More than one-half million babies (12.5%) were born prematurely in the US in 2004. Virtually all infants born  27 weeks’ gestation, ~80% born between 27–30 weeks’ gestation, and 30% born between 30–32 weeks’ gestation require endotracheal intubation and intermittent positive pressure ventilation starting soon after birth1. Although ventilator management of premature infants has improved, with advanced technologies available, there are still no standards for weaning and determining optimal extubation time for those infants. Approximately, 30% of preterm infants will fail extubation, require reintubation and continued mechanical ventilation2. Failed extubation may result in a significant setback for the infant, resulting in requirements for increased support compared to the level required before extubation. Also, the risk of developing co-morbidities may increase with the clinical deterioration. However, decision-making whether or not to extubate a premature infant is extremely complex, involving a large amount of information that is, in part, processed subconsciously by the clinician and based largely on clinical experience. A decision-support tool aiding clinicians in their decision-making would be invaluable during the care and treatment of premature infants in the NICU for the purpose of informing inexperienced clinicians about infants potentially ready for extubation, particularly during times when the NICU is overcrowded and extremely busy. Such a decision-support tool can assist in decreasing the number of false-positive (i.e. infants that were extubated too early) and false-negative (infants that could have been extubated earlier) cases. Most attempts to develop a decision-support tool rely on the use of a single method to determine the most appropriate decision. Since the decision whether or not to extubate a given infant is a complex process, and preliminary work has shown promising, but not yet satisfactory, results with a single method3, the development of a committee of several different methods is proposed. With this committee of several different algorithms that combines the individual predictions into a single prediction, improved decision-making is expected, but must be tested. Through implementation of a decision-support tool distributed through the Internet, the user (clinicians or researchers) will work solely with an interactive easy-to-use website for entering data and/or retrieving predictions for extubation outcome, thus, leaving all software maintenance and updating to the decision-support tool providers. With the deployment of semantic technologies such as XML and RDF, which provide a foundation for interoperability, data can be collected in an independent, distributed fashion and can be shared and analyzed over the Internet facilitating and promoting interaction between researchers4. Furthermore, extensions of these semantic technologies such as query languages (SPARQL) and web ontology language (OWL) have recently become available, which highlight the fact that a supporting R&D community has emerged and mission-critical applications based on that technology are at hand. These language extensions provide reliable access to the data through query and analysis features and allow more sophisticated knowledge representation of and inference from the data. First, we developed an ANN prototype that was enthusiastically received by the scientific community. However, an editorial published alongside the original appearance of these data stated that the need for well-designed studies to assess the effects and cost-effectiveness of clinical decision-support systems are critical but are not eagerly adopted in the clinics due to concerns about sufficient validation5. Sufficient validation not only means that these systems need to be validated with data sets independent of those used for training, but also that machine-learning approaches need to be compared with equally powerful machine-learning methods with different characteristics. This requirement led to the process flow for the development of a decision-support tool to predict extubation outcome with the following four important phases: 1) variable “discovery” carried out in our preliminary studies by extracting knowledge from a group of experts; 2) sound method development using state– of–the-art machine-learning algorithms combined in a powerful committee; followed by 3) method verification that will include preliminary testing of the methods and the feasibility of a study involving these methods in the NICU, as well as obtaining estimates of the parameters required to calculate the appropriate sample size; and finally 4) definitive validation of the clinical effectiveness of the proposed clinical decision-support tool that will require a multi-center randomized clinical trial with a clustered design to ensure a sufficiently large sample size. This sequence of studies is analogous to that used for development of clinically useful biochemical biomarkers6.

1)    Hamilton B, Martin J, Ventura S, Sutton D, Menacker F. Births: Preliminary Data for 2004.     
       National Vital Statistics Reports 2005;54(8)
2)    Kavvadia V, Greenough A, Dimitriou G. Prediction of extubation failure in preterm neonates. 
       Eur J Pediatr 2000;159(4):227-31.
3)    Ennett CM, Frize M, Walker CR. Influence of missing values on artificial neural network 
       performance. Medinfo 2001;10(Pt 1):449-53.
4)    MIT, ERCIM, Keio. World Wide Web Consortium (W3C). In; 2007.
5)    Walker CR, Frize M. Are artificial neural networks "ready to use" for decision making in the 
       neonatal intensive care unit? Commentary on the article by Mueller et al. and page 11.    
       Pediatr Res 2004;56(1):6-8.
6)    Rifai N, Gillette MA, Carr SA. Protein biomarker discovery and validation: the long and uncertain
       path to clinical utility. Nat Biotechnol 2006;24(8):971-83.