Multimodal data classification using signal quality indices and empirical similarity-based reasoning

All bedside monitors are prone to heterogeneity and mis-labeled data, yet each multimodal sample data contains different sets of multi-dimensional attributes. To reduce the incidence of false alarms in the Intensive Care Unit (ICU), a new interactive classifier was proposed. In the algorithm, case was represented with signal quality Indices(SQIs) and RR interval features. With the function wabp, the annotations were obtained from the target signal after preprocessing. Five features were used as the inputs to a case-based reasoning classifier, retrieving the cases with empirical similarity. With the posted 750 records of the PhysioNet/CinC 2015 Challenge, the classifier was trained for answering the alarm types of the query segments. Compared with conventional threshold-based alarm algorithms, the performance of our proposed algthom reduces the maximum number of false alarms while avoiding the suppression of true alarms. Evaluated with the hidden test dataset, both real-time and retrospective, the results show that the overall TPR is 83% and 82% respectively; and TNR 44% and 43% respectively. This algorithm offers a new way of thinking about retrieving heterogeneity patients with multimodal data and classifying the alarm types in the context of mis-labeled cases.