Deriving Treatment Decision Support From Dutch Electronic Health Records by Exploring the Applicability of a Precision Cohort–Based Procedure for Patients With Type 2 Diabetes Mellitus: Precision Cohort Study

TL;DR

This study explores using electronic health records to support treatment decisions for type 2 diabetes patients by analyzing similar patient data.

Contribution

The study adapts a precision cohort analytics workflow for the Dutch primary care setting using real-world EHR data.

Findings

The precision cohort analytics workflow was successfully adapted for the Dutch primary care setting.

Treatment options and outcome analyses were technically feasible but lacked statistical power to show significant outcomes.

Data from over 11,000 type 2 diabetes patients was used to generate treatment-specific filter cohorts and precision cohorts.

Abstract

The rapidly increasing availability of medical data in electronic health records (EHRs) may contribute to the concept of learning health systems, allowing for better personalized care. Type 2 diabetes mellitus was chosen as the use case in this study.

This study aims to explore the applicability of a recently developed patient similarity–based analytics approach based on EHRs as a candidate data analytical decision support tool.

A previously published precision cohort analytics workflow was adapted for the Dutch primary care setting using EHR data from the Nivel Primary Care Database. The workflow consisted of extracting patient data from the Nivel Primary Care Database to retrospectively generate decision points for treatment change, training a similarity model, generating a precision cohort of the most similar patients, and analyzing treatment options. This analysis showed the treatment options that led to a better outcome for the precision cohort in terms of clinical readouts for glycemic control.

Data from 11,490 registered patients diagnosed with type 2 diabetes mellitus were extracted from the database. Treatment-specific filter cohorts of patient groups were generated, and the effect of past treatment choices in these cohorts was assessed separately for glycated hemoglobin and fasting glucose as clinical outcome variables. Precision cohorts were generated for several individual patients from the filter cohorts. Treatment options and outcome analyses were technically well feasible but in general had a lack of statistical power to demonstrate statistical significance for treatment options with better outcomes.

The precision cohort analytics workflow was successfully adapted for the Dutch primary care setting, proving its potential for use as a learning health system component. Although the approach proved technically well feasible, data size limitations need to be overcome before application for clinical decision support becomes realistically possible.

Figures