Expert system supporting an early prediction of the bronchopulmonary dysplasia

This work presents a decision support system which uses machine learning to support early prediction of bronchopulmonary dysplasia (BPD) for extremely premature infants after their first week of life. For that purpose a knowledge database was created based on the historical data gathered including data on 109 patients with birth weight less than or equal to 1500 g. The core of the database consists of support vector machine and logit regression classification results calculated specifically for that system, and obtained by considering 214 different combinations of 14 risk factors. Based on the results obtained and user demands, the system recommends the best methods and the most suitable parameter subset among those currently available to the user. The program is also able to estimate the accuracy, sensitivity and specificity together with their standard deviations. The user is also given information on which additional parameter it is worth adding to his measurement system most and what an increase in prediction efficiency it is expected to trigger. The BPD can be predicted by the system with the accuracy reaching up to 83.25% in the best-case scenario, i.e. higher than for most of the models presented in the literature. This work presents a set of examples illustrating the difficulties in obtaining one single model that can be widely used, and thus explaining why an expert system approach is much more useful in day-to-day clinical practice. In addition, the work discusses the significance of the parameters used and the impact of a chosen method on the sensitivity and specificity.