The goal of my research is to develop novel statistical learning and signal processing methods to enable useful prediction with biomedical data (in particular high-dimensional time series, imaging, and genomic data), with a focus on the brain.
Computational methods for network imaging genomics
The goal of the research strand is to establish links between very different biological levels - large-scale brain networks and nano-scale changes in gene expression or variations in genes (SNPs). Our recent paper in Science show our first results in this area.
12/06: Note the Allen REST API calls for mouse ISH data in the Science paper supplementary have improper line breaks, I will soon provide fixed links here for your convenience.
Imaging markers for neurodegenerative and neuropsychiatric diseases
In this research axis, we are developing novel imaging-based markers for several brain diseases. We are particularly interested in using functional magnetic resonance imaging (fMRI).
For example, we have recently shown that MCI and Alzheimer disease patients have significantly slower cerebrovascular reactivity (blood vessel dilation and contraction) than healthy controls when undergoing a CO2 challenge in fMRI. Since this velocity is also tied to cognitive performance, and the acquisition setup is relatively cheap and easy, we hope that this can be made part of the initial dementia workup in clinical settings.
Modelling and inference with brain graphs
The goal of this EU-funded project (Marie Curie IOF #299500) is to develop predictive methods for early clinical diagnosis of neurodegenerative diseases from brain connectivity data.
We focus on developing theoretical advances supported by empirical experiments on synthetic and real neuroimaging data of MS and AD patients, in particular asking
- How can we extract connectivity graphs from neuroimaging data?
- How can we perform modelling and inference on connectivity graphs?
We also explore the practical limitations of the methods:
- Aside from pathology, what changes in population or processing most likely affect connectivity data?
- What are possible ways of mitigating these confounding factors?
Predictive multivariate brain mapping with linear discriminants
The goal of this line of research is to develop information mapping procedures from multivariate linear classifiers.
Like contrasts in Statistical Parametric Maps, discriminative weight vectors carry differential information, but there are specific issues with using weight vectors for mapping. In particular, we look at the types of information maps that can be computed, the handling of multiple weight vectors in the multi-class case, and statistical thresholding procedures in high-dimensional settings.
Learning with non-i.i.d. multi-site data
In this project funded by the Bavaria California Technology Center we seek to improve statistical learning when imaging data comes from different acquisition hardware.
MR physics is such that it is difficult to successfully combine data from multiple sites, even when the equipment is calibrated, and from the same make and model. To develop the next generation of data-driven tools, advanced methods are needed to successfully exploit multi-site data with minimum loss in statistical power.
Models of resting-state and task-based functional imaging data for dementia
In this project funded by the Swiss National Science Foundation we build predictive models of functional imaging data to permit early diagnosis of dementia.
By using a recently developed memory-specific task with high sensitivity for the diagnosis of mild cognitive impairment subjects, together with resting-state data, we aim at obtaining improved performance for machine learning-based diagnosis. We take a multi-view perspective, where we treat the task data and resting-state data as different observations (views) of the same brain.