Multi-source feature learning for joint analysis of incomplete multiple heterogeneous neuroimaging data

Bi-level multi-source learning for heterogeneous block-wise missing data

Bio-imaging technologies allow scientists to collect large amounts of high-dimensional data from multiple heterogeneous sources for many biomedical applications. In the study of Alzheimer's Disease (AD), neuroimaging data, gene/protein expression data, etc., are often analyzed together to improve predictive power. Joint learning from multiple complementary data sources is advantageous, but feature-pruning and data source selection are critical to learn interpretable models from high-dimensional data. Often, the data collected has block-wise missing entries. In the Alzheimer's Disease Neuroimaging Initiative (ADNI), most subjects have MRI and genetic information, but only half have cerebrospinal fluid (CSF) measures, a different half has FDG-PET; only some have proteomic data. Here we propose how to effectively integrate information from multiple heterogeneous data sources when data is block-wise missing. We present a unified "bi-level" learning model for complete multi-source data, and extend it to incomplete data. Our major contributions are: (1) our proposed models unify feature-level and source-level analysis, including several existing feature learning approaches as special cases; (2) the model for incomplete data avoids imputing missing data and offers superior performance; it generalizes to other applications with block-wise missing data sources; (3) we present efficient optimization algorithms for modeling complete and incomplete data. We comprehensively evaluate the proposed models including all ADNI subjects with at least one of four data types at baseline: MRI, FDG-PET, CSF and proteomics. Our proposed models compare favorably with existing approaches.

Recent advances in imaging Alzheimer's disease

Advances in brain imaging technology in the past five years have contributed greatly to the understanding of Alzheimer's disease (AD). Here, we review recent research related to amyloid imaging, new methods for magnetic resonance imaging analyses, and statistical methods. We also review research that evaluates AD risk factors and brain imaging, in the context of AD prediction and progression. We selected a variety of illustrative studies, describing how they advanced the field and are leading AD research in promising new directions.

Accurate multimodal probabilistic prediction of conversion to Alzheimer's disease in patients with mild cognitive impairment

Accurately identifying the patients that have mild cognitive impairment (MCI) who will go on to develop Alzheimer's disease (AD) will become essential as new treatments will require identification of AD patients at earlier stages in the disease process. Most previous work in this area has centred around the same automated techniques used to diagnose AD patients from healthy controls, by coupling high dimensional brain image data or other relevant biomarker data to modern machine learning techniques. Such studies can now distinguish between AD patients and controls as accurately as an experienced clinician. Models trained on patients with AD and control subjects can also distinguish between MCI patients that will convert to AD within a given timeframe (MCI-c) and those that remain stable (MCI-s), although differences between these groups are smaller and thus, the corresponding accuracy is lower. The most common type of classifier used in these studies is the support vector machine, which gives categorical class decisions. In this paper, we introduce Gaussian process (GP) classification to the problem. This fully Bayesian method produces naturally probabilistic predictions, which we show correlate well with the actual chances of converting to AD within 3 years in a population of 96 MCI-s and 47 MCI-c subjects. Furthermore, we show that GPs can integrate multimodal data (in this study volumetric MRI, FDG-PET, cerebrospinal fluid, and APOE genotype with the classification process through the use of a mixed kernel). The GP approach aids combination of different data sources by learning parameters automatically from training data via type-II maximum likelihood, which we compare to a more conventional method based on cross validation and an SVM classifier. When the resulting probabilities from the GP are dichotomised to produce a binary classification, the results for predicting MCI conversion based on the combination of all three types of data show a balanced accuracy of 74%. This is a substantially higher accuracy than could be obtained using any individual modality or using a multikernel SVM, and is competitive with the highest accuracy yet achieved for predicting conversion within three years on the widely used ADNI dataset.

•

Prediction of MCI to AD conversion using ADNI data and Gaussian processes.

•

74% accuracy, 0.795 area under ROC curve for predicting conversion within 3 years.

•

Gaussian processes allow automatic parameter tuning including multimodal weights.

•

Statistically significant improvement for multimodal vs best unimodal prediction.

