|| Checking for direct PDF access through Ovid
Introduction: Comparing outcomes between hospitals for quality improvement needs risk adjustment for the severity of the stroke. The NIHSS is the standard instrument for this but has difficulties in collection and integration into administrative and other electronic databases. A proxy derived from existing data items would be useful.Methods: We derived a proxy using the 2014 national hospital administrative data for England and the 2012 US Nationwide Inpatient Sample. Diagnosis codes and procedures (performed on admission day) with a prevalence in ischaemic stroke admissions of at least 1% were tabulated and their potential for inclusion into a proxy independently assessed by members of a physician panel. Classification and regression trees helped inform the development of a logistic regression model that was used to finalise and evaluate the performance of the proxy when predicting in-hospital mortality. The proxy was also compared against the actual NIHSS in a sample of 282 US administrative records and 265 English records linked to those patients’ NIHSS from a separate study.Results: For England, there were 69,125 admissions and 9,130 in-hospital deaths; for the NIS there were 86,478 admissions and 3,795 in-hospital deaths. The physician panel agreed on 19 candidate variables, which included intubation, aspiration pneumonitis, dysphasia and aphasia, coma, and others. The models in England and the US retained most variables. The relative importance of each variable was similar in each database, with intubation, aspiration pneumonitis, and coma the most important. The c statistic for discrimination was 0.82 for the NIS and 0.69 for England; adding age, sex and admission source gave 0.84 for NIS and 0.78 for England. The adjusted odds ratio per unit increase in the proxy in the NIS was 1.16 (p<0.001). The correlation between our proxy and the actual NIHSS in our separate sample was 0.52 in England and 0.47 in the US.Conclusions: An NIHSS proxy can be derived using administrative data with either ICD9 or ICD10 and has reasonable prediction for in-hospital mortality.