Enhancing post-kidney transplant prognostication: an interpretable machine learning approach for longitudinal outcome prediction

Abstract

Kidney transplantation offers life-extending treatment for patients with end-stage renal disease, yet long-term risks of graft loss and death persist. Traditional prediction models using only baseline data often fail to capture patients' evolving health status post-transplant. In this study, we propose a two-stage machine learning (ML) framework for dynamic, next-year risk prediction of graft loss and death, updated annually with newly available clinical and laboratory data. Using a multi-center cohort from the Swiss Transplant Cohort Study (STCS), we trained and evaluated five ML models across 13 years of follow-up, demonstrating that incorporating longitudinal data significantly improved predictive performance compared to baseline-only models. LightGBM achieved the strongest performance, with AUROC values up to 0.896 for graft loss and 0.797 for death. Our findings suggest that dynamic, interpretable ML models can enhance personalized risk stratification, offering a practical and scalable tool for guiding follow-up strategies and early interventions in kidney transplant recipients.

Fig. 1. Overview of the event distribution and two-stage modeling approach.

a Event distribution. The first two plots show, for death and graft loss, the number of at-risk patients over successive post-transplant years (left y-axis) together with the number of events observed in each year (right y-axis). The stacked histograms depict the distribution of follow-up lengths: each x-axis bin corresponds to the post-transplant year in which a patient was last observed (years 1–13). Bars are divided into patients who remained event-free (censored; light blue) and those who experienced the event (dark blue); the total column height therefore represents the proportion of the cohort whose follow-up ended in that year because of either censoring or the event. A substantial fraction of events occur within the first five years, although many patients remain at risk much longer. b Two-stage modeling. (Left) The Baseline model uses only pre-transplant data to estimate the risk of an outcome during the first post-transplant year. (Right) For every subsequent event-free year t (t = 1, 2, …, 12), the Follow-up model uses the baseline data plus the follow-up measurements collected during year t (-Year t data-) to estimate the risk of an outcome in year t + 1 (-Year t + 1 outcome risk-). Thus, the first follow-up prediction occurs at the end of Year 1 (using Year 1 data to predict Year 2 risk), and the process repeats annually until an event or censoring.

Fig. 2. Year-by-year performance of the LightGBM model.

Shown for (Left) AUROC and (Right) AUPRC. Light blue points represent Graft Loss, while dark blue points represent Death. In the AUPRC plot, the incidence proportion of the events is also visualized using dashed lines under the curves. The shaded area indicates the 95% confidence intervals to quantify the uncertainty of the prediction.

Fig. 3. SHAP value analysis for the follow-up model predictions.

Each outcome type is visualized using three panels: (Left) SHAP beeswarm plots illustrating feature impact distributions of the top 10 features. The (B) after the feature name indicates a baseline characteristic. High feature values are depicted in dark blue, low values in lighter blue. (Middle) the top 10 features ranked by mean absolute SHAP values. (Right) 2D scatter plots showing the relationship indicates a baseline characteristic between SHAP values and actual feature values for the most important predictor. Panel (a) corresponds to the death prediction, and Panel (b) corresponds to the graft loss prediction. These visualizations provide interpretability by highlighting how individual features influence the model predictions.

Fig. 4. Subcohort analysis for the follow-up model.

Performance is shown across subgroups defined by variables such as sex, donor type, blood group, ethnicity, and transplant center. Panel (a) and (b) AUROC and AUPRC for death prediction respectively; Panel (c) and (d) AUROC and AUPRC for graft loss prediction respectively. Bars indicate the performance metric values in each subgroup, while the diamonds in the AUPRC plots represent the incidence proportion of the outcome within that subgroup. For donor-type specific analyses, ''Living Donor: Related vs. Unrelated” includes only patients with living donor kidneys, while ''Deceased Donor: DBD vs. DCD” includes only deceased donor recipients.