The Science Journal of the American Association for Respiratory Care

2012 OPEN FORUM Abstracts


Hung-Ju Kuo1,2, Mauo-Ying Bien3,4, Chun-Nin Lee1, Hung-Wen Chiu2; 1Division of Pulmonary Medicine, Department of Internal Medicine, Shuang Ho Hospital, Taipei Medical University, Taipei, Taiwan; 2Graduate Institute of Biomedical Informatics, Taipei Medical University, Taipei, Taiwan; 3Division of Pulmonary Medicine, Department of Internal Medicine, Taipei Medical University Hospital, Taipei, Taiwan; 4School of Respiratory Therapy, Taipei Medical University, Taipei, Taiwan

Background: The rapid shallow breathing index (RSBI) is commonly used clinically for predicting the outcome of weaning from mechanical ventilation. However, there are existed different thresholds and sensitivities of RSBI among different populations and measurement conditions. There is no single appropriate and convenient predictor or method can be used to help the clinicians to predict the weaning outcome. The artificial neural networks (ANNs), are the machine-learning models which can change their structures and outputs based on the external or internal information during the learning phase, they had been applied in modeling the medical decision support systems. The purpose of this study was to design an ANN model for predicting the weaning outcome of mechanically ventilated patients. Method: Ninety-five ready for weaning patients living in medical intensive care unit were recruited and randomly divided into training group (n=76) and testing group (n=19). Eight features including patientsÂ’ age, reasons for intubation, duration of using mechanical ventilator, APACHE II score, the mean of inspiratory time, the mean of expiratory time, the mean of respiratory rate and the mean of tidal volume in thirty minutes spontaneous breathing trail under PSV 5 cmH2O with PEEP 5 cmH2O ventilator were selected as the ANN input variables. The performance of ANN model was compared with 1-min RSBI measurement method by using confusion matrix and the receiver operating characteristic curves. Result: The area under the receiver operating characteristic curves (AUROC) of ANN model was 0.95 and that for the RSBI method was 0.51 when the threshold was set to 105 breaths/min/L. Predictions by the testing group of ANN model had a sensitivity of 91.7%, a specificity of 85.7%, and an accuracy rate of 89.5%, compared with 75%, 14.3% and 52.6%, respectively, for the RSBI method. Conclusion: In our study, the ANN model improved the accuracy for prediction of weaning outcome. By applying this ANN model clinically, the clinicians could select the appropriate weaning time as early as possible, which could decrease the chance of unnecessary prolonged ventilator support and premature weaning. Therefore, the incidence of patientsÂ’ complication rate and medical cost related to ventilator support will decrease. Sponsored Research - None Comparison table between ANN model and 1-min RSBI measurement method