Development of an Artificial Intelligence-Based Prediction Model for Chronic Kidney Disease Outcomes

Chronic kidney disease (CKD) is a common condition that imposes a substantial health burden and contributes significantly to global morbidity and mortality. In the United States, 2024 data indicate that about 14% of adults - more than 31 million people - have CKD, costing hundreds of billions of dollars each year. In Vietnam, the estimated prevalence is 12.8%, affecting roughly 10 million people. Because CKD often progresses silently, reliable early prediction of adverse outcomes - end-stage kidney disease (ESKD), disease progression, and death - carries considerable clinical value. Timely intervention in high-risk patients can improve quality of life and reduce morbidity, mortality, and the costs arising from kidney replacement therapy. Several statistical models predict CKD outcomes from variables such as age, sex, eGFR, and albuminuria. However, most were developed predominantly in White populations, and evidence for their generalizability to other ethnic groups, including Vietnamese, remains scarce. Some models omit proteinuria despite its strong prognostic role in CKD, and most do not account for therapies proven to slow progression, such as renin-angiotensin-aldosterone system (RAAS) inhibitors and sodium-glucose cotransporter-2 (SGLT2) inhibitors.

Machine learning (ML), a branch of artificial intelligence, enables computers to learn latent patterns from data and make predictions without being explicitly programmed. Compared with traditional statistics, ML can represent complex, non-linear, and highly collinear relationships that conventional regression may miss, and has recently shown superior predictive performance across many clinical settings.

Contemporary CKD care has advanced substantially: landmark trials have established the renal and cardiovascular benefits of SGLT2 inhibitors regardless of diabetes status, and current KDIGO guidance emphasizes risk-based, individualized management. Prediction models built before this therapeutic era may no longer capture current risk adequately.

The investigators therefore propose to develop an artificial intelligence-based model to predict CKD outcomes suited to the new treatment era in the Vietnamese population. Outcomes comprise disease progression (a ≥ 40% decline in eGFR or ESKD) and renal or cardiovascular death. Predictors are restricted to baseline comorbidities and routine blood and urine tests that are widely recommended for CKD monitoring.

Using a prospective cohort, the investigators will determine the 2-year incidence of these composite events, develop and compare several ML algorithms (logistic regression, random forest, decision tree, Naïve Bayes, k-nearest neighbours, and support vector machine...), benchmark them against existing equations (KFRE and CKD-PC), and select the optimal model, using SHAP-based interpretation to clarify each predictor's contribution. The minimum sample size of 1,182 was derived using the method of Riley.