PRISM is claimed to score disease severity which has attributed an impact on length of PICU stay (LOS).
Primary objective: To determine the impact of PRISM on LOS, and evaluate an Artificial Neural Network's (ANN) performance to estimate LOS from PRISM item patterns.
Research design and methods: Retrospectively we performed correlation and regression analyses on routinely scored PRISM data of all consecutive admissions to our level-III PICU from 1994 to 1999 (n > 2000) with individual LOS. In addition, an ANN was trained on the chronologically first 75% of those data (inputs, PRISM items + age + sex; output, LOS). The ANN's performance was tested on the remaining most recent 25% of the data sets.
Main results: The Spearman and Pearson coefficients of correlation between PRISM and LOS were 0.2 (p < 0.001) and 0.08 (p = 0.0003), the latter being slightly higher when LOS was logarithmically transformed. Pearson's coefficient of correlation between ANN derived LOS estimate and actual LOS was 0.21 (p < 0.001) (LOS logarithmically transformed: 0.34; p < 0.001) in the independent validation sample.
Conclusions: The ANN with its intrinsic ability to detect non-linear correlation, and to relate specific item patterns to LOS, outperformed linear statistics but was still disappointing in estimating individual LOS. It might be speculated that therapeutic intervention modulates the natural course of the disease thus counteracting both disease severity as initially scored by PRISM, and LOS. This being true, the inverse of the correlation between PRISM (or PRISM based LOS estimate) and LOS might be a candidate indicator of quality of care.