals. The third category of pattern recognition technique uses a combination of different data modes. By combining 3 data modes (MRI, positron emission tomography (PET) and cerebrospinal fluid (CSF)), Zhang et al. (2011) accurately classified 76.4% (sensitivity, 81.8%; specificity, 66%) of 151 participants as having either MCI or normal cognitive functioning. However, each of the above-mentioned studies was limited by the sample used, which was either small (Davatzikos et al., 2008, Gerardin et al., 2009, Wee et al., 2011) or clinic-based (Zhang et al., 2011), and in 2 instances also comprised socio-demographically matched subjects (Davatzikos et al., 2008, Wee et al., 2011). From the standpoint of early detection, the ability to reliably distinguish between individuals with aMCI and healthy elderly would be ideal. A further consideration is that having simple and relatively few markers might make such screening more practical.
The present study aims to use a T1-weighted- and DTI-based data-driven approach to identify markers of aMCI in community-dwelling elderly individuals. Reliable and valid volumetric analysis requires accurate segmentation. The probabilistic-based software package FreeSurfer (FS) (Fischl et al., 2002, Fischl et al., 2004) is freely available, but its reliability and validity, especially in relation to subcortical volumetrics, can be improved upon by combining it with the Large Deformation Diffeomorphic Metric Mapping (LDDMM)-based label-propagation method. This FS + LDDMM subcortical segmentation method was proposed by Khan et al. (2008), and has greater reliability and accuracy than FS alone (Khan et al., 2008, Wang et al., 2009). Accordingly, in the present study we employed the FS + LDDMM method to extract brain subcortical volumetric features, which have the potential to capture sensitive differences between normal aging and aMCI. Furthermore, to the best of our knowledge, this is the first study to use a combined T1-weighted and WM DTI approach for the automated identification of aMCI, and we expect that including both morphometric changes and WM alterations will facilitate detection of the subtle brain abnormalities that are likely to characterize this condition.
A secondary aim of the present study, which does not appear to have been previously addressed, is to understand reasons for misclassifications by automated systems designed to discriminate between impaired and normal cognitive functioning in elderly samples. Individuals may be misclassified if exhibiting cognitive impairment without substantial structural changes, or if maintaining normal cognitive functioning despite the presence of these. Either situation could be associated with socio-demographic, lifestyle, health and other factors that exert independent effects on cognition in the elderly (for factors typically investigated for such effects see Hendrie et al., 2006, Plassman et al., 2010). We examined if any of these factors contributed to misclassifications by our algorithm. Identifying such factors may also help to validate our classification schema, and could offer insights into preventing or slowing the progression of aMCI.
Section snippets
Participants
Participants were drawn from the Sydney Memory and Aging Study (MAS), a longitudinal study of non-demented, community-dwelling individuals aged 70–90 years old at baseline. MAS participants were recruited randomly from areas of Eastern Sydney, Australia via the electoral roll, for which registration is compulsory. Individuals were excluded if they had an adjusted Mini-Mental State Examination (Anderson et al., 2007, Folstein et al., 1975) score < 24, a diagnosis of dementia, mental retardation,
Classification performance
Performance characteristics were determined as the average of 10 classification experiments for each of 3 approaches to classification: volumetric measurements alone, DTI alone, and a combination of volumetric measurements and DTI. The accuracy, sensitivity, specificity and AUC of these approaches are shown in Table 2; ROC curves are shown in Fig. 4. Our results clearly show that the best performance was achieved by the combined volumetric/DTI approach.
