On the study of the weaning results of ventilator-dependent patients with closest reasonable centroids

Ventilator has been widely used to support the breathing needs of patients who have difficulty in breathing. For a ventilator-dependent patient, it is hard to tell when is the proper time to remove ventilator from him (or her) and whether the weaning process would be successful or not. Recently, many models of applying artificial intelligence (AI) techniques into the analysis of clinical data have been proposed. Unfortunately, most models provide little help when specific “cause–effect” relation of data is not available, or even known. In this paper, an innovative method, called closest reasonable centroids (CRC), is directed to address this issue. Our present application domain was a clinical data set of 189 weaning records of ventilator-dependent patients. Each record consists of 4 observatory items, 8 physiological items, 10 disease items, 5 infection items, and its weaning result. Experimental result shows that the CRC's differentiability is comparable to those of back-propagation neural networks (BPN), support vector machine (SVM) and artificial neuromolecular (ANM) system, but better than that of Waikato environment for knowledge analysis (WEKA). From the health conditions of patients, the proposed method can roughly differentiate the weaning result and indicate the possible occurrence of particular infection. Additionally, item analysis reveals the most salient items in such differentiation process and infection indication process. All of the above results have been double confirmed by the clinicians, implicating that CRC could be used as assistant tool.

Research highlights▶ This study aims to investigate the “cause–effect” relations for clinical data. ▶ The “closest reasonable centroids” method can afford to explain such relations. ▶ This method can differentiate the weaning result of ventilator-dependent patients. ▶ It is also capable of indicating the possible occurrence of particular infection. ▶ This method could be used as assistant tool for clinical practices.