Supplementary Material for: Machine-learning-aided decision-making model for the discontinuation of continuous renal replacement therapy

ABSTRACT
BACKGROUND. Continuous renal replacement therapy (CRRT) is a primary form of renal support for patients with acute kidney injury in an intensive care unit. Making an accurate decision of discontinuation is crucial for the prognosis of patients. Previous research has mostly focused on the univariate and multivariate analysis of factors in CRRT, without the capacity to capture the complexity of the decision-making process. The present study thus developed a dynamic, interpretable decision model for CRRT discontinuation.
METHODS. The study adopted a cohort of 1234 adult patients admitted to an intensive care unit in the MIMIC-IV database. We used the extreme gradient boosting (XGBoost) machine learning algorithm to construct dynamic discontinuation decision models across four time points. Shapley additive explanation (SHAP) analysis was conducted to show the contribution of an individual feature to the model output.
RESULTS. Of the 1234 included patients with CRRT, 596 (48.3%) successfully discontinued CRRT. The dynamic prediction by the XGBoost model produced an area under the curve of 0.848 and accuracy, sensitivity, and specificity of 0.782, 0.786, and 0.776, respectively. The XGBoost model was thus far superior to other test models. SHAP demonstrated that the features that contributed most to the model results were the sequential organ failure assessment score, serum lactate level, and 24-hour urine output.
CONCLUSIONS. Dynamic decision models supported by machine learning are capable of dealing with complex factors in CRRT and effectively predicting the outcome of discontinuation.
KEYWORDS: acute kidney injury, continuous renal replacement therapy, discontinuation, machine learning

摘要
## 背景
连续性肾脏替代治疗（continuous renal replacement therapy, CRRT）是重症监护病房（intensive care unit, ICU）急性肾损伤（acute kidney injury, AKI）患者的主要肾脏支持手段。精准制定CRRT停疗决策对患者预后至关重要。既往研究多聚焦于CRRT相关影响因素的单因素、多因素分析，无法捕捉决策过程的复杂性。因此，本研究构建了一款可解释的动态CRRT停疗决策模型。

## 方法
本研究纳入MIMIC-IV数据库中1234名成年重症监护病房住院CRRT患者。采用极限梯度提升（extreme gradient boosting, XGBoost）机器学习算法，构建四个时间节点的动态停疗决策模型。通过Shapley可加解释（Shapley additive explanation, SHAP）分析，展示单个特征对模型输出的贡献度。

## 结果
在纳入的1234名CRRT患者中，596名（48.3%）成功终止CRRT治疗。XGBoost模型的动态预测曲线下面积为0.848，准确率、灵敏度、特异度分别为0.782、0.786和0.776，显著优于其他测试模型。SHAP分析显示，对模型结果贡献度最高的特征依次为序贯器官衰竭评估评分、血清乳酸水平及24小时尿量。

## 结论
基于机器学习的动态决策模型能够处理CRRT治疗中的复杂影响因素，有效预测停疗结局。

## 关键词
急性肾损伤，连续性肾脏替代治疗，停疗，机器学习