Prediction of Adherence to Basal Insulin in People with Type 2 Diabetes Initiating Telemonitoring Development of a Multivariable Machine Learning Prediction Model

Abstract

Baggrund: Type 2-diabetes udgør en betydelig global sundhedsudfordring med stigende økonomisk byrde. På trods af behandling bidrager dårlig adherence til insulinbehandling til forværring af sygdomsresultatet og øgede omkostninger ved hospitalsindlæggelse. Formålet med dette studie var at udvikle en machine learning model, til tidlig prædiktion af adhernece og identificere features. Som potentielt kan hjælpe sundhedsprofessionelle med at identificere patienter med behov for ekstra pleje. Metode: Data fra DiaMonT-studiet blev anvendt, hvor patienter insulindoser var telemonitoreret med forbundet insulin smart penne. Baseline information, data fra spørgeskema og blodprøver, fra 149 danske patienter med type 2 diabetes, blev brugt som potentielle features til machine learning modeller. Til forudsigelse af tidlig adherence blev dag 1-21 inkluderet. Sekventiel feature selektion blev anvendt med Logistisk regression klassifikationsmodel. Features scoret efter arealet under kurven med fire-folds krydsvalidering. Logistic regressions modellen blev evalueret på accuracy ved forskellige sensitivitetets niveauer. Resultater: Studiet forudsagde patienternes adherence til insulin behandling med ni identificerede features. Fire-folds krydsvalidering gav en gennemsnitlig ROC AUC score på 0,69 ± 0,03. Modellen viste et moderat præstationsniveau med den højeste accuracy værende 63.10% med en grænseværdi på 52.10% og positiv prædiktiv værdi på 62.78%. Konklusion: Det var muligt at udvikle en prædiktionsmodel for tidlig adherence og feature koefficienter. Anvendelsen af tærskelværdien vil afhænge af den relative betydning af sensitivitet og specificitet i den specifikke kliniske anvendelse.

Background: Type 2 diabetes poses a significant global health challenge, with an increasing economic burden. Despite treatment, poor adherence to insulin therapy contributes to worsening of disease outcome and increased cost in hospitalization. The aim of this study was to develop a machine learning model for early prediction of adherence to basal insulin and identify features, supporting health professionals with identifying patients in need of extra care. Method: Data from the DiaMonT trial was utilized in which patients’ insulin doses were telemonitored with a connected insulin smart pen. Baseline information, questionnaire data, and blood samples from 149 Danish patients with type 2 diabetes were used as potential features for the machine learning model. For prediction of early adherence, days 1-21 were included. Sequential feature selector was used with Logistic regression as the classifier. Features scored by area under the curve with a four-fold-cross validation. The logistic regression model was evaluated using accuracy, based on various sensitivity levels. Results: This study predicted patients’ adherence to insulin with nine identified features. Fourfold cross validation yielded a mean receiving operating characteristic area under the curve of 0.69 ± 0.003. The model showed a moderate level of performance with the highest accuracy at 63.10%, with a threshold of 52.10% and positive predictive value of 62.78%. Conclusion: It was possible to develop a prediction model for early adherence and feature coefficients. The threshold to be used in a clinical application moving forward will depend on the relative importance of sensitivity and specificity in the specific clinical application.

A master thesis from Aalborg University

Prediction of Adherence to Basal Insulin in People with Type 2 Diabetes Initiating Telemonitoring Development of a Multivariable Machine Learning Prediction Model