Can valid and practical risk-prediction or casemix adjustment models, including adjustment for comorbidity, be generated from English hospital administrative data (Hospital Episode Statistics)? A national observational study

Abstract Background NHS hospitals collect a wealth of administrative data covering accident and emergency (A&E) department attendances, inpatient and day case activity, and outpatient appointments. Such data are increasingly being used to compare units and services, but adjusting for risk is difficult. Objectives To derive robust risk-adjustment models for various patient groups, including those admitted for heart failure (HF), acute myocardial infarction, colorectal and orthopaedic surgery, and outcomes adjusting for available patient factors such as comorbidity, using England’s Hospital Episode Statistics (HES) data. To assess if more sophisticated statistical methods based on machine learning such as artificial neural networks (ANNs) outperform traditional logistic regression (LR) for risk prediction. To update and assess for the NHS the Charlson index for comorbidity. To assess the usefulness of outpatient data for these models. Main outcome measures Mortality, readmission, return to theatre, outpatient non-attendance. For HF patients we considered various readmission measures such as diagnosis-specific and total within a year. Methods We systematically reviewed studies comparing two or more comorbidity indices. Logistic regression, ANNs, support vector machines and random forests were compared for mortality and readmission. Models were assessed using discrimination and calibration statistics. Competing risks proportional hazards regression and various count models were used for future admissions and bed-days. Results Our systematic review and empirical analysis suggested that for general purposes comorbidity is currently best described by the set of 30 Elixhauser comorbidities plus dementia. Model discrimination was often high for mortality and poor, or at best moderate, for other outcomes, for example c = 0.62 for readmission and c = 0.73 for death following stroke. Calibration was often good for procedure groups but poorer for diagnosis groups, with overprediction of low risk a common cause. The machine learning methods we investigated offered little beyond LR for their greater complexity and implementation difficulties. For HF, some patient-level predictors differed by primary diagnosis of readmission but not by length of follow-up. Prior non-attendance at outpatient appointments was a useful, strong predictor of readmission. Hospital-level readmission rates for HF did not correlate with readmission rates for non-HF; hospital performance on national audit process measures largely correlated only with HF readmission rates. Conclusions Many practical risk-prediction or casemix adjustment models can be generated from HES data using LR, though an extra step is often required for accurate calibration. Including outpatient data in readmission models is useful. The three machine learning methods we assessed added little with these data. Readmission rates for HF patients should be divided by diagnosis on readmission when used for quality improvement. Future work As HES data continue to develop and improve in scope and accuracy, they can be used more, for instance A&E records. The return to theatre metric appears promising and could be extended to other index procedures and specialties. While our data did not warrant the testing of a larger number of machine learning methods, databases augmented with physiological and pathology information, for example, might benefit from methods such as boosted trees. Finally, one could apply the HF readmissions analysis to other chronic conditions. Fundin