There are many active research projects accessing and applying shared ADNI data. Use the search above to find specific research focuses on the active ADNI investigations. This information is requested annually as a requirement for data access.
| Principal Investigator | |
| Principal Investigator's Name: | Jorge Garcia Condado |
| Institution: | IIS Biocruces Bizkaia |
| Department: | Computational Neuroimaging |
| Country: | |
| Proposed Analysis: | Unsupervised clustering will be used to partition the patient data into groups based on neuroimaging derived phenotypes. Spectral clustering is commonly used however it has a high computational complexity which makes it infeasible for such large datasets. We propose to use deep unsupervised methods combined with multi-modal fusion techniques. IDPs derived from each brain MRI modality are different views of the same data. To take advantage of the richness of data we propose to use multi-modal fusion techniques. A combination of multi-view techniques using Canonical Correlation Analysis [1], Joint Nonnegative Matrix Factorization [2] or Structured Sparsity [3] will be considered. Due to the complexity of the data simple clustering techniques such as k-Means algorithm are not sufficient. We propose to combine the multi-view fusion techniques with deep unsupervised techniques like VaDE [4] which carry out representation learning at the same time as they cluster and are scalable. These have the added advantage that they can generate new samples from each cluster to later on train supervised machine learning algorithms. To determine the number of clusters several techniques will be considered. To determine the number of clusters an initial investigation using the Silhouette coefficient on the latent representation of each subject will be undertaken. Community detection algorithms such as Louvain community detection [5] or Leiden community detection [6] can be used to validate the number of clusters. Finally, once the clusters are generated the similarities and differences between patients within each cluster will be studied. This will be both in terms of the IDPs as well as the diseases outcomes reported. [1] ICML p.129-136 (2009) [2] ICDM p.252-260 (2013) [3] ICML p.354-360 (2013) [4] IJCAI .1965-1972 (2017) [5] Science 328 p.876-878 (2010) [6] Scientific report 9 p.5233 (2018) |
| Additional Investigators | |
| Investigator's Name: | Jesus M. Cortes Diaz |
| Proposed Analysis: | Unsupervised clustering will be used to partition the patient data into groups based on neuroimaging derived phenotypes. Spectral clustering is commonly used however it has a high computational complexity which makes it infeasible for such large datasets. We propose to use deep unsupervised methods combined with multi-modal fusion techniques. IDPs derived from each brain MRI modality are different views of the same data. To take advantage of the richness of data we propose to use multi-modal fusion techniques. A combination of multi-view techniques using Canonical Correlation Analysis [1], Joint Nonnegative Matrix Factorization [2] or Structured Sparsity [3] will be considered. Due to the complexity of the data simple clustering techniques such as k-Means algorithm are not sufficient. We propose to combine the multi-view fusion techniques with deep unsupervised techniques like VaDE [4] which carry out representation learning at the same time as they cluster and are scalable. These have the added advantage that they can generate new samples from each cluster to later on train supervised machine learning algorithms. To determine the number of clusters several techniques will be considered. To determine the number of clusters an initial investigation using the Silhouette coefficient on the latent representation of each subject will be undertaken. Community detection algorithms such as Louvain community detection [5] or Leiden community detection [6] can be used to validate the number of clusters. Finally, once the clusters are generated the similarities and differences between patients within each cluster will be studied. This will be both in terms of the IDPs as well as the diseases outcomes reported. [1] ICML p.129-136 (2009) [2] ICDM p.252-260 (2013) [3] ICML p.354-360 (2013) [4] IJCAI .1965-1972 (2017) [5] Science 328 p.876-878 (2010) [6] Scientific report 9 p.5233 (2018) |
