Development of a Machine Learning Model to Identify Nonadherence in People With Type 2 Diabetes using Connected Insulin Pen Data
Authors
Nielsen, Mads Weiss ; Højer, Line Aas ; Leensbak, Maja Randa
Term
4. term
Publication year
2024
Submitted on
2024-05-31
Pages
59
Abstract
Type 2 diabetes is widespread, and poor medication adherence worsens glycemic control and increases costs. Early identification of nonadherence can enable targeted interventions and is considered more beneficial than developing new treatments. This study aimed to develop a machine learning model to identify insulin nonadherence in people with type 2 diabetes using objective data from a connected insulin pen. Data from 331 individuals were used to build eight supervised models; features were selected via sequential forward feature selection, models were trained with 5-fold cross-validation, evaluated with ROC-AUC, and optimized using grid search, with only the top-performing model reported. In total, 43 features were considered during feature selection. A Random Forest performed best (ROC-AUC: 0.749), with most important predictors including time in range from continuous glucose monitoring, HbA1c, telemonitoring status, systolic blood pressure, overall health status, and insulin regimen type (basal plus bolus vs basal). The study provides a model capable of identifying insulin nonadherence in people with type 2 diabetes, and the findings suggest that telemonitored patients are more likely to be adherent.
Type 2 diabetes er udbredt, og manglende adhærens til medicin forværrer blodsukkerkontrol og øger omkostningerne. Tidlig identifikation af nonadhærens kan muliggøre målrettede tiltag og vurderes mere fordelagtigt end at udvikle nye behandlinger. Dette studie havde til formål at udvikle en maskinlæringsmodel, der kan identificere insulin-nonadhærens hos personer med type 2 diabetes ved at udnytte objektive data fra en sammenkoblet insulinpen. Data fra 331 personer indgik i udviklingen af otte superviserede modeller, hvor features blev udvalgt med sequential forward feature selection, og modellerne blev trænet med 5-fold krydsvalidering og evalueret med ROC-AUC. Alle modeller blev optimeret med grid search; resultaterne for den bedst præsterende model præsenteres. I alt blev 43 features udtrukket til feature selection. Random Forest præsterede bedst (ROC-AUC: 0,749), og de vigtigste features omfattede tid i målområdet for kontinuerlig glukosemonitorering, HbA1c, om patienten var telemonitoreret, systolisk blodtryk, generel helbredstilstand samt insulintype (basal+bolus eller basal). Studiet præsenterer dermed en model, der kan identificere nonadhærens til insulinbehandling hos personer med type 2 diabetes, og resultaterne indikerer, at telemonitorerede patienter er mere tilbøjelige til at være adhærente.
[This apstract has been generated with the help of AI directly from the project full text]