Abstract
Continuous renal replacement therapy (CRRT) is a form of dialysis prescribed to severely ill patients who cannot tolerate regular hemodialysis. However, as the patients are typically very ill to begin with, there is always uncertainty whether they will survive during or after CRRT treatment. Because of outcome uncertainty, a large percentage of patients treated with CRRT do not survive, utilizing scarce resources and raising false hope in patients and their families. To address these issues, we present a machine learning-based algorithm to predict short-term survival in patients being initiated on CRRT. We use information extracted from electronic health records from patients who were placed on CRRT at multiple institutions to train a model that predicts CRRT survival outcome; on a held-out test set, the model achieves an area under the receiver operating curve of 0.848 (CI = 0.822–0.870). Feature importance, error, and subgroup analyses provide insight into bias and relevant features for model prediction. Overall, we demonstrate the potential for predictive machine learning models to assist clinicians in alleviating the uncertainty of CRRT patient survival outcomes, with opportunities for future improvement through further data collection and advanced modeling.
Introduction
Continuous Renal Replacement Therapy (CRRT) is a specialized, resource-intensive form of dialysis used for critically ill patients who cannot tolerate standard hemodialysis. Despite its role as a life-sustaining intervention, approximately 50% of patients initiated on CRRT do not survive. Currently, the decision to start CRRT is often subjective, relying on individual physician assessment rather than standardized guidelines. This uncertainty can lead to futile care, exhausting medical resources and providing false hope to families. This study presents a data-driven machine learning framework designed to predict short-term survival at the point of CRRT initiation, offering a potential tool to assist in these difficult clinical decisions.
Methodology and Data
The researchers developed a machine learning pipeline utilizing a longitudinal dataset of 8,986 patients across two major quaternary care hospital systems: UCLA and Cedars-Sinai. The data included static demographic features and longitudinal records such as laboratory results, vital signs, and medications.
Key technical components of the pipeline include:
- Windowing: A "look-back window" was used to aggregate clinical features over several days prior to CRRT initiation, capturing the patient's trajectory.
- Model Selection: The team compared several algorithms, including Logistic Regression, Random Forest, LightGBM, and XGBoost.
- Explainability: SHAP (Shapley Additive Explanations) values were employed to identify which clinical features most heavily influenced the model's predictions, ensuring the logic aligned with clinical intuition.
Key Results
The comprehensive model, trained on multi-institutional data, demonstrated strong discriminative power and calibration on a held-out test set:
- ROCAUC: 0.848 (CI = 0.822–0.870), indicating high accuracy in distinguishing survivors from non-survivors.
- Precision-Recall AUC (PRAUC): 0.777.
- Calibration: A Brier score of 0.155 confirmed that the model's predicted probabilities closely matched observed outcomes.
Analysis revealed a "generalization gap," where models trained at one institution performed poorly at another due to domain shift in data collection. However, models trained on combined datasets proved robust. Notably, SHAP analysis identified "surgery before CRRT" as the strongest predictor of mortality, validating clinical observations that surgical patients requiring CRRT often face poorer outcomes.
Discussion and Limitations
While the model successfully predicts outcomes at "Day Zero"—the most critical window for decision-making—it is not yet intended for direct clinical use. The study's retrospective nature means survival labels may be influenced by prior transitions to comfort care. Furthermore, data heterogeneity between hospitals remains a challenge for universal deployment. Future work will involve prospective studies to evaluate how clinicians interact with the tool and whether it improves real-time resource allocation and patient outcomes.
Summary updated