QUANTIFYING TIME-VARYING MEDICATION REGIMEN COMPLEXITY USING A MACHINE-LEARNING-BASED PIPELINE IN PATIENTS WITH TYPE 2 DIABETES
Author(s)
Jun Gong, MPharm1, Antoinette Coe, PharmD, PhD1, Michael Dorsch, PharmD, MS1, Sarah Vordenberg, PharmD, MPH1, VG Vinod Vydiswaran, PhD2, Corey Lester, PharmD, PhD1;
1University of Michigan College of Pharmacy, Ann Arbor, MI, USA, 2University of Michigan Medical School, University of Michigan School of Information, Ann Arbor, MI, USA
1University of Michigan College of Pharmacy, Ann Arbor, MI, USA, 2University of Michigan Medical School, University of Michigan School of Information, Ann Arbor, MI, USA
Presentation Documents
OBJECTIVES: Patients with type 2 diabetes (T2DM) experience complex medication regimens. The Medication Regimen Complexity Index (MRCI) is a validated measure based on dosage form, dosing frequency, and additional instructions, yet it’s limited by manual, static calculation. This study employed a machine-learning-based pipeline to automate time-varying MRCI estimation and identify longitudinal trends and key drivers of medication regimen complexity in T2DM patients.
METHODS: In this retrospective cohort study, adults with T2DM treated at an academic medical center (2009-2019) were included. Electronic health records and pharmacy claims were integrated to construct longitudinal medication histories and compute index-standardized MRCI components: 1. Structured medication records mapped from National Drug Codes to RxNorm concept; 2. Unstructured dosing instructions extracted from free-text using a custom-trained named entity recognition (NER) model. MRCI was computed as the weighted sum of the dosage form, dosing frequency, and additional instructions components; updated with each prescription change. Temporal trends and drivers were assessed using linear regression and summarized using monthly population-level statistics.
RESULTS: The cohort included 7,976 patients, with 884,185 pharmacy claims and 490,081 outpatient medication orders. Average follow-up was 5.6 years. Unstructured text extraction was required for 79.8% of dosing instructions, and the NER model demonstrated high accuracy (F1 = 0.987). Mean total MRCI increased from 17.1 at diagnosis to 36.6 at 11 years, with a significant average monthly increase of 0.1 points (p < 0.001). At diagnosis, dosing frequency accounted for 49.6% of the total MRCI, followed by dosage form (40.6%), and instructions (9.8%). Over the follow-up period, both frequency and dosage form scores doubled. Diabetes-specific medications captured 19.6-21.7% of regimen complexity over the study period.
CONCLUSIONS: Medication regimen complexity increased steadily over time among patients with T2DM, reflecting cumulative treatment burden. Implementing scalable, automated MRCI can improve real-world computability of complexity and support population-level monitoring and targeted interventions in T2DM care.
METHODS: In this retrospective cohort study, adults with T2DM treated at an academic medical center (2009-2019) were included. Electronic health records and pharmacy claims were integrated to construct longitudinal medication histories and compute index-standardized MRCI components: 1. Structured medication records mapped from National Drug Codes to RxNorm concept; 2. Unstructured dosing instructions extracted from free-text using a custom-trained named entity recognition (NER) model. MRCI was computed as the weighted sum of the dosage form, dosing frequency, and additional instructions components; updated with each prescription change. Temporal trends and drivers were assessed using linear regression and summarized using monthly population-level statistics.
RESULTS: The cohort included 7,976 patients, with 884,185 pharmacy claims and 490,081 outpatient medication orders. Average follow-up was 5.6 years. Unstructured text extraction was required for 79.8% of dosing instructions, and the NER model demonstrated high accuracy (F1 = 0.987). Mean total MRCI increased from 17.1 at diagnosis to 36.6 at 11 years, with a significant average monthly increase of 0.1 points (p < 0.001). At diagnosis, dosing frequency accounted for 49.6% of the total MRCI, followed by dosage form (40.6%), and instructions (9.8%). Over the follow-up period, both frequency and dosage form scores doubled. Diabetes-specific medications captured 19.6-21.7% of regimen complexity over the study period.
CONCLUSIONS: Medication regimen complexity increased steadily over time among patients with T2DM, reflecting cumulative treatment burden. Implementing scalable, automated MRCI can improve real-world computability of complexity and support population-level monitoring and targeted interventions in T2DM care.
Conference/Value in Health Info
2026-05, ISPOR 2026, Philadelphia, PA, USA
Value in Health, Volume 29, Issue S6
Code
PT11
Topic
Methodological & Statistical Research
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics
Disease
SDC: Diabetes/Endocrine/Metabolic Disorders (including obesity)