Machine learning for classifying affective valence from fMRI: a systematic review and meta-analysis

Abstract The pleasantness or unpleasantness of psychological states, known as hedonic valence is considered a fundamental dimension of emotional experiences. Many studies have applied machine learning techniques to predict valence from fMRI data and reported varying levels of accuracy. In this work, we systematically review studies published up to October 2023 that have applied machine learning as a multi-variate pattern analysis approach to classify valence from fMRI trials of healthy adults. In each trial, a participant was presented with a stimulus expected to induce positive or negative valence. Our objectives were to (1) review and summarize selected studies based on attributes such as experimental design and task (complete list of attributes is provided in the text); (2) summarize the accuracy of valence classification; and (3) investigate how the accuracy of valence prediction is influenced by the experimental paradigm. We searched the databases Scopus, Pubmed, IEEEXplore and ACM Digital Library to retrieve relevant studies. Twenty-three studies met the eligibility criteria and were included in the review. We performed a meta-analysis involving 30 observations from 22 of those studies. The meta-analytic summary of the accuracy for classifying positive vs. negative valence was significantly above chance level. Further analysis showed that studies adopting a block-design achieve significantly higher classification accuracy than those adopting an event-related design. Based on our experiments comparing popular machine learning models across two datasets, we recommend logistic regression for its simplicity, interpretability, and comparable accuracy to more complex models. However, we suggest that future studies also explore deep learning architectures such as convolutional and graph neural networks, which have not yet been applied to classify valence from fMRI data.