Electronic Health Record Data May Sharpen Diabetes Screening

Miriam E Tucker

February 19, 2016

Data from electronic health records (EHR) can be used to improve detection of patients with diabetes, finds a new study that also identified possible novel diabetes risk factors.

The results were published online February 16 in the Journal of Biomedical Informatics by Ariana Anderson, PhD, a research professor and statistician at the University of California, Los Angeles.

In their analysis of nearly 10,000 EHRs, Dr Anderson and colleagues identified several new possible risk factors for type 2 diabetes, including a history of viral, chlamydial, and intestinal infections and sexual and gender-identity disorders. On the other hand, migraine disorder and cardiac dysrhythmias were associated with lower risk. Using those findings, they created a model that significantly raised the prediction for type 2 diabetes beyond that of traditional risk factors.

"We were trying to see whether you could use all this information you can get for free [in the EHR] to help diagnose diabetes," Dr Anderson told Medscape Medical News. "It was very exciting to see that even after the traditional risk factors such as [body mass index] BMI, hypertension, age, and gender, there's still a wealth of information we weren't using to screen for diabetes,"

The study was conducted prior to the US Preventive Services Task Force's October 2015 recommendation to screen overweight and obese adults aged 40 to 70 years for abnormal glucose as part of cardiovascular risk assessment, but Dr Anderson says this new algorithm extends beyond those criteria. "Even when you account for all the things we know should influence diabetes, there's still a lot out there that we could be using but aren't."

Public and private insurers are likely to be the largest adopters of EHR phenotype models, with risk scores created using existing claims databases, the authors note.

Mining the EHR

The study population included approximately 131,000 unique EHR visit entries, containing 9948 patients from 1137 unique sites spanning all 50 United States, collected between 2009 and 2012. Type 2 diabetes was listed as an ICD-9 250.X category diagnosis in 18%. Because there were a great deal of missing data — such as family history and prior medication use — the setup posed a "worst-case" scenario for prediction, Anderson and colleagues write.

Using the EHR data, the authors created three models to assess the current likelihood of type 2 diabetes: conventional model mimicking conventional risk scores including factors such as smoking status, age, sex, BMI, and hypertension status; a full "EHR model" based on the EHR phenotype using 298 features including both traditional and novel risk factors identified in the EHR but excluding diabetes-related complications and their treatments; and an "EHR DX" model using EHR model information but excluding glucose-lowering medications, since a diabetes diagnosis would change that variable.

The full EHR model and the EHR DX models both predicted type 2 diabetes better than the conventional models (P < .001). For the full EHR Model, EHR DX model, and conventional model, areas under the curve were 84.9%, 83.2% and 75.0%, respectively.

The full EHR model identified unexpected factors that were associated significantly with current type 2 diabetes. Sexual and gender-identity disorders increased the risk by approximately 130%, about the same order of magnitude as hypertension. Sexually transmitted disorders raised the risk by 82% and gastrointestinal infections by 88%, both roughly similar to the 101% rise seen with obesity.

Herpes zoster, previously shown to be associated with type 2 diabetes, raised the risk by about 90%.

On the flip side, current migraines were associated with a lower diabetes risk, akin to that of a person 29 years younger. Use of benzodiazepines and depot medroxyprogesterone contraceptives also significantly reduced the risk.

Using New Risk Factors

Since these associations were made by diagnostic code categories, it wasn't possible to unearth more specific information regarding causation, according to the researchers. For example, the use of depot medroxyprogesterone contraceptives might reflect a patient population that engages in other healthful activities that decrease the diabetes risk or a lower likelihood of prescriptions written for women with diabetes, rather than the depot medroxyprogesterone itself being protective.

Dr Anderson said the association with sexual and gender-identity disorders might be related to a variety of factors, including stress from isolation, sexually transmitted disease, drug usage, or hormones used in gender transition. "We can't really dig into that right now and figure out why it's happening, but the fact that it was there and as strong and as significant as hypertension was pretty interesting," she commented.

One theory about the reduced diabetes risk with migraines, from a previous paper that had also described the negative association, suggested that people with diabetic neuropathy may be less likely to feel the headache pain.

More research is needed to examine the nature of these associations, but in the meantime, Anderson and colleagues write, "Incorporating more medical history could increase the accuracy of existing diabetes risk scores in at-risk patient populations, for stepwise screening. The combined efficacy of EHR screening plus focused laboratory testing needs future study."

Moreover, Dr Anderson told Medscape Medical News, by using a "big-data" aggregation approach across multiple practices, the goal of improving screening accuracy can be achieved without any extra data-entering effort on the part of physicians. "As long as we can aggregate these records from enough clinicians, we can still fill in the missing holes and infer the patterns."

She added, "I think this should be encouraging to clinicians to know that even when they're just doing what they can in the time they have, which is really restricted, they're still doing a good enough job to provide enough data for us to do research on. It shows that they don't have to be perfect or absolutely thorough in order for us to make progress….They're doing a great job, even by just checking boxes. "

The study was funded by an award from the Burroughs Wellcome Fund. The authors have no further relevant financial relationships.

J Biomed Inform. 2016;60:162–168. Article


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.