Interpretable machine learning prediction models for 28-day mortality in critically ill patients with atrial fibrillation and acute kidney injury

Comprehensive Summary

Atrial fibrillation (AF) and acute kidney injury (AKI) are two commonly co-occurring complications in ICU patients and amplifies mortality risk by disturbing cardiovascular and renal stability. Both AF and AKI have been shown to increase mortality individually, but limited research has been conducted on ICU patients with both complications. In this study, Gao et al. developed and validated machine-learning (ML) prognostic models for this high-risk population. Using data from over 14,000 patients drawn from two large U.S. ICU databases (MIMIC-IV and eICU-CRD), the researchers compared nine ML algorithms. A Gradient Boosting Machine (GBM) model was found to perform best with AUCs of 0.856 and 0.761 for internal and external validation respectively. Moreover, the GBM model outperformed widely used SOFA scores. SHAP analysis on the model showed the three most important predictors of mortality were max anion gap, mean heart rate, and age. Additionally, the authors developed a free, online risk calculator for individual risk assessment at the bedside. The authors acknowledge key limitations such as retrospective study design, reliance on electronic health records, inability to differentiate between chronic and acute AF, incomplete data on certain treatments, and solely U.S-based datasets which could limit generalizability. Overall, this study is one of the first externally validated ML prognostic models for ICU patients with concurrent AF and AKI, with potential to improve individualized risk assessment and clinical decision-making in this high-risk population.

Outcomes and Implications

This study advances the case for utilizing ML-models in the ICU for decision making support. Gao et al. demonstrate that interpretable, externally validated models specifically trained for high-risk subgroups can outperform conventional severity scores, such as SOFA, and provide clinicians with information needed to make data-driven clinical decisions. The development of a free, online risk calculator is a step towards real-world implementation, where clinicians can use this tool to determine who needs closer monitoring, direct resource allocation, and have informed conversations with patients and families regarding prognosis.The GBM model’s performance over the standard SOFA score also highlights the need to move away from one-size-fits-all prognostic approaches for complex populations with multiple comorbidities – an area where continued research on ML prognostic models may prove greatly beneficial.