Developing Statistical Methods for Estimating Patient Phenotypes Using Electronic Health Record Data

• Download Summary

What was the project about?

Researchers can use data from electronic health records, or EHRs, in studies that compare two or more treatments. In these studies, researchers need to identify all patients with the same phenotype. Phenotypes are a person’s known traits, like height and weight, or known health problems, like diabetes. However, in EHR data, some data on patient traits or health problems may be missing for some patients.

Missing data in EHRs make it hard to correctly identify all patients with the same phenotype. It’s even harder when data are missing due to a patient’s health status. For example, patients with uncontrolled diabetes may need more lab tests than patients with controlled diabetes. As a result, researchers who are looking at lab tests may not identify patients with controlled diabetes as having diabetes.

In this project, the research team developed and tested a new statistical method that accounts for missing EHR data to estimate patient phenotypes.

What did the research team do?

The research team developed the new method. The method combined multiple types of EHR data, such as lab results, diagnostic codes, and patient age and gender to better estimate patient phenotypes. It also combined information on the patterns of missing data and the available data for each patient.

Next the research team tested the method using data created by a computer program. The data looked like real EHR data for 1,000 patients at risk for type 2 diabetes. The team created two types of missing data: data missing by chance and data missing due to a patient’s health status. Then the team looked at how well the method worked compared with four other methods that estimate patient phenotypes.

Doctors, patients, and caregivers helped design the study.

What were the results?

The new method worked better than the other methods to estimate phenotypes. It was 99 percent accurate compared with 94 percent for the next best method.

What were the limits of the project?

The project applied the method to data that met a specific set of conditions. Results may differ for data that don’t meet these conditions.

Future studies can test the methods with data to detect phenotypes that change over time, such as cancer stage.

How can people use the results?

Researchers can use the new method to identify patients with specific health problems for research, even when the EHR has missing data.

Background

In comparative effectiveness research, researchers can use data from electronic health records (EHRs) to identify patients based on their observable traits, which is known as phenotyping. Missing data in EHRs due to variations in clinical assessments across patients make it difficult to accurately assign phenotypes. It is especially challenging when data missingness is not random but instead is related to a patient’s health status. For example, patients with uncontrolled diabetes may have more HbA1c lab tests done, while patients with controlled diabetes have missing data for those tests. As a result, patients with controlled diabetes may not be identified as having diabetes in research studies using EHRs.

Existing methods for handling missing data, such as multiple imputation (MI), assume that data are missing at random. When data are not missing at random, using these methods could lead to biased results. A latent variable approach can address gaps in existing methods by combining data that vary in availability across patients. Such methods account for instances when data availability depends on health status, which may improve accuracy in estimating phenotypes from EHRs.

Objective

To develop and evaluate a latent class model for estimating phenotypes using EHR data

Study Design

Methods

The research team developed a Bayesian latent class model for predicting a patient’s phenotype. The model combined information on data availability and observed data values for each patient to estimate a latent, or unobserved, phenotype. The model assumed that the latent phenotype was correlated with model covariates, like biomarkers, clinical diagnosis codes, prescription medications, age, and gender.

The research team then simulated EHR data for 1,000 patients to resemble a sample of patients at high risk for type 2 diabetes. They introduced two patterns of missing data in the biomarkers: missing at random and missing not at random. The team evaluated the model using the simulated data. They compared the model’s performance with existing phenotype estimation methods based on (1) biomarkers only, (2) clinical codes only, (3) biomarkers and clinical codes, (4) biomarkers with missing values replaced via MI, and (5) biomarkers and clinical codes with missing biomarker values replaced via MI. For each method, the team calculated sensitivity, specificity, and the proportion of patients misclassified relative to an actual type 2 diabetes diagnosis.

Clinicians, patients, and caregivers helped design the study.

Results

The latent class model performed better than existing methods with and without MI. Sensitivity was similar to existing methods (95.9% and 91.9%, respectively) and specificity was higher (99.7% and 90.8%, respectively). Mean classification accuracy was 99% compared with 94% for the next best method that used biomarkers with MI.

Limitations

The latent class model may not work when the parametric assumptions about data distributions and relationships among covariates are not met.

Conclusions and Relevance

The latent class model can help characterize phenotypes when EHRs have missing data associated with patient health status.

Future Research Needs

Future research can apply the methods to phenotypes that vary over time, such as cancer stage.