Home // HEALTHINFO 2023, The Eighth International Conference on Informatics and Assistive Technologies for Health-Care, Medical Support and Wellbeing // View article

Medication Adherence Prediction for Homecare Patients, Using Medication Delivery Data

Authors:
Ben Malin
Tatiana Kalganova
Ejike Nwokoro
Joshua Hinton

Keywords: medication adherence; CNN; RF, healthcare; homecare; adherence prediction

Abstract:
This study aims to predict the risk of medication nonadherence for patients who are newly enrolled into a medication delivery homecare service – an insight that can underpin the design of more impactful patient support programs for patients with long term conditions. In the context of this study, we have defined a nonadherent patient as someone without any prescribed medication available across the month. This is calculated using medication delivery confirmation and prescription data. Convolutional Neural Networks (CNN) and Random Forest (RF) networks are used for this study, with the former shown to be our best-performing model, achieving an 82.8% Area Under the Curve (AUC) on a subset of the patient population who have been on service for 3 to 4 months. When testing the model on the entire patient population (regardless of how long they have been on service), and by using crossvalidation, the AUC improves to 97.4%. The methodology that is applied in our study is novel based on three distinct factors: (1) prediction that is based on a novel visualization of 12 months of patient medication delivery data, (2) taking into consideration the temporal patient communications as well as the possibility of patient stockpiling of prescribed medication and (3) the service level i.e., level of nurse support received by the patient. We find that the inclusion of temporal patient communication data into our analysis improves both the AUC and the nonadherence prediction precision in the CNN model (0.7% and 19.4% respectively); a similar improvement in AUC and prediction precision is not seen in the RF model. The CNN model is therefore identified as the appropriate model for our use case. Furthermore, our results support the claim that temporal communication data are relevant datapoints for predicting adherence in a network that is better-suited to timeseries data.

Pages: 30 to 38

Copyright: Copyright (c) IARIA, 2023

Publication date: November 13, 2023

Published in: conference

ISSN: 2519-8491

ISBN: 978-1-68558-105-3

Location: Valencia, Spain

Dates: from November 13, 2023 to November 17, 2023