Ensemble AI Detects CKD with 91% Sensitivity in Diabetics and 92.5% in Non‑Diabetics

Read the paper →

AI-Driven Ensemble Model Enhances Early Detection of Chronic Kidney Disease in Diabetic and Non-Diabetic Patients Across Latin America

New machine learning models using simple, readily available clinical data achieve up to 91% sensitivity in identifying high-risk chronic kidney disease patients – with distinct approaches for diabetics and non-diabetics in a large multicenter Latin American cohort.

Introduction: The Challenge of Chronic Kidney Disease Detection

Chronic kidney disease (CKD) silently affects over 10% of adults worldwide, often progressing unnoticed until reaching advanced stages. This delayed diagnosis contributes to high morbidity and mortality, particularly in regions with limited access to comprehensive laboratory testing. Early identification of CKD is critical to slow progression and reduce complications, yet traditional screening relies heavily on specialized lab tests like estimated glomerular filtration rate (eGFR) and urine analysis that may not be consistently available in low-resource settings.

Recent advances in machine learning (ML) offer promising tools to identify patients at risk of CKD by analyzing diverse clinical data. However, current algorithms often depend on extensive laboratory data that may be costly or incomplete in many healthcare environments, especially across Latin America. Furthermore, prior models frequently focus solely on diabetic populations, limiting their broader applicability.

In this context, researchers developed an ensemble ML approach tailored for both diabetic (T2D) and non-diabetic (NT2D) patients, leveraging simple clinical parameters routinely collected at point-of-care. Their models achieved strong sensitivity metrics, marking a significant step toward scalable CKD screening in resource-constrained settings.

Study Partnership & Context

This study was a collaborative effort among Arkangel AI, AstraZeneca Colombia, and Universidad de Caldas, spanning healthcare institutions across Colombia’s Caribbean region and Peru. The Latin American setting is uniquely important given the high burden of CKD coupled with limited laboratory infrastructure, making cost-effective, easily deployable screening tools crucial.

The patient population represented a large and diverse cohort including both diabetic and non-diabetic individuals, enabling nuanced, population-specific model development. By attending to this regional context and patient heterogeneity, the study addresses a clear gap in CKD risk identification tools applicable to real-world Latin American healthcare environments.

Study Design and Methodology

The researchers retrospectively analyzed clinical records from three databases collected between undetailed but recent years, encompassing 19,194 diabetic and 169,842 non-diabetic patients. Patients lacking diabetes diagnosis dates or relevant kidney function markers (creatinine, eGFR) were excluded. CKD risk was defined by eGFR < 60 mL/min/1.73 m² or prior official CKD diagnosis.

Input features comprised simple clinical variables such as age, sex, body mass index (BMI), hypertension status, and diabetes duration for diabetics, deliberately excluding complex lab tests to maximize accessibility.

An AI web platform named Arkangel AI was utilized to train, test, and compare various ML algorithms, including Random Forest classifiers and deep neural networks. The best-performing models for each subgroup were combined into ensemble models—weighted averages of model outputs—to enhance predictive sensitivity.

The team also applied SHAP (Shapley Additive explanations) analysis to interpret feature importance and ensure that model decisions aligned with clinical knowledge, confirming that age, hypertension, and sex were influential in non-diabetics, while age, BMI, and diabetes duration dominated diabetic predictions.

Key Results

• For diabetic patients (T2D):
  • Ensemble model (Random Forest + Neural Network with 2:1 weight) achieved 91% sensitivity (vs 81.5% and 97.5% individually for constituent models)
  • Specificity decreased to 39%, trading off some false positives to prioritize catching true CKD cases
  • Area under ROC curve (AUC) was 0.65, reflecting reasonable discrimination
  • Accuracy stood at 69%, with F1-score of 0.77 balancing sensitivity and precision
• For non-diabetic patients (NT2D):
  • Deep neural network alone yielded moderate sensitivity of 92.5%
  • High specificity of 97.2% and very strong AUC of 0.95 indicated excellent performance
  • Precision and accuracy were similarly robust at 93% and 96%, respectively

These results underscore the ensemble approach’s advantage in diabetic patients by increasing sensitivity substantially, which is critical for screening purposes. Meanwhile, the neural network for non-diabetics offered a highly balanced and accurate classification.

Interpretation & Clinical Implications

This study presents practical, low-cost ML tools that reliably identify individuals at high risk of CKD in both diabetic and non-diabetic groups using routinely gathered clinical features. Prioritizing sensitivity is appropriate for screening contexts where missing early CKD cases could lead to delayed care and poorer outcomes.

For clinicians, these algorithms offer a data-driven, interpretable approach to flag patients needing further confirmatory testing and early intervention without relying on extensive laboratory panels. Health systems in Latin America and other resource-limited settings stand to benefit by deploying such models to optimize CKD screening workflows, potentially reducing disease burden and costs.

However, some trade-offs exist, particularly the reduced specificity in diabetic patients’ ensemble model, which may lead to more false positives. Future work should explore calibration based on population CKD prevalence and integration with additional clinical variables to enhance precision.

Deployment & Scalability Considerations

The AI models were developed and tested using Arkangel AI, a web-based application facilitating model training and output generation, positioning this work well for integration into electronic health record workflows. Since the models rely only on simple demographics and clinical data, they circumvent barriers imposed by expensive or missing lab tests, enabling scalable deployment in diverse hospital or primary care settings.

Challenges to implementation may include physician engagement, data quality variability, and ensuring local adaptation to different epidemiological patterns outside the Caribbean and Peru. However, the model’s flexibility to weigh diabetic and non-diabetic data separately highlights adaptability.

This approach could also be adapted for screening other chronic conditions sharing risk factors or clinical features, broadening its utility in population health management.

Conclusion & Future Directions

This multicentric Latin American study demonstrates that ensemble machine learning models trained on readily available clinical data can effectively identify patients at high risk of CKD, particularly improving sensitivity in diabetics while maintaining strong overall performance in non-diabetics. It provides a promising roadmap towards affordable, scalable CKD screening in resource-constrained healthcare systems.

Future steps should include prospective validation in clinical settings, optimization of false positive rates, and expansion to other regions and comorbidities. The integration of transparent, interpretable AI tools like these can empower clinicians and health systems to address the global CKD epidemic more proactively.

Reference: Martinez J, Perez A, Zea J, Llano I, Castaño-Villegas N, Caro D, Arango JJ. Development of an ensemble learning algorithm to detect patients at high risk of Chronic Kidney Disease using readily available clinical features. (2024)