Interpretable time-series neural turing machine for prognostic prediction of patients with type 2 diabetes in physician-pharmacist collaborative clinics

Jie Xiao; Bin Chen; Lei Chen; Qing Wang; Shenglan Tan; Haiyan Yuan; Daxiong Xiang; Bikui Zhang; Xia Li; Shuting Huang; Yuhan Tan; Yining Cheng; Wenzheng Xie; Ping Xu

doi:10.1016/j.ijmedinf.2024.105737

Interpretable time-series neural turing machine for prognostic prediction of patients with type 2 diabetes in physician-pharmacist collaborative clinics

Int J Med Inform. 2024 Nov 29:195:105737. doi: 10.1016/j.ijmedinf.2024.105737. Online ahead of print.

Authors

Jie Xiao¹, Bin Chen², Lei Chen¹, Qing Wang¹, Shenglan Tan¹, Haiyan Yuan¹, Daxiong Xiang¹, Bikui Zhang¹, Xia Li³, Shuting Huang¹, Yuhan Tan¹, Yining Cheng¹, Wenzheng Xie¹, Ping Xu⁴

Affiliations

¹ Department of Pharmacy, The Second Xiangya Hospital, Central South University, Changsha, China; Institute of Clinical Pharmacy, The Second Xiangya Hospital, Central South University, Changsha, China.
² KylinSoft Corporation Ltd, Changsha, China; School of Automatic, Central South University, Changsha, China.
³ Department of Endocrine, The Second Xiangya Hospital, Central South University, Changsha, China.
⁴ Department of Pharmacy, The Second Xiangya Hospital, Central South University, Changsha, China; Institute of Clinical Pharmacy, The Second Xiangya Hospital, Central South University, Changsha, China. Electronic address: xuping1109@csu.edu.cn.

PMID: 39626597
DOI: 10.1016/j.ijmedinf.2024.105737

Abstract

Background: Type 2 diabetes (T2D) has become a serious health threat globally. However, the existing approaches for diabetes prediction mainly had difficulty in addressing multiple time-series features. This study aims to provide an adjunctive tool for the clinical identification of patients in physician-pharmacist collaborative clinics at high risk of poor prognosis.

Methods: This study proposes a novel interpretable time-series Neural Turing Machine (ITS-NTM) to form patient characteristics into feature matrixes to simulate one's disease and treatment process, predicting the prognosis of patients with T2D and alerting early interventions. Model robustness was verified by 10-fold cross-validation, external validation and multi-model comparisons. We also conducted dynamic prediction and feature importance analysis to explore its interpretability.

Results: The study population included patients with T2D attending physician-pharmacist collaborative clinics over 12 months in primary healthcare centers, while clinical features and behavioral indicators at baseline, 3rd, 6th, 9th and 12th months were used to reflect the fluctuation of disease control over time. Compared with five state-of-the-art prediction models, the ITS-NTM obtains 92.0 % in accuracy and 91.8 % F1-score, demonstrating the superiority performance. Feature importance demonstrated that the top 5 features were glycosylated hemoglobin, fasting blood glucose, medication adherence scores, 2-hour postprandial blood glucose and waist-to-hip ratio, which had the greatest impact on the performance of the predictive model.

Conclusions: Proposed ITS-NTM could be used to promote the implementation of physician-pharmacist collaborative clinics, and further prompt the application of artificial intelligence to optimize the allocation of medical resources and improve the quality of care in under-resourced areas.

Keywords: Deep learning; Neural Turing Machines; Physician-pharmacist collaborative clinics; Precision medicine; Type 2 diabetes.