QMVO-SCDL: A new regression model for fMRI pain decoding using quantum-behaved sparse dictionary learning

The exponential growth of functional magnetic reasoning imaging (fMRI) data offers a great opportu-nity for basic and clinical research to explore functional brain activity. Nonetheless, due to the lack of effective and efficient tools for handling high-dimensional and complex fMRI data, this potential has still not been fully explored. A critical issue to be addressed is to identify clinically relevant features from high dimensional fMRI data in a fast and accurate manner. To address this problem, a new quantum-behaved multiverse optimization (QMVO) approach is proposed for fMRI dimensionality reduction and a new prediction approach based on L1-norm sparse coding and adaptive dictionary learning (SCDL) is developed to decode the whole brain fMRI data. QMVO is designed to enhance local search ability, avoid premature convergence, and increase population diversity among individuals. Further, to efficiently re-estimate parameters of the prediction model with new available data, a new SCDL sparse coding is proposed, which requires no training and needs minimal parameters to tune. The proposed QMVO-SCDL model is applied to a pain-evoked fMRI dataset to decode the pain perception level. Results show that QMVO-SCDL can decode pain levels and identify predictive fMRI patterns with high accuracy, high convergence speed, and short consumption time. Moreover, the performance of the proposed model outperforms different recent ML techniques. Therefore, the proposed model has a great promise to be a powerful tool for fMRI decoding. (C) 2022 Elsevier B.V. All rights reserved.

Automated summary

This study proposes a new quantum-behaved multiverse optimization (QMVO) approach and a prediction method based on sparse coding and dictionary learning (SCDL) for handling high-dimensional and complex fMRI data. The proposed model, applied to pain perception fMRI data, demonstrates high accuracy in decoding pain levels and identifying predictive fMRI patterns. Moreover, the performance of the model surpasses that of other machine learning techniques.