Real-time artificial intelligence assisted insulin dosage titration system for glucose control in type 2 diabetic patients: a proof of concept study

Abstract

Objective

This study aims to develop an insulin dosage adjustment model using machine learning of high quality electronic health records (EHRs) notes and then to form an artificial intelligence-based insulin clinical decision support workflow (iNCDSS) implemented in the HIS system to give a real-time recommendation of insulin dosage titration. The efficacy and safety in clinical practice is evaluated in this proof-of-concept study.

Research design and methods

We extracted patient-specific and time-varying features from the original EHRs data and performed machine learning analysis through 5-fold cross validation. In the patient-blind, single-arm interventional study, insulin dosage was titrated according to iNCDSS in type 2 diabetic inpatients for up to 7 d or until hospital discharge. The primary end point of the trial was the difference in glycemic control as measured by mean daily blood glucose concentration during the intervention period.

Results

A total of 3275 type 2 diabetic patients with 38,406 insulin counts were included for the model analysis. The XGBoost model presented the best performance with root mean square error (RMSE) of 1.06 unit and mean absolute relative difference (MARD) of 6.0% in the training dataset, and RMSE of 1.30 unit and MARD of 6.9% in the testing dataset. Twenty-three patients with T2DM (male 14, 60.9%; age 58.8 ± 10.7 years; duration of diabetes 11.8 ± 8.8 years, HbA1c 9.1 ± 1.1%) were enrolled in the proof of concept trial. The duration of iNCDSS intervention was 7.0 ± 0.1 d. The insulin dose recommended by iNCDSS was accepted by physicians in 97.8%. The mean daily capillary blood glucose was markedly improved during the intervention period, with a reduction of mean daily capillary BG from 11.3(8.0, 13.9) mmol/L in the first 24 h to 7.9(6.5,8.9) mmol/L in the last 24 h of the trial (P < 0.001). In addition, the time range below 3.9 mmol/L was decreased from 1.1% to 0.5%.

Conclusions

The clinical decision support system of insulin dosage titration developed using a machine learning algorithm based on the EHRs data was effective and safe in glycemic control in in type 2 diabetic inpatients.

Trial registrations

ClinicalTrials.gov Identifier: NCT04053959.

Funding

National Natural Science Foundation of China,(82000822); Innovative Research Group Project of the National Natural Science Foundation of China,(31830041)