Interpretable Machine Learning Model Based on Radiomic Features of Hippocampal Subfields for Classifying Zoster-associated Pain.

Postherpetic neuralgia (PHN), the most common complication of herpes zoster (HZ), is difficult to prevent due to limited understanding of the mechanisms underlying the transition from acute to chronic pain and the lack of objective neuroimaging biomarkers for individualized risk stratification. In this study, we aimed to develop interpretable machine learning models based on hippocampal subfield radiomics to classify zoster-associated pain (ZAP) patients and identify neuroimaging biomarkers reflecting microstructural alterations in hippocampus associated with ZAP. High-resolution 3D T1-weighted MR images were obtained from 28 PHN patients, 25 HZ patients, and 21 healthy controls (HCs). We extracted radiomic features of ten hippocampal subfields and used five machine learning classification algorithms to identify the most contributive subfield and robust model for classifying ZAP patients. Furthermore, SHapley Additive exPlanations (SHAP) was used to interpret the optimal model. In distinguishing ZAP patients from HCs, the SVM_Linear model based on the left subiculum radiomics achieved the highest performance with an AUC of 0.740, whereas the Logistic_L1 model based on the right CA4-DG and left SR-SL-SM radiomics showed the best performance with an AUC of 0.799 in distinguishing PHN patients from HZ patients. SHAP analysis revealed that SmallAreaEmphasis from the right CA4-DG subfield was the dominant feature for identifying PHN patients, followed by ClusterShade from the left SR-SL-SM subfield. Furthermore, both features showed strong associations with disease duration. These results suggest that hippocampal subfield radiomics can effectively stratify ZAP phases. Specifically, right CA4-DG radiomic features may serve as effective neuroimaging biomarkers associated with disease chronicity and have the potential for early risk stratification and intervention of ZAP.