Researchers use AI to better spot risk factors of Type 1 diabetes

Type 1 diabetes (T1D) is an autoimmune disorder that can affect people at any age, and there is currently no cure. The only hope is to delay or prevent its onset—and this is where our latest machine learning-based research can help.

Our team, composed of scientists from IBM and JDRF—a leading research and advocacy organization for T1D—and five academic research centers in four different countries, has just published a breakthrough study, “Islet Autoimmunity and HLA Markers of Presymptomatic and Clinical Type 1 Diabetes: Joint Analyses of Prospective Cohort Studies in Finland, Germany, Sweden, and The United States.”¹

Appearing in Diabetes Care, it is the first major clinical paper from this collaboration on identifying patients at high risk for the disease. The Type-1 Data Intelligence (T1DI) Study is a large and unique cohort of children followed closely from birth.

Our work has provided insights into development of biomarkers associated with risk of T1D onset in children. We believe that our results could make it easier to find at-risk children for clinical trials focused on delaying, and possibly preventing, the onset of T1D.

Children affected the most

No cure, life-long insulin dependency and possible long-term complications including cardiovascular disease, kidney failure and diabetic retinopathy, which can lead to blindness. This is what T1D is: it's an auto-immune condition that can affect people at any age but is typically diagnosed during childhood or adolescence. In the United States, T1D affects about 1.6 million people, many of whom are young children and adolescents, according to the American Diabetes Association. And this number is on the rise.

The disease typically develops over five to 15 years, with the gradual loss of insulin-producing beta cells in the pancreas. It is this gradual development, over decades, that has prompted scientists to look for ways to help delay or prevent its onset. We are among them, armed with the latest machine learning technology.

First, the IBM Research team created the T1DI Study cohort, the largest of its kind for predictors of childhood T1D. In partnership with JDRF, which brought together a team of over 30 scientists from nine institutions in four countries, we combined the data from five natural history studies of T1D led by those institutions. Some studies began over 30 years ago.

These studies all focused on the development of T1D; however, the study design, duration and data collected was different for each project—meaning we had to integrate the individual datasets in a way that was common to all studies.

The datasets included measurements of islet autoantibodies—biomarkers specific to T1D that can develop and change over time. Biomarkers are measurable substances that can be detected with the help of laboratory testing or other mechanisms, and indicate potential presence or risk of developing a disease. The term 'seroconversion' describes the earliest time point at which such autoantibodies are detected in a blood test and marks the start of autoimmunity.

Since the data was collected over many years and across multiple locations, lab tests for biomarkers used different methods or reporting standards—among sites and over time, as well as genotyping methods differing in resolution. This was an additional complication, so we also harmonized the data in a way that results reported in different ways could be analyzed together.

Advanced machine learning at work

Once data preparation was done, the real analytic work could begin.

We used advanced statistical and machine learning methods, and our research team developed innovative and interactive graphical tools. A paper on one of those visualization tools, DPVis, was published² last year in IEEE Transactions on Visualization and Computer Graphics. Another on simulating population-level screening, using the COOL (Collaborative Open Outcomes tooL), will be presented at the upcoming AMIA 2021 Annual Symposium in November.

Throughout the research, we applied advanced machine learning algorithms to identify predictors of T1D onset. The data we used included patient characteristics such as sex and genotypes known to be associated with onset of T1D. We also relied on laboratory test results collected over time from each study participant, specifically for the three islet autoantibodies associated with development of T1D. Our analysis discovered new patterns of autoantibody development, and their links to other risk factors.

By analyzing the data, we found that the number of islet autoantibodies present at seroconversion, the earliest time point in autoimmunity development, can reliably predict risk of T1D onset in young children for periods up to 10 to 15 years into the future.

For children with multiple autoantibodies—more than one type of islet autoantibody—at the time of seroconversion, the risk of developing T1D is very high, about 90 percent over a 15-year period. Also, the younger the age at which children develop multiple autoantibodies, the greater the risk, peaking between two to four years of age. We also confirmed that genotypes for T1D do not affect the risk in multiple autoantibody-positive children.

Our research has also shed new light on risk of T1D in single autoantibody-positive children. Their 15-year risk of T1D onset is markedly lower, just some 30 percent, especially for children with only a single autoantibody at time of seroconversion and those remaining single after that. Even though this risk seems overall substantial, we found that their individual risk assessment can be improved based on genetic profile and a repeat antibody test in about two years.

In other words, we found that the overall risk for developing T1D for single autoantibody-positive children remains considerably lower, but is especially low if these children do not develop a second autoantibody in the two years that follow seroconversion. Also, children who remain positive for only a single autoantibody, as well as having a low-risk genotype for T1D, have substantially lower risk: just about 12 percent overall, or about one third lower than those with high-risk genotypes.

Helping identify at-risk children

These findings could help identify and stratify participants for trial recruitment based on the number of autoantibodies and genetic results. Similarly, the results could help inform routine screening, monitoring cadence and overall management of at-risk children. Population-based screening and surveillance is typically done for diseases where a cure or immediate treatment is available, which is not yet the case for T1D. However, T1D onset and initial diagnosis is often associated with life-threatening complications of diabetic ketoacidosis (DKA), increasing the importance of early detection.

Early identification of at-risk children could help families and caregivers to better understand the risk and recognize early signs of DKA to reduce its incidence at onset. This is particularly valuable since research has shown reduced DKA rates in study participants who were routinely tested for development of autoantibodies and followed, at least in the constituent T1DI studies.

In light of ongoing research to delay or prevent onset of T1D, such as in the TrialNet consortium, our findings should help inform screening programs to identify high-risk individuals early as potential candidates for such trials. This could benefit not only the children who participate, but also the entire T1D research community.

In addition, our research has validated important previous results³ by the ADA, JDRF and the Endocrine Society. In 2015, these findings led to a proposal for staging T1D based on development of islet autoimmunity as stage 1.

Disease stages are often used in helping clinicians identify when patients should be monitored or treated in different ways. For example, early disease stages may mean less frequent monitoring and limited or no treatment, while later stages may require much more frequent visits and more aggressive treatments or other interventions. Improved staging helps clinicians and caregivers provide the best care for their patients at all stages of disease.

In summary, we have found specific combinations of factors involving autoantibody patterns and genetics that are associated with different rates and probabilities of developing T1D. Our results not only pave the way for better understanding of risk factors for T1D but may also help to develop guidelines for routine screening, monitoring and management of at-risk children.

Such guidelines can help reduce complications at onset and may identify patients who might benefit from participation in ongoing clinical trials intended to delay or prevent onset of T1D.

Hopefully one day they might even help to prevent T1D altogether.

Two of the participating centers in the T1DU collaboration are in the US: DAISY at the University of Colorado, Denver and DEW-IT at the Pacific Northwest Research Institute in Seattle. The other three centers are in Europe: DiPiS, in Sweden at the Department of Clinical Sciences Malmö at Lund University CRC, and Skåne University Hospital in Malmö. The DIPP study was conducted at multiple locations in Finland: the Institute of Biomedicine and Centre for Population Health Research at the University of Turku, Department of Pediatrics at Turku University Hospital in Turku, University of Oulu and Oulu University Hospital, Department of Pediatrics, PEDEGO Research Unit, in Oulu; and in Germany, the BABYDIAB and BABYDIET studies were conducted at Helmholtz Zentrum München in Munich.