Cocaine Use Prediction with Tensor-based Machine Learning on Multimodal MRI Connectome Data
Authors:
Anru R. Zhang,
Ryan P. Bell,
Chen An,
Runshi Tang,
Shana A. Hall,
Cliburn Chan,
Kareem Al-Khalil,
Christina S. Meade
Abstract:
This paper considers the use of machine learning algorithms for predicting cocaine use based on magnetic resonance imaging (MRI) connectomic data. The study utilized functional MRI (fMRI) and diffusion MRI (dMRI) data collected from 275 individuals, which was then parcellated into 246 regions of interest (ROIs) using the Brainnetome atlas. After data preprocessing, the datasets were transformed in…
▽ More
This paper considers the use of machine learning algorithms for predicting cocaine use based on magnetic resonance imaging (MRI) connectomic data. The study utilized functional MRI (fMRI) and diffusion MRI (dMRI) data collected from 275 individuals, which was then parcellated into 246 regions of interest (ROIs) using the Brainnetome atlas. After data preprocessing, the datasets were transformed into tensor form. We developed a tensor-based unsupervised machine learning algorithm to reduce the size of the data tensor from $275$ (individuals) $\times 2$ (fMRI and dMRI) $\times 246$ (ROIs) $\times 246$ (ROIs) to $275$ (individuals) $\times 2$ (fMRI and dMRI) $\times 6$ (clusters) $\times 6$ (clusters). This was achieved by applying the high-order Lloyd algorithm to group the ROI data into 6 clusters. Features were extracted from the reduced tensor and combined with demographic features (age, gender, race, and HIV status). The resulting dataset was used to train a Catboost model using subsampling and nested cross-validation techniques, which achieved a prediction accuracy of 0.857 for identifying cocaine users. The model was also compared with other models, and the feature importance of the model was presented.
Overall, this study highlights the potential for using tensor-based machine learning algorithms to predict cocaine use based on MRI connectomic data and presents a promising approach for identifying individuals at risk of substance abuse.
△ Less
Submitted 21 October, 2023;
originally announced October 2023.