The benefits of feature selection were
Classification via imaging features
There are many reports of structural atrophy being a characteristic brain abnormality of MCI (Davatzikos et al., 2008, Desikan et al., 2009, Fan et al., 2008, Vemuri et al., 2008, Wang et al., 2009). Brain changes measured by DTI have also been related to neurodegenerative disease (Rose et al., 2006, Zhang et al., 2007, Zhuang et al., 2010). In the present study, we utilized measures of both gray and white matter in assessing a pattern recognition technique for discriminating between normal
Conclusions
The present study evaluated an automated, data-driven method for identifying individuals with aMCI in a community-based elderly sample, and to the best of our knowledge is the first to do so using a combination of T1-weighted-derived volumetrics and DTI-derived measures of WM alterations. Our study is also novel in identifying various socio-demographic, lifestyle, health and other factors associated with misclassifications by automated systems designed to identify neurological disorders. These
Acknowledgments
This research was supported by National Health and Medical Research Council (NHMRC) Program Grants (IDs 350833 and 510175), an Australian Research Council Discovery Grant (ARC DP-0774213), and a CSC-Newcastle Scholarship. We thank the MAS participants, interviewers and large study team.
References (68)
- et al.
Effects of sociodemographic and health variables on Mini-Mental State Exam scores in older Australians
Am. J. Geriatr. Psychiatry
(2007) - et al.
Early diagnosis of Alzheimer's disease: contribution of structural neuroimaging
Neuroimage
(2003) - et al.
Age-related changes in white matter lesions, hippocampal atrophy, and cerebral microbleeds in healthy subjects without major cerebrovascular risk factors
J. Stroke Cerebrovasc. Dis.
(2011) - et al.
Diffusion tensor imaging of the posterior cingulate is a useful biomarker of mild cognitive impairment
Am. J. Geriatr. Psychiatry
(2009) - 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.
Detection of prodromal Alzheimer's disease via pattern classification of magnetic resonance imaging
Neurobiol. Aging
(2008) - et al.
Multivariate examination of brain abnormality using both structural and functional MRI
Neuroimage
(2007) - et al.
Spatial patterns of brain atrophy in MCI patients, identified via high-dimensional pattern classification, predict subsequent cognitive decline
Neuroimage
(2008) - et al.
Shape differences of the brain ventricles in Alzheimer's disease
Neuroimage
(2006) - et al.
Whole brain segmentation: automated labeling of neuroanatomical structures in the human brain
Neuron
(2002)
Sequence-independent segmentation of magnetic resonance images
Neuroimage
"Mini-mental state". A practical method for grading the cognitive state of patients for the clinician
J. Psychiatr. Res.
Multidimensional classification of hippocampal shape features discriminates Alzheimer's disease and mild cognitive impairment from normal aging
Neuroimage
The NIH Cognitive and Emotional Health Project. Report of the Critical Evaluation Study Committee
Alzheimers Dement.
Hypothetical model of dynamic biomarkers of the Alzheimer's pathological cascade
Lancet Neurol.
FreeSurfer-initiated fully-automated subcortical brain segmentation in MRI using large deformation diffeomorphic metric mapping
Neuroimage
The effects of age and sex on cortical sulci in the elderly
Neuroimage
Selective vulnerability of corticocortical and hippocampal circuits in aging and Alzheimer's disease
Prog. Brain Res.
The Alzheimer's disease neuroimaging initiative
Neuroimaging Clin. N. Am.
Homocysteine, cerebrovascular disease and brain atrophy
J. Neurol. Sci.
Tract-based spatial statistics: voxelwise analysis of multi-subject diffusion data
Neuroimage
Relationship between hippocampal volume and memory ability in healthy individuals across the lifespan: review and meta-analysis
Neuropsychologia
Alzheimer's disease diagnosis in individual subjects using structural MR images: validation studies
Neuroimage
Enriched white matter connectivity networks for accurate identification of MCI patients
Neuroimage
Development and validation of a geriatric depression screening scale—a preliminary-report
J. Psychiatry Res.
Multimodal classification of Alzheimer's disease and mild cognitive impairment
Neuroimage
White matter integrity in mild cognitive impairment: a tract-based spatial statistics study
Neuroimage
White matter damage in Alzheimer disease and its relationship to gray matter atrophy
Radiology
Non-linear optimisation
Non-linear registration, aka spatial normalization
Neuropathological stageing of Alzheimer-related changes
Acta Neuropathol.
A tutorial on support vector machines for pattern recognition
Data Min. Knowledge Discov.
