Canadian in-hospital mortality for patients with emergency-sensitive conditions: a retrospective cohort study

Abstract

BACKGROUND: The emergency department (ED) sensitive hospital standardized mortality ratio (ED-HSMR) measures risk-adjusted mortality for patients admitted to hospital with conditions for which ED care may improve health outcomes. This study aimed to describe in-hospital mortality across Canadian provinces using the ED-HSMR. METHODS: Hospital discharge data were analyzed from April 2009 to March 2012. The ED-HSMR was calculated as the ratio of observed deaths among patients with emergency-sensitive conditions in a hospital during a year (2010-11 or 2011-12) to the expected deaths for the same patients during the reference year (2009-10), multiplied by 100. The expected deaths were estimated using predictive models fitted from the reference year. Aggregated provincial ED-HSMR values were calculated. A HSMR value above or below 100 respectively means that more or fewer deaths than expected occurred within a province. RESULTS: During the study period, 1,335,379 patients were admitted to hospital in Canada with an emergency-sensitive condition as the most responsible diagnosis. More in-hospital deaths (95% confidence interval) than expected were respectively observed for the years 2010-11 and 2011-12 in Newfoundland [124.3 (116.3-132.6); & 117.6 (110.1-125.5)] and Nova Scotia [116.4 (110.7-122.5) & 108.7 (103.0-114.5)], while mortality was as expected in Prince Edward Island [99.9 (86.5-114.8) & 100.7 (87.5-115.3)] and Manitoba [99.2 (94.5-104.1) & 98.3 (93.5-103.3)], and less than expected in all other provinces and territories. CONCLUSIONS: Our study revealed important variation in risk-adjusted mortality for patients admitted to hospital with emergency-sensitive conditions among Canadian provinces. The ED-HSMR may be a useful outcome indicator to complement existing process indicators in measuring ED performance.

Machine ​learning algorithms for claims data‐based prediction of in‐hospital mortality in patients with heart failure

“…The three development data subsets were evaluated using five algorithms: logistic regression [generalized linear models (GLM)], random forest (RF), gradient boosting machine (GBM), single‐layer neural network (NNET), and extreme gradient boosting (XGBoost). Even though ML algorithms can implicitly take into account interactions between variables and non‐linearity, we decided not to specify interactions or non‐linear effects in the GLM for several reasons: (i) to keep the model specification as simple as possible for comprehensibility, (ii) to reflect the approaches already existing in the literature, and (iii) because non‐linearity was not expected from a clinical perspective 16,17 . The algorithms were tuned using a grid search with a k‐fold approach, using three repetitions of 10‐fold each.…”