Researchers from Victoria University of Wellington have used data from 10,000 patients to develop an algorithm to predict the onset of type 2 diabetes. The next step is turning what they learned into a tool healthcare professionals can use.
With a tool that can help predict the onset of type 2 diabetes, professionals can initiate interventions or modify existing treatment plans, as well as better understand the progression of the disease. This tool can drive a personalised approach to medicine that will result in improved healthcare quality.
Type 2 diabetes is responsible for considerable mortality, morbidity and strain on healthcare systems worldwide. Globally, it is estimated 425 million people have diabetes and this is predicted to increase to 629 million by the end of 2045.
One possible method of reducing this number is tailored prevention. Current treatment methods are inadequate and costly so prevention is better for the population and healthcare system.
Our study looks at creating tailored prevention for type 2 diabetes by using electronic health records. We have combined these records with a state-of-the-art machine learning algorithm to better predict the factors that lead to the onset of type 2 diabetes.
Electronic health records have the potential to store information from millions of patients across many health institutions, including demographics, medical data and clinical notes. They can be used to develop tools to help with different types of medical care and predict the onset of different diseases.
We used the wide and deep learning model developed by Google to accurately predict type 2 diabetes using anonymised health data from nearly 10,000 patients. We knew 1904 of those patients had developed type 2 diabetes and we wanted to see if our deep learning model could learn features related to developing type 2 diabetes and thus predict its onset.
We removed information that clearly indicated which patients had diabetes - for example, medication or lab tests for glucose. Once that information was removed, we tested our deep learning model to see if it could identify common characteristics among patients that led to a type 2 diabetes diagnosis. These characteristics included high body weight, a high waist circumference and being older - all risk factors for type 2 diabetes.
Our model had an accuracy rate of 84.28 percent in identifying which characteristics led to a diagnosis of type 2 diabetes - a very promising result. Despite some issues due to incomplete data, this deep learning model was able to outperform other machine learning models.
Because our model was accurate and had high sensitivity (the ability to correctly identify disease) and specificity (the ability to correctly identify the absence of disease), clinicians could use the model to more easily and effectively target more patients who are likely to develop type 2 diabetes and help prevent the onset of this disease - and they could be confident they were targeting the right people rather than attempting to target all patients.
This was the first study to apply wide and deep learning for the prediction of type 2 diabetes using electronic health records. Although there is still work to be done in refining the model and adding ability to predict important risk factors, these initial results are promising.
Now we have developed a more accurate algorithm to predict the onset of type 2 diabetes, the next step is to turn these findings into a tool to help healthcare professionals and policy-makers.
With access to a tool that can help predict the onset of type 2 diabetes, professionals can initiate interventions or modify existing treatment plans, as well as better understand the progression of the disease. This tool can drive a personalised approach to medicine that will result in improved healthcare quality.