Predicting Type 2 Diabetes Risk: A Non-negative Matrix Factorization Approach for Feature Selection

The use of machine learning (ML) models for predicting type 2 diabetes (T2D) is growing rapidly due to the availability of large patient datasets and advanced ML frameworks. Most studies focus on supervised classification, which relies on labeled examples, but the absence of true negative examples can weaken predictions. Toovercome this, we propose a novel unsupervised method to estimate the likelihood of T2D in undiagnosed individuals. Our study suggests that comorbid conditions such as hyperparathyroidism, iron deficiency anemia, hyperlipidemia, and vitamin D deficiency may contribute to T2D development. Clinical relevance: Our study has the potential to enable early detection of T2D in undiagnosed patients by utilizing their co-occurring health conditions and medication history. This approach may assist healthcare providers in implementing timely interventions, ultimately leading to improved patient outcomes