The study by Choi et al. [5] is a leading effort in applying machine learning and deep learning to predict PTDM within 1 year after KT, using data from over 3,200 recipients in the nationwide KOTRY (Korean Organ Transplantation Registry) cohort. There have been several previous attempts to predict PTDM applying deep learning or machine learning techniques to clinical data, imaging, or metabolomics (Table 1) [5–8]. The strength of the present study is that it represents one of the largest and most methodologically advanced studies on PTDM risk prediction to date. The authors used a well-designed and powerful system to handle the data. They curated 236 variables from recipient and donor profiles. They then performed strategic feature selection to distill 72 clinically meaningful variables (41 pretransplant and 31 posttransplant) based on missing data thresholds and medical relevance. The dataset was then split into training, validation, and test sets with stratified five-fold cross-validation, ensuring model robustness [5]. Four machine learning algorithms—XGBoost, CatBoost, lightGBM, and logistic regression—were benchmarked against a deep learning architecture that combined multilayer perceptrons and long short-term memory networks to process both static and sequential data. Performance metrics included area under the curve (AUC), accuracy, precision, recall, and F1 score.

The XGBoost model emerged as the top performer, with an AUC of 0.738, an accuracy of 0.86, and an F1 score of 0.42. Notably, CatBoost demonstrated the highest precision (0.79), although its recall was limited. Deep learning showed superior recall (0.64) but suffered from a low precision (0.27), reflecting its sensitivity at the cost of specificity [5]. These findings affirmed that in the context of chronic diseases such as PTDM, models with a balanced predictive capacity, such as XGBoost, might be more clinically useful than those favoring either sensitivity or specificity alone. The authors highlighted key predictors of PTDM, with top factors including recipient age, baseline body mass index (BMI), HbA1c, high-density lipoprotein cholesterol (HDL-C), 6-month triglyceride and uric acid levels, white blood cell (WBC) count, and tacrolimus trough levels at discharge. These predictors aligned well with known pathophysiological mechanisms of PTDM, underscoring the model’s biological plausibility. For example, old age and BMI are well-established PTDM risk factors. Elevated triglycerides and low HDL-C reflect components of metabolic syndrome. Tacrolimus, a calcineurin inhibitor, is notorious for inducing posttransplant hyperglycemia via beta-cell toxicity and insulin resistance [1,3]. The inclusion of uric acid and WBC count might also reflect systemic inflammation and oxidative stress as emerging contributors to posttransplant metabolic derangements [9]. By providing reasonable insights into individual risk factors, these models are expected to enable clinicians to engage in meaningful shared decision-making with patients and potentially guide interventions such as lifestyle counseling, medication selection, and closer glycemic surveillance.

Despite its merits, this study has some limitations. First, PTDM was diagnosed exclusively using fasting plasma glucose levels, which might have underestimated the true incidence compared to oral glucose tolerance testing [10]. Additionally, the lack of data on cumulative corticosteroid exposure and time-averaged tacrolimus levels—two key contributors to PTDM—might have limited model granularity. Second, while the model performed well in internal validation, external validation across diverse populations, immunosuppressive protocols, and healthcare systems is essential to establish generalizability. We need to consider racial and ethnic differences in PTDM risk, health conditions, and access to medical care before using it widely. Third, although deep learning has some advantages for working with complex time-based data, it did not perform very well in this study. This suggests that we might need bigger and more detailed long-term data to see its full power. Finally, more research is needed to find the best way to combine physician’s knowledge with the model’s performance. Hybrid models that combine expert knowledge with data could make results more accurate and easier to accept. Nevertheless, clinical implications of this study are profound. First, integrating predictive models such as XGBoost into electronic medical records could enable automated risk scoring at the point of care, prompting early screening and intervention. Second, transplant centers could use these models to tailor immunosuppressive regimens, potentially selecting agents with lower diabetogenic potential for high-risk patients. Moreover, in a research setting, these models could stratify patients for interventional trials aiming to prevent PTDM, thereby improving study efficiency and outcome clarity. They could also help healthcare systems allocate resources more efficiently by targeting follow-up care and lifestyle modification programs to those most in need.

In conclusion, Choi et al. [5] have set a new standard for predictive analytics in transplantation. Their study not only demonstrates the feasibility of using machine learning to forecast PTDM but also offers a framework for other posttransplant complications. As machine learning becomes increasingly embedded in clinical workflows, such models will serve as critical tools for moving from reactive care to proactive prevention. Ultimately, the goal is not merely to predict PTDM, but to prevent it. By combining predictive power AI with the clinical insight of an experienced transplant nephrologist, we are getting closer to making personalized medicine a reality in KT.