•

Probabilistic interpretation of results to better reflect continuum of disease.

Applying tensor-based morphometry to parametric surfaces can improve MRI-based disease diagnosis

Many methods have been proposed for computer-assisted diagnostic classification. Full tensor information and machine learning with 3D maps derived from brain images may help detect subtle differences or classify subjects into different groups. Here we develop a new approach to apply tensor-based morphometry to parametric surface models for diagnostic classification. We use this approach to identify cortical surface features for use in diagnostic classifiers. First, with holomorphic 1-forms, we compute an efficient and accurate conformal mapping from a multiply connected mesh to the so-called slit domain. Next, the surface parameterization approach provides a natural way to register anatomical surfaces across subjects using a constrained harmonic map. To analyze anatomical differences, we then analyze the full Riemannian surface metric tensors, which retain multivariate information on local surface geometry. As the number of voxels in a 3D image is large, sparse learning is a promising method to select a subset of imaging features and to improve classification accuracy. Focusing on vertices with greatest effect sizes, we train a diagnostic classifier using the surface features selected by an L1-norm based sparse learning method. Stability selection is applied to validate the selected feature sets. We tested the algorithm on MRI-derived cortical surfaces from 42 subjects with genetically confirmed Williams syndrome and 40 age-matched controls, multivariate statistics on the local tensors gave greater effect sizes for detecting group differences relative to other TBM-based statistics including analysis of the Jacobian determinant and the largest eigenvalue of the surface metric. Our method also gave reasonable classification results relative to the Jacobian determinant, the pair of eigenvalues of the Jacobian matrix and volume features. This analysis pipeline may boost the power of morphometry studies, and may assist with image-based classification.

Maximizing power to track Alzheimer's disease and MCI progression by LDA-based weighting of longitudinal ventricular surface features

We propose a new method to maximize biomarker efficiency for detecting anatomical change over time in serial MRI. Drug trials using neuroimaging become prohibitively costly if vast numbers of subjects must be assessed, so it is vital to develop efficient measures of brain change. A popular measure of efficiency is the minimal sample size (n80) needed to detect 25% change in a biomarker, with 95% confidence and 80% power. For multivariate measures of brain change, we can directly optimize n80 based on a Linear Discriminant Analysis (LDA). Here we use a supervised learning framework to optimize n80, offering two alternative solutions. With a new medial surface modeling method, we track 3D dynamic changes in the lateral ventricles in 2065 ADNI scans. We apply our LDA-based weighting to the results. Our best average n80-in two-fold nested cross-validation-is 104 MCI subjects (95% CI: [94,139]) for a 1-year drug trial, and 75AD subjects [64,102]. This compares favorably with other MRI analysis methods. The standard "statistical ROI" approach applied to the same ventricular surfaces requires 165 MCI or 94AD subjects. At 2 years, the best LDA measure needs only 67 MCI and 52AD subjects, versus 119 MCI and 80AD subjects for the stat-ROI method. Our surface-based measures are unbiased: they give no artifactual additive atrophy over three time points. Our results suggest that statistical weighting may boost efficiency of drug trials that use brain maps.

Sparse Methods for Biomedical Data

Following recent technological revolutions, the investigation of massive biomedical data with growing scale, diversity, and complexity has taken a center stage in modern data analysis. Although complex, the underlying representations of many biomedical data are often sparse. For example, for a certain disease such as leukemia, even though humans have tens of thousands of genes, only a few genes are relevant to the disease; a gene network is sparse since a regulatory pathway involves only a small number of genes; many biomedical signals are sparse or compressible in the sense that they have concise representations when expressed in a proper basis. Therefore, finding sparse representations is fundamentally important for scientific discovery. Sparse methods based on the [Formula: see text] norm have attracted a great amount of research efforts in the past decade due to its sparsity-inducing property, convenient convexity, and strong theoretical guarantees. They have achieved great success in various applications such as biomarker selection, biological network construction, and magnetic resonance imaging. In this paper, we review state-of-the-art sparse methods and their applications to biomedical data.