Healthcare

AI Offers Hope for Earlier Screening for Type 1 Diabetes

Share this post:

Some 1.25 million Americans suffer from type 1 diabetes (T1D), an autoimmune disease characterized by antibodies in the blood against insulin-producing cells made by the pancreas1. The exact cause of the disease is unknown, there’s currently no cure, and scientists are hard-pressed to explain why some people develop T1D while others don’t.

Working with JDRF, a leading nonprofit devoted to T1D research, we believe that better understanding disease etiology may hold the potential for new interventions to better manage increasing rates of T1D, which account for $16 billion in healthcare expenditures and lost income annually2. By 2050, the number of people in the U.S. with T1D is expected to surge to 5 million, including nearly 600,000 people under the age of 203.

Although the reasons for the rise in T1D is still a mystery, there is hope in being able to better identify children at-risk for the disease earlier to help mitigate the impact.

Using artificial intelligence (AI), our team of researchers built models that identified patterns of specific antibodies in the blood to progression timelines of how quickly T1D would develop in a particular person – which can vary greatly from individual to individual. Our machine learning algorithms analyzed massive volumes of anonymized data, collected with consent over time from people at-risk of developing T1D, to identify specific progression patterns correlated among biomarkers, genetic risk factors, symptoms and other clinical data.

Key Findings

The scale of the combined data set – including clinical, anonymized data from more than 22,000 individuals collected from four different research sites across three countries – and the availability of advanced analytic techniques enabled us to dig deeper than previous studies around T1D antibodies.

This research is the first time AI has analyzed such a large and diverse body of data about T1D antibodies and the clues these markers give about not only an individual’s risk of developing diabetes, but how quickly it will develop. Screening for T1D antibodies is currently not widely accepted or standard in clinical practice or testing, such as newborn screenings, yet research has increasingly shown its value in predicting the onset of the disease4.

Our study marks the first in-depth, large scale study of time courses when different antibodies appear, and their correlations to specific progression patterns of T1D. Specifically, the research team found that modeling the dynamics of four autoantibodies (IAA, IA-2A, GADA, ZnT8A) improved the accuracy of T1D onset time prediction. In particular, the team found that autoantibody development interacts with the HLA-DR-DQ genotype, as well as interactions between at least one DR3-DQ2.5 haplotype and autoantibodies GADA and ZnT8A. Additionally, connections were found between the IA-2A autoantibody and seroconversion age, or the time period when an antibody develops and becomes detectable in the blood. The AI model showed that these interactions seem to accelerate the progression of onset to T1D.

Ultimately, this lays the groundwork for the potential design and development of earlier screening methods in children, so clinicians can not only determine their risk for T1D, but also how quickly or slowly they are likely to develop the disease, and better design personalized follow-up when individuals are still in pre-symptomatic phases.

As T1D onset typically appears between the ages of 4 and 14 years old, predicting when and how quickly it will show in an individual, could offer clinicians the information they need to monitor children as appropriate, with the goal of helping to reduce complications and potentially enabling individuals to enroll in life-changing trials for preventative therapies.

For example, although insulin therapy can help keep blood-glucose levels within the recommended range after T1D develops, it is not a cure and in some cases is unable to protect people from T1D’s serious effects, which may include damage to the heart, kidneys, eyes and other organs4. If clinicians can more rapidly identify signs that the disease will develop quickly, they may be able to better develop strategies to help mitigate those effects.

These AI-based T1D progression models we have built with JDRF potentially hold enormous potential for the treatment and management of T1D. Although much work remains to be done and validated, we look forward to exploring how this research lays the groundwork for earlier T1D detection and screening, and how it can help clinicians to build more personalized screening and effective treatment plans.

 We are presenting these and other findings this weekat the American Diabetes Association’s 79th Scientific Sessions in San Francisco, one of the world’s largest diabetes conferences. In addition to our work with JDRF, the Sugar.IQ diabetes assistant was launched with Medtronic last year at ADA, and real-world findings will be presented during this year’s Scientific Sessions. These demonstrate how Sugar.IQ is helping users to more proactively manage their diabetes with meaningful, personalized insights everyday.

For example, new data revealed that people using the Guardian™ Connect system with the Sugar.IQ™ app experienced approximately one extra hour per day within target range, and found the personalized insights from Sugar.IQ helpful in managing their diabetes 91 percent of times. The machine learning models in the Sugar.IQ app are also helping users predict the likelihood of a hypoglycemic event up to four hours in advance. This level of predictive insight can help users alleviate the stress of not knowing when they are at-risk of an impending diabetes-related event and action before going out of target range.

Looking Ahead

Our research could open the door for more precise diagnostic T1D testing and screening. There is currently no standard clinical practice for screening for T1D antibodies, and even if risk factors are determined based on genetics, clinicians often tend to use a “one-size-fits-all” model for how frequently to monitor individuals – even though the disease’s progression can vary drastically from person to person.

Additionally, this research could potentially help doctors zero in on clusters of those living with T1D and potentially pave the way for other studies that help researchers better understand the root causes of the disease. The exact cause of T1D is still shrouded in mystery, although researchers are studying the impact that hereditary and environmental factors may have on the disease’s onset.

As studies into the complexities of T1D continue to expand, we look forward to this work playing a role in the quest to help alleviate T1D, its complications, and the burden it places on the people it affects.

  1. CDC National Diabetes Statistics Report, 2017
  2. “Economic Costs of Diabetes in the U.S. in 2017,” Diabetes Care, 2018
  3. JDRF:Impreatore, et. Al 2012. Diab Care 35: 2512-2520; Dabelea, et. al, 2014, JAMA
  4. http://care.diabetesjournals.org/content/38/10/1964

 

IBM Fellow and Global Science Leader, AI for Healthcare, IBM Research

More Healthcare stories

Novel AI tools to accelerate cancer research

At the 18th European Conference on Computational Biology and the 27th Conference on Intelligent Systems for Molecular Biology, IBM will present significant, novel research that led to the implementation of three machine learning solutions aimed at accelerating and guiding cancer research.

Continue reading

Quantitative Immunohistochemistry: Adding new sight to a historic technique

Using microfluidic probe, IBM scientists perform microscale IHC on tissue sections with a gradient of incubation times for the primary antibody.

Continue reading

AI Models Predict Breast Cancer with Radiologist-Level Accuracy

Our team of IBM researchers published research in Radiology around a new AI model that can predict the development of malignant breast cancer in patients within the year, at rates comparable to human radiologists.

Continue reading