Scholarlink

Background <p>Acute kidney injury (AKI) is common in intensive care units (ICUs) and often leads to kidney replacement therapy (KRT). Precise prediction of KRT requirements could improve resource allocation and patient outcomes. We evaluated multiple machine learning (ML) models for their ability to predict KRT in critically ill patients with AKI, comparing their performance with experienced clinicians.</p> Methods <p>We conducted a single-center, retrospective study at Mohammed VI University Hospital, including adult patients (≥ 18&#xa0;years) admitted between April 2023 and December 2024 who developed AKI according to Kidney Disease: Improving Global Outcomes (KDIGO) criteria. Five ML classifiers (random forest, logistic regression, extreme gradient boosting, <i>K</i>-nearest neighbors, and artificial neural network) were trained on 80% of the dataset and tested on the remaining 20%. A panel of five intensivists, blinded to model outputs, independently predicted the need for KRT in the test cohort. Model performance and the clinician consensus were assessed via accuracy, recall, precision, F1-score, F1.5-score, and area under the receiver operating characteristic curve (AUROC).</p> Results <p>Of 2148 screened ICU admissions, 425 patients with AKI were included (22.1% required KRT). Random forest demonstrated the best balance of metrics (accuracy = 0.882; AUROC = 0.931), tied with XGBoost. The clinician consensus performed well (AUROC = 0.898; recall = 0.947; precision = 0.562), slightly below RF. Shapley additive explanations (SHAP) analysis identified serum creatinine, potassium, and blood urea nitrogen among the strongest predictors. Subgroup analysis showed no major bias by age or sex.</p> Conclusions <p>Machine learning models, particularly random forest, reliably predicted KRT needs in critically ill AKI patients and slightly exceeded experienced clinicians’ performance. Integrating ML outputs with clinical judgment may enhance timely decision-making, especially in resource-limited ICUs, yet external prospective validation is essential.</p> Registration <p>Registered on the Open Science Framework (<a href="https://doi.org/10.17605/OSF.IO/YBQ5C">https://doi.org/10.17605/OSF.IO/YBQ5C</a>), including protocol and analytical code.</p> Graphical abstract <p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Predicting kidney replacement therapy in critically ill patients: machine learning models versus clinician judgment

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年10月16日

查看全文
  • Ounci Es-Saad,
  • Mohamed Berrada,
  • Amine Elhasnaoui,
  • Karim Lakhdar,
  • Choukri Bahouh,
  • Youssef Motiaa,
  • Rajae Alkouh,
  • Adil Najdi,
  • Smael Labib

摘要

Background

Acute kidney injury (AKI) is common in intensive care units (ICUs) and often leads to kidney replacement therapy (KRT). Precise prediction of KRT requirements could improve resource allocation and patient outcomes. We evaluated multiple machine learning (ML) models for their ability to predict KRT in critically ill patients with AKI, comparing their performance with experienced clinicians.

Methods

We conducted a single-center, retrospective study at Mohammed VI University Hospital, including adult patients (≥ 18 years) admitted between April 2023 and December 2024 who developed AKI according to Kidney Disease: Improving Global Outcomes (KDIGO) criteria. Five ML classifiers (random forest, logistic regression, extreme gradient boosting, K-nearest neighbors, and artificial neural network) were trained on 80% of the dataset and tested on the remaining 20%. A panel of five intensivists, blinded to model outputs, independently predicted the need for KRT in the test cohort. Model performance and the clinician consensus were assessed via accuracy, recall, precision, F1-score, F1.5-score, and area under the receiver operating characteristic curve (AUROC).

Results

Of 2148 screened ICU admissions, 425 patients with AKI were included (22.1% required KRT). Random forest demonstrated the best balance of metrics (accuracy = 0.882; AUROC = 0.931), tied with XGBoost. The clinician consensus performed well (AUROC = 0.898; recall = 0.947; precision = 0.562), slightly below RF. Shapley additive explanations (SHAP) analysis identified serum creatinine, potassium, and blood urea nitrogen among the strongest predictors. Subgroup analysis showed no major bias by age or sex.

Conclusions

Machine learning models, particularly random forest, reliably predicted KRT needs in critically ill AKI patients and slightly exceeded experienced clinicians’ performance. Integrating ML outputs with clinical judgment may enhance timely decision-making, especially in resource-limited ICUs, yet external prospective validation is essential.

Registration

Registered on the Open Science Framework (https://doi.org/10.17605/OSF.IO/YBQ5C), including protocol and analytical code.

Graphical abstract