How Machine Learning Can Help Prevent Type 2 Diabetes (And Its Complications).

Diabetes is one of the most potentially devastating ailments in existence, yet one of the most common: According to the American Diabetes Association (ADA), nearly 40 million Americans (or more than 10 percent of the population) had diabetes in 2019. According to the ADA, it’s a number that rises 1.4 million every year.

Things aren’t much better from a global perspective, with 463 million diabetics worldwide in 2019 (or 9.3 percent of the global population). This proportion will to grow to 11 percent (or 700 million people) by 2045.

Diabetes also accounts for 12% of the world’s healthcare expenditure across research, analysis, and treatments.

Even more striking than the above numbers are the more than 90 million Americans in 2019 with an early (and completely curable) form of the disease called “prediabetes”. Indeed, one in two people living with the disease around the world doesn’t even know they have diabetes. 

To ensure positive healthcare outcomes and keep costs as low as possible for healthcare systems and payers, it’s imperative to intervene as early as possible to keep at-risk patients from developing the disease. 

What is diabetes?

A lack of insulin ultimately causes diabetes. According to the ADA, most cases of type 1 diabetes are triggered by an autoimmune disorder that attacks insulin-creating beta cells and renders the pancreas incapable of producing insulin. 

The ADA says type 2 diabetes – which afflicts between 90 to 95 percent of people with diabetes – is caused mainly by lifestyle factors such as being overweight or obese; having diets high in fats, sugars, and carbohydrates; and a lack of physical inactivity. 

In both types of diabetes, genetic factors can also play an outsized role. The likelihood of developing either type increases significantly if a first-degree relative also has the condition.

Traditional methods of diabetes prevention typically center around proactive education on the risks associated with obesity, smoking, a sedentary lifestyle, and other factors. Even among those with the disease, such lifestyle choices can cause further complications such as retinopathy, nephropathy, and neuropathy.

Despite these efforts, however, the proportion of diabetics continues to rise. And among those with the disease, studies show that a large proportion (31%) of diabetics stop using their medication within three months (and nearly 60% after one year).

Indeed, the Achilles heel of diabetes prevention and treatment has always been the absence of real-time, key health data – and the ability of the patient to make informed, data-driven health decisions on their own. 

But that’s changing with the rapid development of artificial intelligence (AI) and machine learning (ML) tools for diabetes prevention and treatment. 

How AI and machine learning models can help prevent diabetes

The healthcare industry generates a ton of data – around 30 percent of the entire world’s volume of newly created data. According to RBC Capital Markets, healthcare data’s compound annual growth rate will hit 36 percent (a 10 percent faster growth rate than the financial services industry) by 2025.

That’s a heck of a lot of data – more than researchers and clinicians can ever hope to go through manually. Adding to the challenge is that many healthcare data sources, such as doctor’s notes and internet of things (IoT) data, are unstructured.

But AI and ML models alongside other technologies, such as computer vision for scanning images and NLP for reading free text, can comprehend vast amounts of unstructured data. This can help detect patterns, provide personalized healthcare recommendations, help with early detection, and even predict which currently healthy people might develop diabetes.

When combined with wearable technology, social media, and other online technologies, ML models can help empower patients to either self-manage diabetes more effectively or make better lifestyle choices to prevent developing the disease in the first place.

Indeed, several AI- or ML-based medical devices to aid clinical diagnosis are already on the market in the U.S. Although some recent studies concede that ML models currently can’t predict diabetes any better than conventional models, they also say the explosion in healthcare data and improved computing power will improve the accuracy of ML prediction models.

Machine learning models for diabetes prevention

One such study was on display at the University of Toronto in mid-2021, when researchers analyzed the health data of 2.1 million people using ML models to accurately predict those most likely to develop type 2 diabetes within the next five years. 

“We know that identifying people who are at risk of developing type 2 diabetes is really important because there are things we can do to prevent the onset of the disease,” said one of the researchers, U of T associate professor Laura Rosella. “There is a demonstrated advantage to intervening early when people are at risk of type 2 diabetes.”

The study, which demonstrated an accuracy rate of around 80 percent, was published in JAMA Network Open.

Other examples have used primary care physician records to analyze body mass index, genetic history, and hemoglobin A1C diagnostic test results to potentially warn patients of risk and devise early intervention strategies.

Still other technologies to help prevent or manage diabetes include virtual counseling powered by conversational AI models – essentially personal health coaches that are always available. 

Types of ML models used in preventative diabetic medicine

Many predictive models exist around type 2 diabetes, with the earliest attempts based on classic statistical learning techniques such as linear regression. 

However, several new techniques now allow the prediction of new diabetes cases based on pattern recognition in training data. Models are often combined to create ensemble models for even better performance. 

According to recent systematic reviews, some of the top-performing diabetes detection or prediction models include decision tree, random forests, and swarm optimization. The most frequently used models include deep neural networks (particularly well-suited for big data applications), gradient boosting, random forest, and support vector machines.  

Some recent studies have combined ML models with additional techniques – including resampling and feature selection – to improve performance further.

CapeStart: The healthcare machine learning experts 

CapeStart works with leading healthcare organizations worldwide to improve the accuracy and efficiency of their research through AI, machine learning, natural language processing, and other force-multiplying technologies. Contact us today to set up a quick discovery call with one of our machine learning engineers to learn how our custom and pre-built machine learning models, ML training datasets, and AI expertise can help your organization.