Acute kidney injury (AKI) is a prevalent complication among patients in the Intensive Care Unit (ICU), affecting up to 50% of admitted individuals. The condition is characterized by a sudden spike in serum creatinine and a decrease in urine volume. AKI has been linked to adverse outcomes, including acute kidney disease (AKD), chronic kidney disease (CKD), and mortality. Despite advances in AKI treatment, mortality rates remain high, particularly in severe stages, and some survivors may require long-term renal replacement therapy.
To improve outcomes, it is essential to develop prediction models that can assess patient risks and identify subgroups that require follow-up care. Conventional approaches, such as Logistic Regression, have been used to develop prediction models for AKI outcomes. However, machine learning algorithms have become increasingly popular in healthcare settings due to their ability to handle large quantities of high-dimensional data found in electronic health records (EHRs). These models can capture complex interactions between features in datasets, providing more accurate and reliable prediction models.
In a study published in Scientific Reports, researchers developed machine learning-based prediction models to assess outcomes following AKI stage 3 events in ICU patients. The team used medical records to develop two models that could predict patients at risk of developing CKD after three and six months after experiencing AKI stage 3. They also developed two survival prediction models using random survival forests and survival XGBoost to assess the risk of patient mortality.
The team evaluated established CKD prediction models using AUCROC and AUPR curves and compared them with baseline logistic regression models. They also evaluated mortality prediction models with an external test set and compared their C-indices to baseline COXPH. The researchers included 101 critically ill patients who experienced AKI stage 3, and they expanded the training set for the mortality prediction task by incorporating an unlabeled dataset.
The study found that the models developed using the random forest algorithm and XGBoost outperformed the baseline models in predicting CKD and mortality, respectively, with an AUPR of 0.895 and 0.848 and a c-index of 0.8248. The study also highlighted the benefits of incorporating unlabeled data into survival analysis tasks.
The authors note that while there have been advances in leveraging EHR data for predictive modeling, there is still room for further exploration and development of machine learning models that fully tap into the abundance and diversity of EHR-derived data.
The study’s findings have significant implications for patient care and outcomes. By developing predictive models that can identify patients at risk for negative outcomes, clinicians can make informed decisions and provide timely interventions. The study provides valuable insights into the long-term prognosis of critically ill patients, enhancing clinical decision-making, and ultimately improving patient outcomes.
