[Paper Review] Binary Gaussian Copula Synthesis: A Novel Data Augmentation Technique to Advance ML-based Clinical Decision Support Systems for Early Prediction of Dialysis Among CKD Patients

Hamed Khosravi, Srinjoy Das+2 more|arXiv (Cornell University)|Mar 1, 2024

Artificial Intelligence in HealthcareHealth Professions3 citations

TL;DR

This paper proposes Binary Gaussian Copula Synthesis (BGCS), a two-stage data augmentation framework tailored for binary electronic health record (EHR) data in chronic kidney disease (CKD) patients. BGCS uses a Gaussian copula to model pairwise dependencies among binary clinical features and applies a fine-tuned GPT-2 classifier to filter out implausible synthetic samples, significantly improving minority-class recall (median 0.78–0.87) for 90-day dialysis prediction while preserving data distribution fidelity (mean p-value = 0.68).

ABSTRACT

The Center for Disease Control estimates that over 37 million US adults suffer from chronic kidney disease (CKD), yet 9 out of 10 of these individuals are unaware of their condition due to the absence of symptoms in the early stages. It has a significant impact on patients' quality of life, particularly when it progresses to the need for dialysis. Early prediction of dialysis is crucial as it can significantly improve patient outcomes and assist healthcare providers in making timely and informed decisions. However, developing an effective machine learning (ML)-based Clinical Decision Support System (CDSS) for early dialysis prediction poses a key challenge due to the imbalanced nature of data. To address this challenge, this study evaluates various data augmentation techniques to understand their effectiveness on real-world datasets. We propose a new approach named Binary Gaussian Copula Synthesis (BGCS). BGCS is tailored for binary medical datasets and excels in generating synthetic minority data that mirrors the distribution of the original data. BGCS enhances early dialysis prediction by outperforming traditional methods in detecting dialysis patients. For the best ML model, Random Forest, BCGS achieved a 72% improvement, surpassing the state-of-the-art augmentation approaches. Also, we present a ML-based CDSS, designed to aid clinicians in making informed decisions. CDSS, which utilizes decision tree models, is developed to improve patient outcomes, identify critical variables, and thereby enable clinicians to make proactive decisions, and strategize treatment plans effectively for CKD patients who are more likely to require dialysis in the near future. Through comprehensive feature analysis and meticulous data preparation, we ensure that the CDSS's dialysis predictions are not only accurate but also actionable, providing a valuable tool in the management and treatment of CKD.

Motivation & Objective

• Address the severe class imbalance in CKD patient data, where only a small fraction progress to dialysis, limiting machine learning model performance.
• Overcome the limitations of existing data augmentation methods that are not designed for the binary structure of clinical EHR features.
• Develop a method that preserves the statistical dependencies among binary clinical variables while generating clinically plausible synthetic samples.
• Improve the performance of machine learning models in predicting 90-day dialysis onset using real-world EHR data from 15,169 CKD patients.
• Integrate the best-performing model into an interpretable clinical decision support system to enhance risk stratification and clinical utility.

Proposed method

• Apply a Gaussian copula framework to model pairwise dependencies among binary clinical features, capturing multivariate relationships in EHR data.
• Generate synthetic minority-class samples (patients progressing to dialysis) using the copula model, preserving the correlation structure of real data.
• Fine-tune a GPT-2 language model on clinical notes and feature patterns to classify synthetic samples as clinically plausible or implausible.
• Filter out implausible synthetic samples using the GPT-2 classifier to improve data quality and clinical realism.
• Integrate the filtered synthetic data into training pipelines for multiple machine learning classifiers, including XGBoost, Random Forest, Logistic Regression, and SVM.
• Evaluate model performance using repeated 5-fold cross-validation across 25 independent runs to ensure robustness and reliability.

Experimental results

Research questions

• RQ1 Can a data augmentation method specifically designed for binary clinical features improve minority-class recall in early dialysis prediction for CKD patients?
• RQ2 How does BGCS compare to established methods like SMOTE, CTGAN, and standard Gaussian Copula in terms of predictive performance and data fidelity?
• RQ3 To what extent does the integration of a fine-tuned LLM (GPT-2) for filtering synthetic samples enhance the clinical plausibility and utility of augmented data?
• RQ4 What are the most influential clinical features for predicting 90-day dialysis onset when using BGCS-augmented models?
• RQ5 Can BGCS-enabled models be effectively embedded into interpretable decision support systems for clinical use?

Key findings

• BGCS achieved the highest median minority-class recall across all evaluated classifiers, ranging from 0.78 to 0.87 for 90-day dialysis prediction.
• The method demonstrated strong distributional fidelity to real data, with a mean p-value of 0.68 across features in Kolmogorov-Smirnov tests, indicating high similarity to real-world distributions.
• BGCS consistently outperformed SMOTE, CTGAN, and standard Gaussian Copula across all 25 independent runs and all four classifiers in both AUC and recall metrics.
• The best-performing BGCS-augmented model was successfully integrated into an interpretable decision tree-based clinical decision support system.
• Electrolyte imbalances, cardiovascular comorbidities, and renal monitoring indicators were identified as the most influential predictive features in the interpretable model.
• The use of a fine-tuned GPT-2 classifier for filtering synthetic samples significantly improved the clinical plausibility of generated data, reducing noise and spurious patterns.