BAYESIAN DEEP LEARNING ALGORITHMS TO PREDICT THE RATE OF DETERIORATION IN CHRONIC KIDNEY DISEASE

Introduction:

Chronic Kidney Disease (CKD) poses a significant global health challenge, directly increasing the risk of End-Stage Renal Disease (ESRD) and indirectly raising the risk of cardiovascular disease.

Despite the prevalence of early-stage CKD within the population, these stages are often clinically silent and can only be detected through laboratory methods. Early detection and accurate prediction of disease progression are of paramount importance, as they allow for the implementation of preventive strategies to reduce the incidence of ESRD, cardiovascular disease, and their associated substantial financial burdens.

Given the intricate challenges associated with early identification and management of CKD, this study aims to develop and validate Bayesian Deep Learning Algorithm for predicting the rate of CKD deterioration. This novel approach provides clinicians with a powerful tool for timely intervention, potentially enhancing patient outcomes and reducing the burden of CKD on healthcare systems.

Methods:

The study was conducted at seven major Apollo Hospitals in India using a retrospective cohort design. Over 60k CKD Diagnosed patient records aged between 20 and 80 years were initially considered. After applying specific exclusion criteria, a final cohort of 1,529 patients was obtained. These patients showed disease progression over a period of 3 years, characterized by a decrease in estimated glomerular filtration rate (eGFR). The clinical parameters considered in the study included age, gender, type 2 diabetes, hypertension, dyslipidemia, urological problems, systemic diseases, significant cardiovascular disease, and lab parameters such as Creatinine, Albumin, Urea, Uric Acid, and serum electrolytes. Bayesian Neural Networks (BNNs) were employed for modelling to predict the rate of eGFR deterioration. The models were developed and validated using a K-fold cross-validation technique (K=5) to prevent overfitting or underfitting.

Results:

Multivariate Odds Ratio showed significance for diabetes (OR 3.2 CI-2.85-3.55), age (OR 2.6 CI-2.14-2.93), hypertension (OR3.3 CI-2.97-3.63) amongst the clinical features. Similarly, feature importance was high for creatinine, eGFR, urea, hemoglobin, albumin and sodium levels with the outcome variable being the Rate of deterioration or Progression of CKD. The results for regression equations are provided in Figure 1 with R2 being notable in early CKD changes (G1 - >90 ml/min/1.73 m). Overall Bayesian Deep Learning Regression results with K-Fold show an average R2 - 0.4265 with Mean Squared Error - 0.0147. A Bayesian Neural Network (BNN) Classifier is used to extend NN by incorporating uncertainty estimates related to eGFR deterioration along with K-Fold Validation. It's AUC at threshold 0.05 shows 0.85 (AP at 0.83) and 0.10 is at 0.88 (AP at 0.73) with varying Thresholds of eGFR deterioration at 0.05, 0.10, 0.15 & 0.20 (ml/min/1.73m)/day multiple classification models were Developed.

Conclusions:

This research highlights the effectiveness of Bayesian Deep Learning in accurately predicting CKD progression, enabling early detection and intervention. The study findings if validated in subsequent trials will lead to significant medical, financial, and resource allocation benefits. It will also have a potential for broader AI applications in Chronic Disease Management.

I have no potential conflict of interest to disclose.

I did not use generative AI and AI-assisted technologies in the writing process.