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: | Federico Nemmi |
| Institution: | Inserm |
| Department: | UMR1214 - ToNIC |
| Country: | |
| Proposed Analysis: | Previous studies have combined multivariate analysis and multimodal MRI (e.g. combination of structural, diffusion and resting state fMRI), or multivariate analysis and PET imaging (e.g. FDG and AV45 imaging) in order to discriminate between healthy controls (HC), patients with Mild Cognitive Impairment (MCI) and Alzheimer's Disease (AD) patients. Some studies have also combined MRI and PET imaging, but usually using a region-of-interest approach. We have recently adapted a multimodal multivariate data-driven voxel-wise pipeline (Nemmi et al., Human Brain Mapping, 2019; Nemmi et al., Neuroimage:Clinical, 2019) that we applied to MRI data. We would like to expand on our previous work and further adapt the pipeline to include PET data, biological data (e.g. CSF markers) and clinical/neuropsychological data. In details, our pipeline is composed of the following steps: - Matrix reshaping and range normalization Separately for each modality, images are reshaped from 3D matrix to 2D matrix with subjects x voxels dimensions after being masked for the relevant masks (e.g. grey matter mask for gm and FDG, white matter mask for FA). These matrices are then normalized so that the values are comprised between 0 and 1. - Variance thresholding We adopt a simple variance features reduction step in which, for each modality, we eliminated the 25% of features with the lowest variance. - Relieff based features selection Relieff is a features selection algorithm that is widely used in the machine learning literature. It estimates a weight for each feature by comparing, for each case, the distance of the closest intra and inter-class cases in that feature space and increasing the weight if the distance is greater for the inter-class than for the intra-class case. For each modality, we submit the features surviving the variance threshold to the Relieff algorithm. In order to select the most relevant features we use a scree test approach. We calculate the selection threshold as the first minimum of the second derivative of the sorted and smoothed Relieff weights. This is equivalent to find the point at which the speed of the function approach to zero - Spatial clustering of the features Features from brain imaging are intrinsically non-independent due to the fact that voxels that lie close usually belong to the same anatomical/functional region. In the light of this, spatially cluster features (i.e. voxels) that are close to each other is an effective and meaningful way of reducing the number of features. For each modality, we submit the features surviving the Relieff threshold to a spatial clustering algorithm: contiguous voxels are assigned the same cluster. Finally, we extracted the average signal for each cluster, thus effectively reducing the number of features for each modality from hundreds to tens. - Merging of modalities and subset selection After the spatial clustering step, we merge all the modalities in one matrix (having dimensions [subjects] x [N of clusters from all modalities]. Even after Relieff selection and spatial clustering, some clusters may be not very informative, and some clusters may convey redundant information. For this reason, we perform subset selection based on correlation. This selection step is aimed to find the subset of features (i.e. clusters in the case at hand) that maximises the predictive power relative to the outcome while minimizing redundancy among clusters (measured as collinearity) The outcome of this step is a subset of clusters in the order of tens (or lower): the low number of clusters help the interpretability of the model while maximising the discriminative power. - Fitting of the model Finally, the model is fitted using the sequential minimal optimization (SMO) algorithm with a polynomial kernel. The advantage of our pipeline is that it both maximise performance and interpretability of the model, since at the end point of the pipeline between model fitting we end up with less than 10 clusters per modality, thus having compact and interpretable models. All the steps previously described have been optimized for MRI data and we are working on the optimization for neuropsychological data using an in-house dataset. We still need to optimize the steps for biological and PET data. ADNI would be for us a great opportunity to have a dataset of conspicuous size that include MRI, PET, biological data and neuropsychological/clinical data. We will use MRI data (T1, DWI, rs-fMRI), PET data (FDG, AV45) biological data (CSF) and neuropsychological data and we will adapt our pipeline for 4 different tasks: -discrimination between HC, MCI and AD (and possibly between early and late MCI) -prediction of concurrent clinical scale (e.g. MMSE, ADAS13) -prediction of conversion -prediction of severity of cognitive decline Note that the coding for the pipeline is already written and already publicly available at https://github.com/fnemmi-tonic. |
| Additional Investigators |
