An optimization of the Waterlow score using regression and artificial neural networks

Abstract

Objectives: To optimize the ability of the Waterlow Scale to predict individuals vulnerable to developing pressure ulcers. Design: Prospective cohort study. Setting: Two acute care UK National Health Service (NHS) providers. Subjects: Four hundred and twenty-two inpatients across five specialities (general medicine, general surgery, orthopaedics, oncology and rehabilitation). Interventions: Waterlow scores recorded weekly for 14 days post admission to hospital. Main outcome measure: Development of a pressure ulcer. Results: Nonlinear analysis using neural networks did not outperform linear methods. Only five items out of 11 in the Waterlow Scale appeared to have any classification ability in this patient population. Conclusions: The Waterlow score when modelled as a linear equation appears as effective as more complicated nonlinear mappings using neural networks. Only a subset of the variables of the Waterlow Scale have predictive value in this patient population, but this is a different subset to those found in a previous study of a different client group (wheelchair users).