LAUSR

Learning Objectives: A major barrier to international research is the lack of a mortality risk score validated across countries. With this in mind, we established a consortium to create a global open-source severity of illness score (GOSSIS). Together we sought to develop a benchmarking model that is robust to international variation, avoiding discrimination and calibration issues suffered by popular US-centric models. The initial version of GOSSIS (v0.1) is developed on data collected from >380,000 intensive care patients admitted to 366 hospitals in 2014–15, across Australia (AU), New Zealand (NZ), and the US. Methods: In GOSSIS v0.1, we used data from the Australia New Zealand Intensive Care Society (ANZICS) and eICU Collaborative Research Datasets, contributing data from AU (225k), NZ (24k) and the US (131k). The cohort included patients >16 years on admission with a ≥4 hour ICU stay. Inclusion also required presence of an Acute Physiology and Chronic Health Evaluation IV (APACHE-IV) score. Hospital mortality was 8.1% in AU/NZ and 9.0% in the USA. Our model comprised minimum and maximum measurements over the first 24h for 36 commonly-available variables, including most APACHE-IV variables, as well as serum lactate, calcium, blood pressures, prothrombin time, platelet count, and hemoglobin. Missing data were imputed using all other available predictors. Continuous variables were modeled using natural cubic splines via logistic regression, while nested random effects were used for diagnoses. Performance was evaluated on held-out data using area under the receiver operator characteristic curve (AUROC) for discrimination and standardized mortality ratios (SMR) for calibration. Results: Despite variation in data sources and quality, GOSSIS achieved an AUROC of 0.914 for prediction of in-hospital mortality. Discrimination was maintained in AU/NZ (0.921) and the US (0.900), surpassing the level achieved by APACHE-IIIj in AU/ NZ (0.902) and APACHE-IVa in the US (0.868). Calibration was excellent in AU/NZ (SMR 1.00 [0.99–1.02]) and the US (0.99 [0.97–1.02]), compared to (0.60 [0.59–0.61] for APACHE-IIIj in AU/NZ and (0.77 [0.75–0.78] for APACHE-IVa in the US. GOSSIS also performed favorably when compared to equivalent models trained and evaluated on a single country. Conclusions: In summary, our results demonstrate potential for GOSSIS as a free, open model for international benchmarking. Future work will use additional data to develop a family of scores covering low and middle-income countries. CCMCritical Care MedicineCrit Care Med0090-3493Lippincott Williams & WilkinsHagerstown, MD CCM