Decoding pain: Uncovering the factors that affect the performance of neuroimaging-based pain models

Neuroimaging-based pain biomarkers, when combined with machine learning techniques, have demonstrated potential in decoding pain intensity and diagnosing clinical pain conditions. However, a systematic evaluation of how different modeling options affect model performance remains unexplored. This study presents the results from a comprehensive literature survey and benchmark analysis. We conducted a survey of 57 previously published articles that included neuroimaging-based predictive modeling of pain, comparing classification and prediction performance based on the following modeling variables - the levels of data, spatial scales, idiographic vs population models, and sample sizes. The findings revealed a preference for population-level modeling with brain-wide features, aligning with the goal of clinical translation of neuroimaging biomarkers. However, a systematic evaluation of the influence of different modeling options was hindered by a limited number of independent test results. This prompted us to conduct benchmark analyses using a locally collected functional magnetic resonance imaging dataset (N = 124) involving an experimental thermal pain task. The results demonstrated that data levels, spatial scales, and sample sizes significantly impact model performance. Specifically, incorporating more pain-related brain regions, increasing sample sizes, and averaging less data during training and more data during testing improved performance. These findings offer useful guidance for developing neuroimaging-based biomarkers, underscoring the importance of strategic selection of modeling approaches to build better-performing neuroimaging pain biomarkers. However, the generalizability of these findings to clinical pain requires further investigation.