Implementing a clinical prediction model in primary care data to identify individuals with misclassified diabetes and increased hypoglycaemia and DKA risk

Problem

Correct classification of type 1 (T1D) and type 2 diabetes (T2D) is essential to ensure correct treatment. However, distinguishing between them can be challenging, leading to misclassification and inappropriate treatment which may increase the risk of poor patient outcomes. Clinical prediction models are available to aid clinicians where there is uncertainty around diabetes type; in this study we assess their feasibility in identifying misclassified diabetes in primary care record data and whether misclassified individuals are at higher risk of hypoglycaemia and DKA.

Approach

We adapted a validated T1D prediction model based on clinical features (age, BMI, lipids) to run in primary care data (CPRD Aurum: n=97,476 adults currently registered with adult-onset, insulin-treated diabetes; 21% diagnosed as T1D). Misclassified individuals were defined as those with a T2D diagnosis and T1D probability ≥70% (misclassified as T2D), or T1D diagnosis with T1D probability ≤5% (misclassified as T1D). Hypoglycaemia and DKA outcomes were compared to controls without evidence of misclassification (T2D controls: T2D with T1D probability <70%; T1D control: T1D with T1D probability >5%).

Findings

94% of patients had all features required for the T1D prediction model. 794 (1.1%) of those diagnosed with T2D had high T1D probability indicating potential misclassification. Compared to T2D controls, these individuals had higher rates of hospitalisation due to hypoglycaemia (8.4% vs 4.1%) and DKA (13% vs 3.2%) and were more likely to commence insulin within 1 year of diagnosis (32% vs 12%, all p<0.001). 3,721 (19.4%) of those diagnosed with T1D had low T1D probability, and had lower rates of hospitalisation for hypoglycaemia (5.9 vs 8.9%) and DKA (19% vs 25%, all p<0.001) than T1D controls. However, those misclassified as T1D had the highest rates of DKA at diagnosis (8.5% vs 5.3% for T1D controls, vs 0.9% for misclassified as T2D, vs 0.5% for T2D controls, all p<0.001). All four groups had similar median HbA1c (67-68 mmol/mol).

Consequences

Individuals with misclassified diabetes can be identified in primary care data using an adapted T1D prediction model, and those misclassified as T2D have significantly poorer hypoglycaemia and DKA outcomes than those without evidence of misclassification. Those misclassified as T1D have a high rate of DKA at diagnosis which may contribute to misclassification by clinicians. We are currently assessing the feasibility of implementing the clinical prediction model as an automated search in GP practice records.