Sparse learning and stability selection for predicting MCI to AD conversion using baseline ADNI data

Center for Evolutionary Medicine and Informatics, The Biodesign Institute, Arizona, State University, Tempe, AZ, USA. .
BMC Neurology (Impact Factor: 2.04). 06/2012; 12(1):46. DOI: 10.1186/1471-2377-12-46
Source: PubMed


Patients with Mild Cognitive Impairment (MCI) are at high risk of progression to Alzheimer’s dementia. Identifying MCI individuals with high likelihood of conversion to dementia and the associated biosignatures has recently received increasing attention in AD research. Different biosignatures for AD (neuroimaging, demographic, genetic and cognitive measures) may contain complementary information for diagnosis and prognosis of AD.

We have conducted a comprehensive study using a large number of samples from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) to test the power of integrating various baseline data for predicting the conversion from MCI to probable AD and identifying a small subset of biosignatures for the prediction and assess the relative importance of different modalities in predicting MCI to AD conversion. We have employed sparse logistic regression with stability selection for the integration and selection of potential predictors. Our study differs from many of the other ones in three important respects: (1) we use a large cohort of MCI samples that are unbiased with respect to age or education status between case and controls (2) we integrate and test various types of baseline data available in ADNI including MRI, demographic, genetic and cognitive measures and (3) we apply sparse logistic regression with stability selection to ADNI data for robust feature selection.

We have used 319 MCI subjects from ADNI that had MRI measurements at the baseline and passed quality control, including 177 MCI Non-converters and 142 MCI Converters. Conversion was considered over the course of a 4-year follow-up period. A combination of 15 features (predictors) including those from MRI scans, APOE genotyping, and cognitive measures achieves the best prediction with an AUC score of 0.8587.

Our results demonstrate the power of integrating various baseline data for prediction of the conversion from MCI to probable AD. Our results also demonstrate the effectiveness of stability selection for feature selection in the context of sparse logistic regression.

Download full-text


Available from: Victor Lobanov
  • Source
    • " ~ cjlin/libsvm/ ) and the regularization parameter was always selected based on CV in the training set. The stability selection was based on an in-house Matlab implementation following the guidelines of (Ye et al 2012) and it was followed by SVM classification. For this reason, when referring to Elastic-Net or LASSO later on, we do not typically mean stability selection. "
    [Show abstract] [Hide abstract]
    ABSTRACT: We present a comparative split-half resampling analysis of various data driven feature selection and classification methods for the whole brain voxel-based classification analysis of anatomical magnetic resonance images. We compared support vector machines (SVMs), with or without filter based feature selection, several embedded feature selection methods and stability selection. While comparisons of the accuracy of various classification methods have been reported previously, the variability of the out-of-training sample classification accuracy and the set of selected features due to independent training and test sets have not been previously addressed in a brain imaging context. We studied two classification problems: 1) Alzheimer’s disease (AD) vs. normal control (NC) and 2) mild cognitive impairment (MCI) vs. NC classification. In AD vs. NC classification, the variability in the test accuracy due to the subject sample did not vary between different methods and exceeded the variability due to different classifiers. In MCI vs. NC classification, particularly with a large training set, embedded feature selection methods outperformed SVM-based ones with the difference in the test accuracy exceeding the test accuracy variability due to the subject sample. The filter and embedded methods produced divergent feature patterns for MCI vs. NC classification that suggests the utility of the embedded feature selection for this problem when linked with the good generalization performance. The stability of the feature sets was strongly correlated with the number of features selected, weakly correlated with the stability of classification accuracy, and uncorrelated with the average classification accuracy.
    Full-text · Article · Jan 2016 · Neuroinformatics
  • Source
    • "Reported areas under the ROC curve (AUC) range from .647 to .859. Importantly , both the study that reports the highest accuracy (Devanand, et al., 2008) and the study that reports the highest AUC (Ye, et al., 2012) include both cognitive and biological measures in their classifiers. "
    [Show abstract] [Hide abstract]
    ABSTRACT: Objective We constructed random forest classifiers employing either the traditional method of scoring semantic fluency word lists or new methods. These classifiers were then compared in terms of their ability to diagnose Alzheimer disease (AD) or to prognosticate among individuals along the continuum from cognitively normal (CN) through mild cognitive impairment (MCI) to AD. Method Semantic fluency lists from 44 cognitively normal elderly individuals, 80 MCI patients, and 41 AD patients were transcribed into electronic text files and scored by four methods: traditional raw scores, clustering and switching scores, “generalized” versions of clustering and switching, and a method based on independent components analysis (ICA). Random forest classifiers based on raw scores were compared to “augmented” classifiers that incorporated newer scoring methods. Outcome variables included AD diagnosis at baseline, MCI conversion, increase in Clinical Dementia Rating-Sum of Boxes (CDR-SOB) score, or decrease in Financial Capacity Instrument (FCI) score. ROC curves were constructed for each classifier and the area under the curve (AUC) was calculated. We compared AUC between raw and augmented classifiers using Delong’s test and assessed validity and reliability of the augmented classifier. Results Augmented classifiers outperformed classifiers based on raw scores for the outcome measures AD diagnosis (AUC 0.97 vs. 0.95), MCI conversion (AUC 0.91 vs. 0.77), CDR-SOB increase (AUC 0.90 vs. 0.79), and FCI decrease (AUC 0.89 vs. 0.72). Measures of validity and stability over time support the use of the method. Conclusion Latent information in semantic fluency word lists is useful for predicting cognitive and functional decline among elderly individuals at increased risk for developing AD. Modern machine learning methods may incorporate latent information to enhance the diagnostic value of semantic fluency raw scores. These methods could yield information valuable for patient care and clinical trial design with a relatively small investment of time and money.
    Full-text · Article · Jun 2014 · Cortex
  • Source
    • "Neuroimages allow the identification of brain changes and have been used for automated diagnosis of AD and MCI (Silveira and Marques, 2010; Ye et al., 2012). Due to the high variability of the pattern of brain degeneration in AD and MCI, the analysis of brain images is a very difficult task. "
    [Show abstract] [Hide abstract]
    ABSTRACT: Alzheimer's disease is a type of dementia that mainly affects elderly people, with unknown causes and no effective treatment up to date. The diagnosis of this disease in an earlier stage is crucial to improve patients' life quality. Current techniques focus on the analysis of neuroimages, like FDG-PET or MRI, to find changes in the brain activity. While high accuracies can be obtained by combining the analysis of several types of neuroimages, they are expensive and not always available for medical analysis. Achieving similar results using only 3-D FDG-PET scans is therefore of huge importance. While directly applying classifiers to the FDG-PET scan voxel intensities can lead to good prediction accuracies, it results in a problem that suffers from the curse of dimensionality. This paper thus proposes a methodology to identify regions of interest by segmenting 3-D FDG-PET scans and extracting features that represent each of those regions of interest, reducing the dimensionality of the space. Experimental results show that the proposed methodology outperforms the one using voxel intensities despite only a small number of features is needed to achieve that result.
    Full-text · Conference Paper · Jan 2014
Show more