A comparison of statistical and connectionist models for the prediction of chronicity in a surgical intensive care unit

Abstract

Objective: To compare statistical and connectionist models for the prediction of chronicity which is influenced by patient disease and external factors.

Design: Retrospective development of predictive criteria and subsequent prospective testing of the same predictive criteria, using multiple logistic regression and three architecturally distinct neural networks; revision of predictive criteria.

Setting: Surgical intensive care unit (ICU) equipped with a clinical information system in a +/- 1000-bed university hospital.

Patients: Four hundred ninety-one patients with ICU length of stay 3 days who survived at least an additional 4 days.

Interventions: None.

Measurements and main results: Chronicity was defined as a length of stay > 7 days. Neural networks predicted chronicity more reliably than the statistical model regardless of the former's architecture. However, the neural networks' ability to predict this chronicity degraded over time.

Conclusions: Connectionist models may contribute to the prediction of clinical trajectory, including outcome and resource utilization, in surgical ICUs.