Search

For effective patient therapy and improved patient safety, it is critical to administer medication accurately in accordance with doctor's prescription. However, accidents owing to the erroneous programing of infusion pumps caused by users have been consistently reported in several documents. In this study, the authors propose a novel surveillance technique for infusion pumps to continuously monitor the variations in panel digits using a convolutional neural network model, and evaluate the performance of the implemented technique. During the experimental evaluation, 1st-step ROIs and 2nd-step ROIs were successfully extracted from the frame images regardless of the ambient lighting conditions. The final accuracies of the implemented CNN model are 99.9% for both the training (172,800 images) and validation (1080 images) dataset while the final losses for the training and validation datasets are 0.48 and 0.45 after 13th epoch, respectively. In the 24-h continuous monitoring test, the accuracy of the model for volume recognition considering all the 1440 measurements (960 for day-lighting and 480 for night-lighting ) is 95.5%, whereas in day-lighting and night-lighting modes the accuracies of the model are 98.2% and 90.0%, respectively. Based on these experimental results, the proposed surveillance technique incorporating infusion pumps is expected to improve the safety of patients who need long-term treatments via infusion pumps, reducing the burden on the nurses and hospitals.

An adaptive network based fuzzy inference system (ANFIS) model was employed to predict effluent chemical oxygen demand (COD(eff)) and ammonia nitrogen (NH(4)(+) (eff)) from an anaerobic/anoxic/oxic (A(2)/O) process, and meanwhile a self-adapted fuzzy c-means clustering algorithm was used to identify the model's architecture and optimize fuzzy rules. When constructing the model or predicting, the on-line monitoring parameters, namely hydraulic retention time (HRT), influent pH (pH), dissolved oxygen in the aerobic reactor (DO) and mixed-liquid return ratio (r), were adopted as the input variables. Compared with the artificial neural network (ANN) model whose weight vector was optimized by a real-code genetic algorithm (GA), the ANFIS presented better estimate performance. When predicting, the mean absolute percentage errors (MAPEs) of 1.8458% and 2.8984% for COD(eff) and NH(4)(+) (eff) could be achieved using ANFIS; the root mean square errors (RMSEs) for COD(eff) and NH(4)(+) (eff) were 1.6317 and 0.1291, respectively; the correlation coefficient (R) values of 0.9928 and 0.9951 for COD(eff) and NH(4)(+) (eff) could also be achieved. The results indicated that reasonable monitoring A(2)/O process performance, just using on-line monitoring parameters, has been achieved through the ANFIS.