[Development and research of an AI-assisted decision-making platform in treatment of insomnia with acupuncture of acupoint prescription].

OBJECTIVE

To construct and validate a predictive model for the therapeutic effect of acupuncture at prescription (acupoint prescription for promoting the circulation of the governor vessel and nourishing the heart) on insomnia, so as to develop an open-access interactive artificial intelligence (AI)-assisted decision-making platform.

METHODS

Clinical data of 139 insomnia patients treated with acupuncture therapy were included. All the patients had received acupuncture at Baihui (GV20), Yintang (GV24), bilateral Shenmen (HT7), and bilateral Sanyinjiao (SP6); and electric stimulation was attached to Baihui (GV20) and Yintang (GV24), using a continuous wave and a frequency of 2 Hz. The treatment was delivered once every other day, 3 treatments a week, and for 2 consecutive weeks. Patients with Pittsburgh sleep quality index (PSQI) score reduction rate <50% were classified as the "no response group", and those with ≥50% were as the "response group". Outliers were addressed using the 1.5×IQR rule, and missing values were imputed via predictive mean matching. Key features were selected by intersecting the feature importance results from eXtreme Gradient Boosting (XGBoost) and random forest algorithms. After balancing class distribution using the Synthetic Minority Over-sampling Technique (SMOTE), 20% of the data was reserved as a validation set. The remained data underwent the stratified sampling iterations to generate 200 pairs of 3∶1 training-test sets, which was employed for training and internal validation of 8 machine learning algorithms. The optimal algorithm and data partitioning strategy were selected to construct the final model, followed by external validation. The best-performing model was deployed online via Streamlit to create an interactive AI platform.

RESULTS

Key predictive features for model construction included insomnia duration, the total PSQI score, PSQI sleep efficiency subscore, the proportion of N1 and N2 sleep stages in total sleep duration, and the maximum pulse rate during sleep. The CatBoost-based model achieved an AUC of 0.92, the average precision of 0.77, and accuracy, average recall, and average F1-score of 0.75 on the test set. On the validation set, it attained an AUC of 0.84, with accuracy, average precision, average recall, and average F1-score all at 0.72, demonstrating robust predictive performance. An interactive AI platform was subsequently developed (https://tdyx-catboost.streamlit.app/).

CONCLUSION

This study successfully establishes and validates a CatBoost-based efficacy prediction model for acupuncture therapy in treatment of insomnia. The developed AI platform provides data-driven decision support for acupuncture-based insomnia management.