Ben Cooper – Project I

Modelling Uncertainty in Surgical Risk Estimation

Surgical procedures are common, with around 5 million operations being conducted per year in the UK, of which 1 million constitute major surgery. The risk of mortality after surgery is predictable and around 80% of postoperative deaths occur in a “higher risk” subset of around 15% of surgical patients. A plethora of risk prediction models have been published but are not widely used, despite national guidance that quantitative mortality risk estimates should be used to guide decision making.  

Depending on the exact nature of the data we are able to access, we intend to address either of the following possible causes for the above mismatch between guidance and clinical practice: 

  • The Models inadequately reflect the uncertainty inherent in clinical practice 

Several logistic regression models have been published which accurately predict mortality risk, yet these models only yield a point estimate of risk, which cannot account for model uncertainty. This uncertainty can arise from limitations of the model training data, or from risk factors which are not known when predicting risk. We intend to use a Bayesian modelling approach to provide a best estimate of individual risk within a range of plausible values. This “credible interval” therefore reflects the uncertainty for a given estimate, providing a clear improvement upon existing models, and allowing us to produce a tool that we hope will be more readily used by clinicians in practice. 

  • Translational problems exist 

Models to date have been based on variables that aren’t always known before surgery (e.g. the intraoperative blood loss) or which aren’t routinely used in the NHS (e.g. a grading system for payment used in the private sector), We intend to validate existing models and recalibrate them for Leeds, as well as “translate” them to using applicable variables that allow the models to be operationalised easily.