A machine learning model for optimizing treatment of patients with poorly controlled type 2 diabetes

Cardiorenal-protective sodium-glucose cotransporter-2 inhibitors (SGLT-2i) and glucagon-like peptide-1 receptor agonists (GLP-1RA) lack selection guidance. We aimed to build a SGLT-2i/GLP-1RA Decision Score (TiP DecScore) to tailor selection between them.

We developed the TiP DecScore in type 2 diabetes (T2D) patients receiving either therapy from the China Metabolic Analytics Project (derivation dataset: n = 24,322; validation dataset: n = 1,459), using gradient boosting decision tree and 15 features. The primary outcomes were glycated haemoglobin (HbA_1c) control (<7%) and HbA_1c levels at 6 and 12 months. The model's clinical effectiveness was evaluated by comparing HbA_1c control between concordant (receiving the predicted optimal therapy) and discordant groups (receiving the predicted non-optimal therapy).

Here we show the derivation cohort has mean (SD) age 53.7 (11.5) years, 63.0% males. Model validation shows good predictive performance (the receiver operating characteristic curve 0.71-0.78). GLP-1RA is favored over SGLT-2i (57.6% vs. 24.2% at 6 months; 57.9% vs. 28.6% at 12 months). At 6 months, compared with SGLT-2i, GLP-1RA is prioritized for patients with a shorter diabetes duration and higher fasting C-peptide, alanine aminotransferase, body mass index (BMI), and low-density lipoprotein cholesterol levels. At 12 months, patients with higher baseline HbA_1c and BMI levels are more likely to be recommended GLP-1RA than SGLT-2i. Higher rates of HbA_1c control are observed in concordant versus discordant groups, especially in younger patients (<55 years; 64.1% vs. 46.2%, P = 0.001) and males (58.6% vs. 45.6%, P = 0.018) at 12 months.

The TiP DecScore effectively guides personalized selection between SGLT-2i and GLP-1RA therapies for T2D patients.