# Artificial Intelligence-Driven Risk Stratification in Chronic Kidney Disease Progression: Minimizing Bias via Race-Specific Algorithms

## TL;DR

This study shows that using race-specific AI models can improve fairness and accuracy in predicting chronic kidney disease progression compared to race-blind models.

## Contribution

The novel modular deep-learning architecture enables fairer risk prediction by correcting miscalibration and predictive value disparities in CKD progression.

## Key findings

- Modular race-specific models eliminated calibration bias and aligned positive predictive values across races.
- Pooled models underestimated risk in Black patients and showed unequal predictive performance.
- Modular models preserved discrimination while improving fairness metrics like predictive parity.

## Abstract

Background

Chronic kidney disease (CKD) is a prevalent condition that affects a substantial portion of the adult population and progresses unevenly across different demographic groups. Recent updates to estimated glomerular filtration rate (eGFR) estimation have removed race adjustments to promote greater equity. Yet, the impact of such changes on model performance and fairness across populations remains uncertain.

Objective

To ascertain whether, in comparison to a traditional pooled (or "race-blind") model, a race-specific, modular deep-learning architecture can enhance clinical utility and fairness in a five-year CKD-progression prediction.

Methods

We retrospectively pooled ~30,000 patients with stage 1-4 CKD from databases such as the National Health and Nutrition Examination Survey (NHANES), UK Biobank, and Chronic Renal Insufficiency Cohort Study (CRIC), and two U.S. health-system electronic health records (EHRs). The endpoint was ≥40% sustained eGFR decline, ≥5 ml/min/1.73 m²/year drop, or kidney-failure event within five years. Two fully connected neural-network strategies were trained: (i) a pooled model on all races without race as an input; (ii) a modular model comprising separate subnetworks for Black and White patients, sharing architecture but trained on race-specific data. Performance was evaluated by discrimination (area under the curve or AUC), calibration, decision-curve net benefit, and fairness metrics (predictive parity, equalized odds, statistical parity).

Results

Overall AUCs were comparable (pooled 0.79, modular 0.80). The pooled model systematically underestimated risk in Black patients (calibration-in-the-large -3.8 percentage points (pp)) and yielded unequal positive predictive value (PPV 67.5% Black vs 58.6% White patients). The modular model virtually eliminated calibration bias (intercept ≤0.5 pp) and aligned PPV across races (~64% each) while preserving discrimination. Decision-curve analysis showed a small but consistent net-benefit gain for the modular approach at clinically relevant thresholds (10-35% risk). Trade-offs remained in equalized-odds: the modular model showed higher sensitivity for Black patients (510/840, 60.7%) than for White patients (294/900, 32.7%), though at the cost of a larger false-positive-rate disparity (365/2,160, 16.9% vs 144/3,600, 4.0%). Overall, CKD progression occurred in 1,820/7,500 (24%) patients - 840/3,000 (28%) Black and 900/4,500 (20%) White patients.

Conclusions

Ongoing monitoring and stakeholder-guided threshold setting are crucial to balance competing fairness criteria. Race-specific modular artificial intelligence (AI) models offer a practical route toward fairer, precision risk stratification by correcting miscalibration and PPV inequities inherent in pooled, race-blind CKD risk tools without sacrificing accuracy.