CORTEX Complication Recognition Through EHR and eXplainable AI

Background and Rationale Surgical site infection (SSI) is a serious postoperative complication after hip fracture surgery, affecting approximately 3–4% of the 80 000 annual hip fracture patients. It is associated with worse recovery, increased mortality, reoperation, and extended hospital stays, with an estimated cost exceeding £10,000 per case. However, the low statistical incidence of SSI makes randomised trials to reduce infection difficult and expensive to power. Current methods for SSI identification rely on manual follow-up, limiting the feasibility of large-scale evaluations. There is an urgent need for validated, scalable, low-burden approaches to detect SSI across routine NHS systems that can support both surveillance and pragmatic trials. Aim and Design This Programme Development Grant (PDG) will develop and validate methods for detecting SSI using routinely collected NHS data, including both structured and unstructured formats. We will benchmark these methods against gold-standard adjudicated SSI outcomes from the (World Hip Trauma Evaluation)WHiTE 8 trial—a prospective, multicentre study of nearly 5,000 hip fracture patients. Methods At Oxford University Hospitals, we will link WHiTE 8 data with structured NHS data (e.g. International Classification of Diseases (ICD) codes, microbiology, prescribing, labs) to define rule-based phenotypes and evaluate machine learning models (e.g. XGBoost). In parallel, we will apply natural language processing (NLP) to unstructured clinical text—such as discharge summaries and operation notes—training models using ClinicalBERT and similar architectures. Both approaches will be validated against WHiTE 8 outcomes to assess sensitivity, specificity, and predictive value, with subgroup analyses to examine variation by patient and care characteristics. Generalisability and Implementation To assess generalisability, models developed at Oxford will be tested at two additional NHS Trusts involved in WHiTE 8. We will evaluate model transferability, local documentation variation, and feasibility for NHS deployment. Real-world constraints such as compute resources and data flows will be assessed. Stakeholder Engagement Patients, carers, clinicians, NHS digital leads, and public health agencies (e.g. UKHSA) will be engaged through structured co-design workshops. These will inform ethical and governance frameworks, define acceptable outputs, and explore issues of transparency, consent, and trust. The NIHR Musculoskeletal Trauma PPI Group will co-lead this work to ensure diverse perspectives are represented. Outputs and Next Steps This PDG will produce validated, explainable SSI detection tools, phenotyping logic, and technical guidance. These outputs will be used to design a full Programme Grant for Applied Research (PGfAR), delivering a suite of embedded randomised feasibility trials within the established multicentre WHiTE trial platform. The trials will evaluate candidate infection prevention interventions (e.g. extended antibiotics) and test the use of routine data pipelines for outcome capture. Impact and Relevance This project directly addresses a James Lind Alliance top 10 research priority and supports NIHR and NHS aims to enable efficient, data-enabled trials and improve patient safety. By enabling detection of SSI using routine data, it will support more inclusive research, reduce participant burden, and enable cost effective interventional studies to reduce this devastating complication.