Explainable machine learning integrating bioelectrical impedance for 6-month cardiovascular risk in peritoneal dialysis

Clinical Snapshot

Key Findings

• P Value: Not reported in abstract

• Effect Size: AUC = 0.88 for random forest model

• Primary Outcome: Major adverse cardiovascular events within 6 months

• Nnt Or Sensitivity: Sensitivity 0.84 with BIS integration (improved from 0.68 with medical history alone)

• Confidence Interval: Not reported in abstract

Clinical Application

Moderate - requires BIS equipment and trained personnel, but uses clinically accessible features Relevant to Australian nephrology practice where PD is established treatment; BIS increasingly available in major dialysis centres; aligns with Kidney Health Australia guidelines for cardiovascular risk management in CKD Clinically stable peritoneal dialysis patients in tertiary care settings with access to bioelectrical impedance spectroscopy

Abstract

Peritoneal dialysis (PD) is a common treatment for end-stage renal disease, yet cardiovascular disease (CVD) remains a major cause of morbidity. Inadequate fluid management can elevate CVD risk, and bioelectrical impedance spectroscopy (BIS) is increasingly used to assess fluid status. This study aimed to develop an artificial intelligence model integrating BIS measurements and medical history to predict major adverse cardiovascular events (MACEs) within six months in clinically stable PD patients. Data were stratified by MACE occurrence and subject, with 80% assigned to training and 20% to testing. Class imbalance was addressed using the synthetic minority over-sampling technique. Four algorithms - logistic regression, random forest, XGBoost, and deep neural networks - were trained using fivefold cross-validation and grid search for optimal hyperparameters. Model performance was evaluated with area under the ROC curve (AUC), calibration plots, and decision curve analysis (DCA). Feature ablation experiments compared models using all 15 features, only 11 BIS-recorded features, and only 4 medical history features. The random forest model achieved the highest performance (AUC = 0.88), with CVD history as the most influential predictor. Isotonic regression calibration improved probability alignment (brier score = 0.0451). DCA suggested potential clinical benefit. While the model relied on four medical history features, its initial sensitivity was modest (0.68). Integrating BIS features significantly enhanced diagnostic sensitivity (0.84). The random forest model, based on 15 clinically accessible features, accurately predicts the risk of MACE within six months in clinically stable PD patients, demonstrating strong discriminatory ability with a performance reaching 88%.

References

1. 1.Tsai, Y. L., Wu, C. L., & Chiang, J. H. (2026). Explainable machine learning integrating bioelectrical impedance for 6-month cardiovascular risk in peritoneal dialysis. Renal Failure, 48(1). https://doi.org/10.1080/0886022X.2026.2660005