An improved medical admissions risk system using multivariable fractional polynomial logistic regression modelling

Abstract

Aim: To develop and validate an in-hospital mortality risk prediction tool for unselected acutely ill medical patients using routinely collected physiological and laboratory data. Design: Analysis of all emergency medical patients admitted to St James's Hospital (SJH), Dublin, between 1 January 2002 and 31 December 2007. Validation using a dataset of acute medical admissions from Nenagh Hospital 2000–04. Methods: Using routinely collected vital signs and laboratory findings, a composite 5-day in-hospital mortality risk score, designated medical admissions risk system (MARS), was developed using an iterative approach involving logistic regression and multivariable fractional polynomials. Results are presented as area under receiver operating characteristics curves (AUROC) as well as Hosmer and Lemeshow goodness-of-fit statistics. Results: A total of 10 712 and 3597 unique patients were admitted to SJH and Nenagh Hospital, respectively. The final score included nine variables [age, heart rate, mean arterial pressure, respiratory rate, temperature, urea, potassium (K), haematocrit and white cell count]. The AUROC for 5-day in-hospital mortality was 0.93 [95% confidence interval (CI) 0.92–0.94] for the SJH cohort (Hosmer and Lemeshow test, P = 0.32) and 0.92 (95% CI 0.90–0.94) for the external Nenagh hospital validation cohort (Hosmer and Lemeshow test, P = 0.28). Conclusions: In-hospital mortality estimation using only routinely collected emergency department admission data is possible in unselected acute medical patients using the MARS system. Such a score applied to acute medical patients at the time of admission, could assist senior clinical decision makers in promptly and accurately focusing limited clinical resources. Further studies validating the impact of this model on clinical outcomes are warranted.