A New Method to Create Personalized Treatment Guidance Using Patient Health Data

• Download Summary

What was the project about?

Researchers can use data on patient traits such as age, health problems, and treatment preferences, to create personalized treatment rules, or PTRs. PTRs provide doctors with guidance on how to treat patients’ health problems based on their traits. But PTRs based on a single data source may not apply to all patients. For example, if researchers create a PTR using data from older people with heart failure, it may not apply to younger people with heart failure.

To avoid this problem, researchers can create PTRs by combining data from many sources. PTRs based on many data sources can help guide treatment for patients with different traits.

In this study, the research team created and tested a new method for creating PTRs using data from multiple sources.

What did the research team do?

First, the research team developed the new method for creating a PTR that applies to many different patients. The method uses data from a source data set and a target data set.

Next, the research team tested the new method. They used health record data from 6,361 patients who received intensive care for sepsis, or wide-spread infection in the body. The team created source and target data sets. Using the source data set, they applied the new method to create a PTR for sepsis treatment. Then they used a computer program to check how well the PTR worked with the target data set under six scenarios. For example, they checked how the method worked when the patient traits between the two data sets were similar or different. They also compared the new method with three existing methods.

A patient, a caregiver, a doctor, a nurse, and an expert in data analysis helped design the study.

What were the results?

Overall, the new method performed better than other methods. The method also worked better than other methods when patient traits in the source and target data sets differed a lot.

What were the limits of the project?

If data on certain patient traits that affect treatment are not in the data sets, then the new method will not create an accurate PTR. If the data reflects bias in how patients of different races and ethnicities receive care, then PTRs may not improve health for all patients.

Future research could look at how well the method works for creating PTRs for patients of different races and ethnicities.

How can people use the results?

Researchers can use the new method to create PTRs using patient data from many sources. These PTRs could help doctors create treatment plans that can work for different patients.

Background

Personalized treatment rules (PTRs) are a set of criteria that can help doctors tailor treatments based on patients’ characteristics, health conditions, and preferences. Researchers often develop PTRs using a single data source, such as electronic health records (EHRs) for patients at one healthcare system. Using multiple data sources may improve PTRs and their generalizability to other patient populations. But variation in the distribution of patient characteristics across data sources, or covariate shift, creates methodological challenges for developing robust and generalizable PTRs.

Objective

To develop statistical methods for creating generalizable PTRs from multiple data sources

Study Design

Methods

The research team developed a weighting method that could use data from a source data set and a target data set to address challenges arising from covariate shifts and confounding. The method is applicable when the source data set includes patient information on covariates, treatment, and health outcomes, but the target data set only includes individual-level data on patient covariates, such as age and comorbid conditions. The new method applied a statistical approach, called sampling weights, to make the covariate distribution in the source data set resemble the covariate distribution in the target data set.

To see if the method could identify which patients would benefit from a specific treatment, the research team tested the method using EHRs from 6,361 patients receiving intensive care for sepsis. They divided the data into source and target data sets. The team first used computer simulations to create six scenarios that mimicked different covariate shifts and confounding. Under each scenario, the team applied the new method to the source data set to develop a generalizable PTR for transthoracic echocardiography as a treatment option for patients with sepsis. Next, the team assessed how generalizable the PTR was for the target data set. They estimated a statistic, called the modified value function, to determine the 28-day mortality rate of the target population under the PTR, compared to the rate based on another PTR that offered a different treatment. The team used the modified value function to compare the PTR developed under the new method with PTRs developed using three other weighting methods.

A patient, a caregiver, a clinician, a nurse, and a director of health system analytics helped design the study.

Results

The new method produced the best modified value, indicating the lowest mortality rate, in four out of six scenarios and performed close to the best in the other two. The new method performed better than other methods if covariate shift or extra confounding was present in the source data set.

Limitations

If the data reflect bias in how patients from certain races and ethnicities receive care, the PTR may worsen disparities in health outcomes.

Conclusions and Relevance

Researchers can apply the new weighting method to develop generalizable PTRs for different treatments.

Future Research Needs

Future research could assess whether the new method works for creating generalizable PTRs for patients of different races and ethnicities.