Type 1 diabetes (T1D) and type 2 diabetes (T2D) are very different diseases. From their basic biology (T1D is autoimmune and T2D is metabolic) to how they’re treated (T1D requires exogenous insulin and T2D is usually treated with lifestyle and non-insulin drug interventions. Advanced T2D does require insulin treatment. There’s a lot that separates the two in terms of management, care, and long-term planning.

Unfortunately, there’s one significant trait both have in common: They both cause dysglycemia, or abnormal blood sugars. This factor, in addition to the misconception that T1D only occurs in children, often leads to adults with T1D to be diagnosed with T2D. This can have tragic consequences.

This is not a small problem, either. Researchers found that up to 20 percent of people with T1D can be misdiagnosed with T2D [1]. And recall that nearly 50 percent of new T1D diagnosis happens in adults. So, that is a margin of error that can affect a lot of people!

JDRF and IQVIA teamed up to see if artificial intelligence, or AI, can be used to make an algorithm that can identify individuals with T1D to avoid misdiagnosis as T2D, and the results were just published.

Finding Misdiagnoses

IQVIA is a leading global provider of advanced analytics, technology solutions, and clinical research services to the life sciences industry. In this JDRF-funded project, they used machine learning to scour IQVIA’s Ambulatory Electronic Medical Records (AEMR) database for individuals diagnosed with T2D and then, later, diagnosed with T1D within a specific time frame.

When analyzing the individuals, a few things stuck out between the individuals confirmed to have T2D and those misdiagnosed. Compared to the confirmed T2D, misdiagnosed adults, on average:

• Weighed less
• Were younger
• Were less likely to have high blood pressure
• Less likely to be coded with obesity
• Were more likely to be Caucasian

Other factors over time, like HbA1c and insulin refills, also correlate with having T1D.

Developing an Algorithm

IQVIA took these findings, used them to develop an algorithm, and applied them to a larger data set for testing and validation. Their model showed it can look at these medical records and identify individuals who are diagnosed with T2D but actually have T1D. In theory, this model could be used in real time to correct misdiagnoses.

Turning this model into a diagnostic tool used in the clinic, however, is not straightforward. It is dependent on accurate electronic medical records, which are often incomplete and compiled using different standards and formats. There is often data missing, and they do not capture the entire medical history of the individuals in them. This algorithm also relies on subtle associations between myriad variables that are not easily translatable into clinical guidelines. They are, however, a great starting point.

Artificial Intelligence: The First, But Not the Last

This publication does a job of underscoring the severity of the problem, demonstrating how advanced analysis using AI and health care data can find trends and identify misdiagnoses, and highlighting the potential for future development into T1D screening guidelines and/or a clinical decision support tool.

Go here to read the publication.

[1] Thomas NJ, Lynam AL, Hill AV, Weedon MN, Shields BM, Oram RA, McDonald TJ, Hattersley AT, Jones AG. Type 1 diabetes defined by severe insulin deficiency occurs after 30 years of age and is commonly treated as type 2 diabetes. Diabetologia. 2019 Jul; 62 (7): 1167-1172.

[2] R. Cheheltani, N. King, S. Lee, B. North, D. Kovarik, C. Evans-Molina, N. Leavitt, S. Dutta, Predicting Misdiagnosed Adult-onset Type 1 Diabetes Using Machine Learning, Diabetes Research and Clinical Practice (2022), doi: https://doi.org/10.1016/j.diabres.2022.110029