Further Perspectives on Predictive Modeling for Diabetes Subtypes Classification in India

This article is in response to Bhardwaj and Rastogi.[1] We thank the authors for their thoughtful letter in response to our article, Predictive Modeling for Diabetes Subtype Classification in India: A machine learning approach.[2] The initial point of concern is the handling of missing data. Due to the size of the dataset and to maintain its integrity, we decided to exclude records with missing values. Moreover, among the 20% of excluded data, more than 16% (16.4%) were due to the short duration and inability to classify. Since the missing data is <5%, as a rule of thumb, the complete case analysis may be used as a primary analysis.[3-5] Regarding the characteristics of excluded versus included participants to assess and mitigate this bias, we have presented the characteristics of the available sample in each model in Supplementary Tables 1 and 2 of our article.[2] We fully agree that the limited access to glutamic acid decarboxylase (GAD) antibody testing, especially in resource-limited settings, is a major challenge. To address this, we developed and validated four models without GAD data. Models 3 and 4 in our article,[2] which included clinical features, family history, lipids, and glycated hemoglobin, showed performance comparable to the GAD-based model, offering feasible alternatives where GAD testing is not available. We recognize the need for accessible, cost-effective biomarkers to improve model utility. While C-peptide is a potential alternative, we excluded it since it was used in defining type 1 diabetes mellitus (T1DM) and type 2 diabetes mellitus (T2DM). In future work, we aim to explore clinical and biochemical features that could replace GAD antibody status. We value the author’s insights on study design and duration. Although the study spans nearly three decades, we minimized heterogeneity from evolving clinical practices by applying consistent diagnostic criteria, rather than relying on historical classifications. This standardized reclassification reduced variability and ensured dataset consistency, limiting the impact of temporal changes on model performance. We also recognize the need for prospective validation and plan periodic updates of the model for continued relevance in modern clinical settings. Thank you for raising the point about the 10- to 30-year age range. We intentionally focused on this group, as misclassification between T1DM and T2DM is most common in this age group[6] due to rising obesity in youth with T1DM[7] and earlier onset of T2DM, especially in India.[8] Several Indian epidemiological studies and clinical guidelines define “young-onset diabetes” as age before 25 years[9,10]; however, T1DM is almost always diagnosed before 10 years,[11] therefore, we set 10 years as the lower limit. To capture more overlapping subjects. We extended the upper age limit to 30 years. We acknowledge the absence of regional and ethnic breakdowns in the article. Our models were developed specifically for the Indian population to fill a critical gap, as there is currently no widely accepted diagnostic model tailored to the Indian context. While our dataset was multi-regional, future studies will assess model performance across regions through stratified analyses to improve the model’s accuracy and clinical relevance across India’s diverse population. We acknowledge that the models have not yet been externally validated using an independent dataset, which is essential for confirming generalizability. This will be a key focus of future work. As for a user-friendly clinical tool, our tool is platform-independent and can be accessed anywhere on any device. We also provided the full model equation to allow manual probability calculation when digital access is limited. We accept that the relatively low prevalence of T1DM in the dataset (approximately <10%) could affect the model stability, but in our dataset, it was over 20% (22.1%). While class imbalance was not severe, we used SMOTE to upsample T1DM subjects and retrained all models. Performance remained comparable, indicating robustness. To further ensure model reliability, we implemented multiple validation techniques, including k-fold cross-validation, calibration plots, log-likelihood ratio tests, net reclassification index, and finally sensitivity analyses using GAD available data. These steps collectively support the model’s validity despite class imbalance and high feature dimensionality. Financial support and sponsorship Nil. Conflicts of interest There are no conflicts of interest.