Study on multi-modal feature fusion machine learning classification model of Chinese and Western medicine for depressive symptoms in middle-aged and older patients with chronic low back pain
TL;DR
Both GBM and RF models incorporating multimodal features from traditional Chinese and Western medicine demonstrated promising preliminary classification performance for screening depressive symptoms in middle-aged and older patients with CLBP, suggesting their potential utility as screening tools in clinical settings.
Abstract
Objective To construct a machine learning classification model integrating multi-modal features of Chinese and Western medicine for screening the risk of depressive symptoms in middle-aged and older patients with chronic low back pain (CLBP). Methods A cross-sectional design included 370 CLBP patients aged ≥45 years admitted to our hospital from January 2022 to June 2024. Based on PHQ-9 scores, they were classified into depression (≥10) and non-depression (<10) groups. Chinese and Western medicine characteristics were collected and randomly split into training and validation sets (7:3 ratio). In the training set, Least Absolute Shrinkage and Selection Operator(LASSO) and multivariate logistic regression identified risk factors for depressive symptoms. Gradient boosting machine (GBM) and random forest (RF) models were then built using these factors. Model performance was assessed via AUC, sensitivity, specificity, and calibration curves, with DeLong tests comparing models. Additionally, a temporal validation cohort of 156 patients admitted from July 2024 to June 2025 was used to evaluate model stability over time. Results The prevalence of depressive symptoms in the training set was 28.20%. Multivariate analysis identified VAS score, ODI score, PSQI score, qi deficiency constitution, and blood stasis constitution as independent risk factors for depressive symptoms, while sleep duration served as a protective factor (all P<0.05). The GBM model achieved AUCs of 0.927 and 0.837 in the training and validation sets, respectively. The RF model achieved AUCs of 0.998 and 0.811 in the training and validation sets, respectively. Both models demonstrated good calibration in both sets. The DeLong test indicated no statistically significant difference in performance between the two models in the training or validation sets (P>0.05). In temporal validation, the AUCs for the GBM and RF models were 0.956 and 0.952, respectively, with no statistically significant difference between the two models (P = 0.751). Conclusion Both GBM and RF models incorporating multimodal features from traditional Chinese and Western medicine demonstrated promising preliminary classification performance for screening depressive symptoms in middle-aged and older patients with CLBP, suggesting their potential utility as screening tools in clinical settings. Further multicenter external validation is needed to confirm the generalizability of these findings.