The development of a model of prognostication for patients with chronic obstructive pulmonary disease (COPD)

Despite known factors linked to increased mortality in patients with COPD, prognostication is not routinely undertaken. Using hospital administrative datasets, this study aimed to develop a prognostication model for these patients.

A previous study of 30179 admitted COPD patients identified factors linked to death within 6 months using a dataset compiled by hospitals containing information coded at the end of each admission. These factors included: age, sex, admission type, co-morbidities, length of stay and recent hospitalisations. A model was developed by:
 1. Plotting age and length of stay against death
 2. Logistic regression analysis performed for categorical variables to determine ROC curve.
 3. Redundant predictors determined by application of a backwards/forwards removal process
     using a filter to reduce number of variables but maintain the best ROC curve
 4. Cohort sorted into quantiles of estimated probabilities of risk of dying within 6 months,
     comparing with observed risk.

The regression analysis revealed a curve with ROC of 0.784, hence for any given randomly selected pair of patients, the model could predict which patient would die within 6 months in 78% of cases. The BIC filter application resulted in the removal of a number of variables with minimal change in ROC curve (0.782). The remaining variables combined in the model and predictive of death included increased age, male gender, emergency admission, heart or renal failure, dementia, cancer local and metastatic, vascular disease and admissions in previous 6 months. Meanwhile concurrent diagnosis of asthma and mechanical ventilation were protective against death. The difference between the observed and predicted risk in the cohort quantiles ranged between 0.2 and 4.7%.

A model based upon administrative data can be used to predict death within 6 months of admission for patients with COPD with potential for alerting clinicians of risk by a computer generated prompt. Using such a model, clinicians could determine what level of risk for death within 6 months would prompt them to reconsider the approach to patient care, and program the model to alert them automatically when a patient reaches such risk.