Mortality prediction and acuity assessment in critical care

Accurate mortality prediction in intensive care units (ICUs) allows for the risk adjustment of study populations, aids in patient care and provides a method for benchmarking overall hospital and ICU performance. ICU risk-adjustment models are primarily comprised of an integer severity of illness score which increases with increasing patient risk of mortality. First published in the 1980s, the improvements to these scores primarily consisted of increasing the dimensionality of the model, and hence also increasing their complexity. This thesis aims to improve upon these models. First, the field is surveyed and the major models for risk-adjusting critically ill patient cohorts are identified including the acute physiology score (APS) and the simplified acute physiology score (SAPS). A key component of model performance is data preprocessing. The effect of preprocessing ICU data is quantified on a dataset of 8,000 ICU patients, and it is shown that after preprocessing to remove extreme values a logistic regression (LR) model performed competitively (AUROC of 0.8633) with the more complex machine learning model; a support vector machine (SVM) which had an AUROC of 0.8653. For validation, model development was repeated in a larger database containing over 80,000 patients admitted to 89 ICUs in the United States. Results were similar (AUROC of 0.8895 for the LR vs 0.8917 for the SVM) but showed the performance gain when using automated outlier rejection is less pronounced in well quality controlled datasets (0.8883 for LR without rejection). It is hypothesised from this that simpler models can perform competitively with more complicated models, while having a greatly reduced burden of data collection. A severity score is developed on the large multi-center database using a Genetic Algorithm and Particle Swarm Optimisation. The severity score, named the Oxford Acute Severity of Illness Score (OASIS), is shown to outperform the APS III (AUROC 0.837 vs 0.822) and perform competitively with APACHE IV when used as a covariate in a regression model (AUROC 0.868 vs 0.881). The severity score requires only 10 variables (58% as many as APS III), reducing the burden of quality control and data collection. These variables are routinely collected in critical care by continuous monitors and do not include comorbidities, diagnosis or laboratory measurements. The severity score is then externally evaluated in an American hospital and shown to discriminate well (AUROC 0.790 vs. 0.782 for the APS III) with excellent calibration. Finally, the severity score was evaluated in an English hospital and compared to other severity scores. OASIS again had excellent calibration and discrimination (AUROC 0.776 vs 0.750 for APS III) whilst requiring a much smaller number of variables. OASIS has many applications, including both simplifying data collection for studies and improving the risk assessment therein.