A machine learning framework for assessing gut microbiota similarity in Alzheimer's disease: A proof-of-concept application to a Jing Si herbal tea intervention cohort

Introduction Gut microbiota dysbiosis has been associated with Alzheimer's disease (AD) pathophysiology. Besides, Traditional Chinese medicine products such as Jing Si Herbal Tea (JSHT) are being explored as potential interventions targeting dysbiosis. However, no quantitative framework exists to objectively assess whether such interventions modulate gut microbiota of AD patients toward a cognitively normal-like microbial profile. This study presents a proof-of-concept machine learning approach to evaluate microbiota similarity and its application to JSHT intervention. Methods A supervised machine learning framework was trained on six publicly available AD gut microbiome datasets, which were uniformly reprocessed using QIIME2 and DADA2 pipelines. Four classifiers were evaluated under stratified nested cross-validation, selecting support vector machine (SVM) as the optimal classifier, evaluated on AUC as the performance metrics. This optimal classifier generated a calibrated Microbiota Similarity Score (MSS), which was applied to an independent 24-week JSHT intervention cohort (n = 18 AD participants; Group TA: 3 g daily; Group TB: 6 g daily; n = 9 per group) to evaluate gut microbiota compositional similarity of our AD participants to a cognitively normal reference profile. Results The SVM classifier achieved a cross-validated AUC of 0.83 ± 0.06 on external datasets, with an accuracy of 0.79, precision of 0.72, recall of 0.71, and F1-score of 0.71. Baseline MSS values in the JSHT cohort were significantly lower (TA: mean = 0.066, TB: mean = 0.064, p = 0.043). However, TB group demonstrated a more pronounced increase in MSS at week 12 and 24 (TA (WK12): mean = 0.115; TA (WK24): mean = 0.180; TB (WK12): mean = 0.334; TB (WK24): mean = 0.528; p = 0.128). The alpha diversity did not change significantly in either groups whereas beta diversity differed nominally between groups at week 24 (p = 0.035). Model explainability with SHAP identified Faecalibacterium, Akkermansia , and Tyzzerella as key features, with nominally consistent increase in abundance during intervention. Discussion This proof-of-concept study demonstrates the feasibility of applying a machine learning framework to assess microbiota similarity and evaluate compositional modulations associated with a dietary intervention in AD patients. The MSS provides a continuous, interpretable index of microbiota compositional similarity that extends beyond static case-control comparisons and enables longitudinal monitoring of microbiota configuration. Given the small intervention cohort, the observed changes associated with JSHT intake are exploratory findings that should be interpreted as proof-of-concept. Future randomized controlled trials with large datasets, integrated cognitive and metabolomic profiling are needed to validate the clinical utility of the MSS framework.