LIBSVM: a library for support vector machines
Feature ranking using linear SVM
Cited by (56)
Artificial intelligence in brain MRI analysis of Alzheimer's disease over the past 12 years: A systematic review
2022, Ageing Research ReviewsCitation Excerpt :These include, but are not limited to, the independency and the size of training, validation, and testing subsets, MRI features based upon which the decision is deduced, and methods and data for cross-validation, on which risks of bias can be sufficiently understood (Appendix C) (Wolff et al., 2019). The use of broad variations can increase model generalizability (e.g., applicable to various MCI subtypes), although may lead to comparatively lower testing accuracy as specificity and sensitivity suffer in a small sample (Cui et al., 2012). Insufficient sample size can inevitably accompany deep learning, especially concerning highly complicated brain MRI, even with the continued effort of open-access data collection and sharing.
Deep Fusion of Brain Structure-Function in Mild Cognitive Impairment
2021, Medical Image AnalysisArtificial intelligence and neuropsychological measures: The case of Alzheimer's disease
2020, Neuroscience and Biobehavioral ReviewsCitation Excerpt :The risk of bias tools highlighted some frequent limitations: only a small set of studies included sufficient details about the selection process (i.e., Cui et al., 2012, Quintana et al., 2012, Schmid et al., 2013, Beltrachini et al., 2015, Tabaton et al., 2010, Runtti et al., 2014); some studies relied on data samples of small size (i.e., less than 40 subjects overall, according to Belleville et al. (2017), or less than 20 subjects per diagnostic class for binary comparisons in order to perform proper training of ML algorithms); among the works included in this meta-analysis, the papers by König et al. (2015), Fasano et al. (2018), and Jarrold et al. (2014) fall in this category;
Diffusion-weighted imaging (DWI) tractography and Alzheimer’s disease
2020, Diagnosis and Management in Dementia: The Neuroscience of Dementia, Volume 1Quantifying deep grey matter atrophy using automated segmentation approaches: A systematic review of structural MRI studies
2019, NeuroImageCitation Excerpt :Although FreeSurfer focusses on surface-based analyses and cortical surface reconstruction using deformable 3D meshes (Fischl, 2012), it additionally performs segmentation of the DGM using an aligned probabilistic atlas and known intensity statistics and spatial relationships (observed from a training set) to compute a probability of DGM class (Fischl et al., 2002). DGM segmentation using FreeSurfer has been used for investigating changes related to MS (Al-Radaideh et al., 2018; Galego et al., 2015; Krämer et al., 2015; Ramasamy et al., 2009), AD (Cui et al., 2012; Lehmann et al., 2010; Oliveira et al., 2010; Westman et al., 2012), PD (Bilgic et al., 2012; Choi et al., 2015; Lee et al., 2017; Messina et al., 2011; Tinaz et al., 2011; Vasconcellos et al., 2018; Wang et al., 2018), schizophrenia (Bollettini et al., 2018; Goldman et al., 2008; Juuhl-Langseth et al., 2012; Rich et al., 2016; van Erp et al., 2016; Vaskinn et al., 2015), bipolar disorder (Abramovic et al., 2016; Arumugham et al., 2017; Sacchet et al., 2015), HD (Majid et al., 2011; Rupp et al., 2012), temporal lobe epilepsy (Alhusaini et al., 2013; McDonald et al., 2008), human immunodeficiency virus (Corrêa et al., 2016b, 2016a), traumatic brain injury (Hellstrøm et al., 2017), paediatric brain injury (Bigler et al., 2010) and healthy controls (Dennison et al., 2013; Grazioplene et al., 2015; Kang et al., 2015; Kremen et al., 2010; Liem et al., 2015; Long et al., 2012; Ostby et al., 2009; Potvin et al., 2016; Winkelmann et al., 2016). Test-retest data on healthy participants showed FreeSurfer had a high reliability for DGM segmentations despite varying pulse sequence (Wonderlick et al., 2009).
Determination of Imaging Biomarkers to Decipher Disease Trajectories and Differential Diagnosis of Neurodegenerative Diseases (DIsease TreND)
2018, Journal of Neuroscience Methods