[Study on the portrait construction of suitable population in the treatment of insomnia with "Tongdu Tiaowei" acupoint prescription].

Abstract

Objectives: To establish a predictive model of acupuncture treatment of insomnia and to create a profile of suitable populations for acupuncture schemes, so as to help improve clinical efficacy.

Methods: The data was sourced from a prospective clinical study on acupuncture treatment of insomnia by "Tongdu Tiaowei" acupoint prescription (Baihui [GV20], Yintang [EX-HN3], bilateral Shenmai [BL62] and bilateral Zhaohai [KI6]). Data from 113 insomnia patients were included in the analysis of the present study, with the reduction rate of the Pittsburgh Sleep Quality Index (PSQI) served as the overall clinical efficacy evaluation. First, the feature selection was performed using univariate logistic regression and Boruta algorithm, and the prediction accuracy of the three boosting algorithms - adaptive boosting, gradient boosting, and extreme gradient boosting (XGBoost) - was compared for selecting the best algorithm. The grid search and ten-fold cross-validation were used to optimize the hyperparameters of the best algorithm. The optimal dataset partitioning method was selected using stratified random partitioning, and the best cut-off value was determined based on the Youden index. The predictive model for the therapeutic efficacy was constructed and its performance was evaluated. Finally, SHAP (shapley additive explanation) analysis was used to visually interpret the model.

Results: The features included in the model were the proportion of stage N1 to total sleep duration, the proportion of stage N2 to total sleep duration, R latency from lights out, stage N2 latency from lights out, the awake time after sleep onset, PSQI sleep efficiency score, and the presence of an old tongue (a tongue picture of a dry, rough texture and an old body). XGBoost was identified as the best algorithm, with the optimal probability threshold of 0.76, a corresponding precision of 0.91, a recall of 0.91, a F1 score of 0.91, an accuracy of 0.91, and an area under curve (AUC) of 0.82. Patients who meet the following conditions are more likely to respond to "Tongdu Tiaowei" acuoint stimulation:the proportion of N1 phase was about 6%-70% of the total sleep duration, N2 phase latency was less than about 40 min from the time when the lights were off, the wakefulness time was less than about 75 min or 100-300 min after falling asleep, the R phase latency was more than about 75 min from the time when the lights were off. The N2 phase were about 20%-50% of the total sleep duration, PSQI sleep efficiency score was 2 or 3, and there was no appearance of "old tongue".

Conclusions: The predictive model of the efficacy of acupuncture treatment for insomnia established using XGBoost, along with the preliminary profile of the suitable population constructed using SHAP, provides a reliable auxiliary decision-making tool for acupuncture treatment of insomnia.