PhD Studentship: Using data to predict risk of hospital-acquired acute kidney injury (HA-AKI)

Background

Acute kidney injury (AKI) is where kidneys suddenly stop working properly. It can range from minor loss of kidney function to complete kidney failure. Without quick treatment, abnormal levels of salts and chemicals can build up in the body, which affects the ability of other organs to work properly. According to UK renal registry, the overall AKI episode rate was over 12,000 per million population. Among the AKI events, 30-33% are hospital-acquired acute kidney injury (HA-AKI). The National Institute for Health and Care Excellence reported that the estimated annual cost of HA-AKI ranges from £430 million to over £1 billion to the NHS. In the US, the EPIC risk of HA-AKI model is a machine learning tool developed within the EPIC systems to predict the risk of HA AKI in hospitalised patients.

However, the model was trained on US populations, and may not be directly applicable to the UK context due to differences in patient demographics, clinical practice, and data infrastructure. Recent external validation of the EPIC HA-AKI model reported moderate predictive performance, and it predicts more accurately low risk but less so in high risk patients. These limitations highlight the need for a HA-AKI risk prediction model trained on UK population. This project aims to develop and validate a HA-AKI prediction model tailored to the UK population using modern statistical and machine learning approaches.

Objectives and methodology

1. Data curation from electronic health records. We will curate large-scale, analysis-ready datasets from NHS electronic health records. HA-AKI outcomes will be defined using validated medicine code lists. Candidate predictors will include demographics, pre-existing comorbidities, treatment history, laboratory biomarkers, and prescribed medications. We will develop preprocessing pipelines to ensure data quality, reproducibility, and clinical relevance.
2. Model development. We will build prediction models using information available at the time of hospital admission. A range of statistical and machine learning methods will be explored, including logistic regression and penalised regression models, tree-based methods, random forests, and gradient boosting. Feature selection strategies will be applied to reduce dimensionality and enhance clinical interpretability.
3. Model validation. Internal validation will be conducted using cross-validation and bootstrapping. External validation will be performed across different NHS sites to assess Engineering and Physical Sciences Research Council Doctoral Landscape Award generalisability. Model performance will be evaluated in terms of discrimination (ability to classify patients into outcome categories) and calibration (agreement between predicted and observed risk). We will conduct decision curve analysis to further assess clinical utility.
4. Model implementation and accessibility. Based on the results of the developed and validated models, we will create an accessible tool using open-source software, to estimate individual risk for HA-AKI. We will conduct stakeholder engagement activities to guide usability within clinical workflows. Open science and impact. All analyses will adhere to TRIPOD guidelines for prediction model research, with reproducible pipelines and open-source software. The project will develop a HA AKI risk prediction model tailored to the UK population, with the potential to improve early detection and patient outcomes.

Funded by: EPSRC Engineering and Physical Sciences Research Council