Multi-Source Sparse Learning to Identify MCI and Predict Decline



 DESCRIPTION (provided by applicant): Patients with Mild Cognitive Impairment (MCI) are at high risk of progression to dementia. MCI offers an opportunity to target the disease process early. Clinicians and researchers are intensifying their efforts to detect MCI pre-symptomatically in order to develop preventive treatments. These efforts generate a large amount of data - brain images of multiple modalities, and proteomics, genetic, and neurocognitive data that provide unprecedented opportunities to investigate MCI-related questions with greater precision and predictive power. Understanding its importance, NIH in 2003 funded the Alzheimer's Disease Neuroimaging Initiative (ADNI) to facilitate scientific evaluation of various biomarkers for the onset and progression of MCI and AD. To realize such an ambitious vision, there is an urgent need for multi-source fusion and disease biomarker discovery frameworks. While promising, large volumes of incomplete data from multiple heterogeneous data sources pose huge challenges to scientists and engineers. For instance, the ADNI-1 data (like many other large datasets) exhibit a block-wise missing pattern: most subjects have MRI, genetic information; about half of the subjects have CSF measures; a different half of the subjects have FDG-PET; and some subjects have proteomics data. Although many bioinformatics tools are available, no existing tools offer an effective way to fuse multi-source incomplete data for disease biomarker discovery. Here we aim to develop a novel computational framework to integrate and analyze multiple, heterogeneous, large volume, incomplete biomedical data for early detection of MCI. Our 4 primary aims are: (1) Develop novel structured sparse learning formulations for multi-source fusion. The computational methods will identify biomarkers to correlate multi-source data with MCI. Novel sparse screening methods will be developed to scale the proposed formulations to very high-dimensional data. (2) Develop computational methods to integrate network data. We will develop novel methods for incorporating existing biological knowledge such as pathways represented as networks into the prediction model. The network structure will be used as prior knowledge to constrain model parameters, to further improve predictive power. (3) Develop computational methods to integrate multiple incomplete data sources. The proposed computational framework will integrate multiple heterogeneous data with a block-wise missing pattern. The proposed framework formulates the multiple incomplete data source fusion problem as a multi-task learning problem by first decomposing the prediction problem into a set of tasks, then building the models for all tasks simultaneously. (4) Develop and disseminate software tools for multi-source fusion and biomarker identification. The software tools will be used for early detection of MCI and will be validated by several clinical research projects. Our open source software will be made freely available to the research communities, including our large community of existing users. One of our current packages, SLEP, has ~4,500 active users from ~25 countries. Our software tools will be easily adaptable for analyzing multi-source data from other neurological and psychiatric disorders.