SVM based group data analysis for drug abuse disorder ASL perfusion study

University Of Pennsylvania, Philadelphia PA

Investigators

Linked publications & trials

Abstract

[unreadable] DESCRIPTION (provided by applicant): Arterial spin labeling (ASL) perfusion MRI provides a potentially extremely useful tool for drug abuse/addiction studies due to the noninvasive cerebral blood flow (CBF) measurement and immunity to low frequency MR signal drift effects that degrade functional MRI based on BOLD contrast over time. However, data processing through conventional univariate general linear model (GLM) based methods has proved extremely challenging for ASL data due to its intrinsic low SNR, the increased motion and agitation typically present during drug craving states, and the "patchy flow" defects commonly found in the brains of patients with certain types of chronic substance abuse (e.g., cocaine). Assuming a linear brain response to the functional stimuli and ignoring the abundant spatial brain activity coupling, the standard GLM is initially sub-optimal for fMRI data analysis. Modeling each voxel's time series with a canonical hemodynamic response function (HRF), it may be further ill-posed to drug abuse studies since the actual shape of HRF may differ significantly in the addicted patient's brain from the canonical one and may differ significantly from patient to patient, from voxel to voxel. A more powerful data analysis method is fundamentally demanded for drug abuse/addiction ASL perfusion studies. The goal of this project is to develop a multivariate and brain response modeling-free fMRI data processing method and to use it for revealing the drug craving related brain activation patterns within existing drug abuse/addiction ASL perfusion fMRI data in our center. A machine-learning algorithm, the support vector machine (SVM), will be used to extract a spatial discriminance map between different experimental conditions for each subject, and a statistic framework will be provided to give a population inference about the extracted discriminance (Aim 1). We hypothesize that this machine-learning based data processing will increase the detection sensitivity of ASL perfusion fMRI as compared to the conventional GLM approach because of the data driven nature and multivariate processing of SVM. To verify this hypothesis and to validate the sensitivity, specificity and reliability of the proposed methods, we propose to acquire 40 normal controls' null-hypothesis ASL perfusion fMRI data and sensory-motor task data in Aim 2. The last but the most important aim (Aim 3) is to apply the proposed method to analyze the existing drug abuse/addiction ASL perfusion fMRI data in our center. In addition to the basic science implications for ASL perfusion fMRI and BOLD fMRI, we believe this line of work will provide critical information, e.g., more precise diagnostic measurement, and prediction of treatment or medication response for the current cocaine and nicotine addiction/relapse vulnerability studies. The output of these more powerful ASL data analysis methods will also pave the way for general application of ASL perfusion fMRI methods across multiple brain disorders, and for sustained states within the normal brain. [unreadable] [unreadable] [unreadable]