Identification of patient clusters using APACHE II parameters, and their correlation to outcomes

Introduction: Our main hypothesis is the identification of specific patient subgroups, using multivariate statistical analysis of APACHE II parameters, will allow improved clinical management of these patients in the ICU. We plan to test the hypothesis that ICU patient populations are formed of 'patient clusters' sharing important properties. We attempt to show that the identification of these clusters will yield improved outcome predictions compared to using the entire patient cohort. Improving the correlation between patient-specific outcomes and patient clusters may provide critical care clinicians with more powerful tools for the clinical management of their patients. Methods: 1,282 complete records of all patients admitted to a tertiary hospital medical ICU over a 9 month period were analyzed with respect to admission APACHE II parameters, verified diagnosis, and outcomes. CART (i.e., a nonparametric, tree-based statistical model) was used to construct decision trees to further stratify patients into homogeneous subgroups. Results were validated using independent data. Results: CART analyses identified homogeneous subgroups of patients using subsets of the 15 Apache II parameters. In one analysis, 19 patients with arterial pH ≤ 7.25 and Glasgow coma scale (GCS) ≤ 7.5, and 32 patients with arterial pH ≤ 7.25 and GCS from 7.5-13.5 were identified. All 19 patients in the first group died and 20 of the 32 patients in the second group died which is statistically different (p = 0.0017). Conclusion: Specific patient subgroups, identified using multivariate statistical analysis of APACHE II parameters, can be correlated to patient outcomes. These preliminary data suggest that cluster analysis may improve upon the identification of patients receiving potentially futile care as well as those likely to survive their ICU stay.