Causal networks for modeling health technology utilization in intensive care units


This study presents the application of Bayesian networks (Bn) to explain Neonatal Intensive Care Unit relationships. Information was compiled retrospectively from the medical records at two neonatal intensive care units of 523 neonates (63 deaths). A total of 31 variables were used for the model, eleven to characterize admission conditions and severity of illness as well as the 20 technologies. With mortality as the output variable, the K2 search algorithm and Geiger-Heckerman quality measures were used in the training that generated the Bn. Evidence propagation was used to assess the training, which yielded a sensitivity of 77.78% and a specificity of 91.30%, in the classification of mortality. Clinical criteria, correlations and logistical regression were used to analyse the relationships the model provided. The Bn found clinically coherent relationships as recognizable conditions that directly affect mortality such as congenital malformations are seen and it exposes the least effective technologies among those studied, bicarbonate treatment. © Springer-Verlag : 2004.

Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Max Chacón
Max Chacón
Full Professor