Predictive Modeling for Diabetes Subtype Classification in India: A Machine Learning Approach

Dear Editor, I read with great interest the article by Venkatesan et al.[1] “Predictive Modeling for Diabetes Subtype Classification in India: A Machine Learning Approach,” published in the Journal of Diabetology (2025). This study significantly advances diabetes care in India by introducing a machine learning model to differentiate type 1 and type 2 diabetes in adolescents and young adults. The authors’ use of a large, diverse dataset and stepwise modeling is commendable, addressing a critical clinical challenge. However, several additional limitations and opportunities for improvement could further enhance its impact and relevance. The study’s exclusion of approximately 20% of the initial cohort due to missing values or short disease duration, along with the inclusion of only participants with complete model data, raises concerns about potential selection bias and limits generalizability of findings.[2] Additionally, glutamic acid decarboxylase (GAD) antibody data were available for only about one-third of participants due to financial constraints, which, despite enhancing model accuracy, restricts its clinical applicability.[3] Future research should employ advanced imputation techniques and compare characteristics of excluded versus included participants, as well as explore alternative biomarkers or validate model performance without relying on the GAD antibody status to improve broader relevance. The retrospective design of the study, spanning nearly three decades, introduces potential heterogeneity in diagnostic criteria, laboratory methods, and clinical practices. Such variability could influence the consistency of the data and, consequently, the model’s performance.[4] Prospective validation in a contemporary cohort would help confirm the robustness and relevance of the model in current clinical practice. The study’s age range limitation—restricted to individuals aged 10–30 years—leaves its applicability to younger children and older adults untested, reducing clinical relevance across India’s diverse diabetes onset demographics. While described as multi-regional, the lack of detail on regional and ethnic representation weakens the model’s generalizability; stratified analysis is needed to ensure consistent performance across India’s varied subpopulations. Furthermore, the model was validated internally but not on an independent external dataset. External validation is essential to establish the generalizability of the model before it can be widely adopted in clinical practice.[5] Additionally, while the stepwise modeling approach is methodologically sound, the lack of a user-friendly clinical tool, such as an online calculator or mobile application, may limit the model’s adoption by healthcare providers. Developing such tools would greatly facilitate integration into routine clinical workflows.[6] Despite using cross-validation, the small number of type 1 diabetes cases and many features may still risk overfitting; validating with larger cohorts and applying regularization could strengthen the reliability. Overall, Venkatesan et al.’s study marks meaningful progress in classifying diabetes subtypes in India, and addressing these added limitations will enhance its clinical value. Financial support and sponsorship Nil. Conflicts of interest There is no conflict of interest.