Researchers develop a machine learning model that accurately predicts diabetes using routinely collected and linked health data

A newly developed model can accurately predict type 2 diabetes using routinely collected health data, according to new research from ICES, a non-profit research institute that uses population-based health information to produce knowledge on a broad range of healthcare issues.

“Effective preventive care and specific interventions exist to prevent type 2 diabetes; but there is often a challenge ensuring those approaches are targeted to those who need it most in the context of a health system,” says Dr. Laura Rosella, lead author of the study, and scientist at ICES.

The study published today in the JAMA Network Open tests new machine learning-based approaches on routinely collected data that examines the entire population of Ontario. This is one of the largest uses of these machine learning approaches on routinely collected health data for diabetes prediction and demonstrates the potential of these methods on similar health data collected around the world.

“Prediction models for type 2 diabetes are specifically useful for informing more effective and efficient targeting of health system interventions that support the prevention of type 2 diabetes,” adds Rosella.

The researchers used linked administrative health data from Ontario from 2006 to 2016. The model trained on data from nearly 1.7 million patients and validated that data on more than 240,000 patients and then tested the data on 236,000 patients to develop a validated algorithm. The approach was able to accurately predict the incidence of diabetes in the population using the routinely collected data.

“The model was tested for accuracy for predicting diabetes onset with 5 years overall and among important demographic and socioeconomic sub-groups to ensure it works well for the full population,” says Rosella.

The model showed that while the number of patients with diabetes was estimated to be 785,000 with an associated cost of $3.5B in 2009, these figures rose to 1,144,000 and $5.4B respectively, only seven years later. The cohort with diabetes grows at an average of 51,800 new patients per year between 2009 and 2016, adding on average $242M per year to the financial burden of diabetes. The model also showed that patients predicted at the highest risk account for the largest portion of healthcare costs: moderate-risk and high-risk patients are five per cent of the population but represent 26 per cent of the total diabetes cost.

While the researchers add that the goal of the model is not to be applied in the context of individual patient care, it’s main purpose if to inform population-health planning and management for prediction of diabetes that incorporates health equity and can be used to guide health system planning.

The study “Development and validation of a machine learning model using administrative health data to predict onset of type-2 diabetes,” was published in JAMA Network Open.

Author block: Mathieu Ravaut, Vinyas Harish, Hamed Sadeghi, Kin Kwan Leung, Maksims Volkovs, Kathy Kornas, Tristan Watson, Tomi Poutanen and Laura C. Rosella.

FOR FURTHER INFORMATION PLEASE CONTACT:

Deborah Creatura
Interim Director Communications, ICES
[email protected]
(c) 647-406-5996