Artificial intelligence based predictive tools for identifying type 2 diabetes patients at high risk of treatment Non-adherence: A systematic review

Abstract

Aims

Several Artificial Intelligence (AI) based predictive tools have been developed to predict non-adherence among patients with type 2 diabetes (T2D). Hence, this study aimed to describe and evaluate the methodological quality of AI based predictive tools for identifying T2D patients at high risk of treatment non-adherence.

Methods

A systematic search was conducted across multiple databases including, EMBASE, Cochrane Library, MedLine, and Google Scholar search. The Prediction model Risk Of Bias ASsessment Tool (PROBAST) was used to assess the quality of studies. The performances of tools were assessed by Area Under the Curve (AUC), precision, recall, C-index, accuracy, sensitivity, specificity or F1 score.

Results

Most studies measured predictive ability using AUC (75 %), and some only reported precision (25 %), recall (12.5 %), C-index (12.5 %), accuracy (37.5), sensitivity (12.5 %), specificity (12.5 %) or F1 score (25 %). All tools had moderate to high predictive ability (AUC > 0.70). However, only one study conducted external validation. Demographic characteristics, HbA1c, glucose monitoring data, and treatment details were typical factors used in developing tools.

Conclusions

The existing AI based tools holds significant promise for improving diabetes care. However, future studies should focus on refining the existing tools, validating in other settings, and evaluating the cost-effectiveness of AI-supported interventions.