World Journal of Traditional Chinese Medicine · Published 2026-02-23 · DOI 10.4103/wjtcm.wjtcm_74_25
Objective: This study aimed to develop a machine learning model for the diagnosis of chronic renal failure (CRF) with spleen–kidney qi deficiency (SKQD) syndrome by investigating facial chromatic biomarkers based on traditional Chinese medicine (TCM) theory. Materials and Methods: This cross-sectional study analyzed TCM-defined facial regions in 303 patients with CRF (166 SKQD and 137 non-SKQD) and 25 healthy controls using active appearance models and RGB/Lab color spaces. Key chromatic features were selected using Chi-square automatic interaction detection (CHAID) decision trees to construct a multibranch deep convolutional neural network (CNN), nomogram, and hybrid model. The model performances were compared using the DeLong test. Results: The SKQD group showed a higher glomerular filtration rate, albumin level, and facial R-value with a lower serum creatinine level (all P < 0.05), accompanied by a higher prevalence of shortness of breath with reluctance to speak, and cold extremities (P < 0.05) compared with the non-SKQD group. Chromatic analysis revealed elevated R/a values in the kidney region using the natural facial color region method (P < 0.05), and distinctively higher R-values in the bladder-uterus region and a-values in the right small intestine region using the Mingtang facial color region method (P < 0.05). The CHAID algorithm selected four diagnostic determinants, including the overall facial R-values. The CNN achieved superior discriminative performance (test area under the receiver operating characteristic curve = 0.74 vs. nomogram 0.68, DeLong test P < 0.001) with no added benefit from model integration (P > 0.05). Conclusions: Facial chromatic biomarkers effectively differentiated SKQD in patients with CRF when analyzed using TCM-defined regions. The superior performance of the CNN validates AI-driven TCM diagnostics as a clinically actionable tool for precise syndrome classification.
Abstract from DOAJ. Public domain (CC0 1.0).
Read the article at the publisher →