1 MBEC_1277 Online-Classification of Capnographic Curves Using Artificial Neural Networks Marcus Bleil 1 , Alexander Opp 2 , Roland Linder 3 , Soehnke Boye 4 , Hartmut Gehring 4 , Ulrich G. Hofmann 2 1 Fernuniversität Hagen, Hagen 2 Institut für Medizintechnik, Universität zu Lübeck, Lübeck 3 Technisch-Naturwissenschaftliche Fakultät, Universität zu Lübeck, Lübeck 4 Klinik für Anästhesiologie, Universitätsklinikum Schleswig-Holstein, Campus Lübeck, Lübeck Abstract— Computer assisted capnometry is a tool for advanced patient monitoring with potential towards a decision support system. CO 2 values as represented in a capnogram reflect information about ventilation and gas exchange. Even more important, the shape of capnograms may reflect certain unexpected clinical relevant situations caused by ventilation or pathophysiological pulmonary reactions like bronchospasm. For the present study capnograms of different types of ventilation have been acquired and templates for the normal capnograms (as judged by experienced anaesthesiologists) were generated. A threshold method was used to extract single capnograms out of the CO 2 -monitor’s continuous data. Labeled capnograms of pathological events are gathered. For generating templates a simple correlation algorithm was used to classify capnograms [1]. The algorithm was further improved by rules inspired by visual inspection of the templates. In a second step an artificial neural network (ANN) was trained to assign capnograms to such pre-defined templates. The network consisted of 25 input and 10 hidden neurons. As training algorithm resilient propagation was used [2]. The performance of classification differed with regard to the type of ventilation and the kind of template. In most combinations more than 2/3 of classifications were correct, i.e., the ANN was able to recognize the correct class of the capnogram in question. For practical use the correlation algorithm and the artificial neural network were implemented on a PocketPC. The software can easily classify capnograms using both the ANN as well as the correlation algorithm. The aim of a decision support system to diagnose certain ventilation related diseases or problems in the area of anaesthesia or pneumology seems within reach by now. Keywords— capnometry, neural network, monitoring, anaesthesia, pneumology, capnogram, decision support system I. INTRODUCTION Capnometer continuously display CO 2 - concentrations of exhaled air, measured by appropriate sensors. However, experienced anaesthesiologists are even able to relate the according CO 2 -curves or shapes (the capnogram) to clinical relevant situations, like obstructed airways, bronchospasms [3], or pulmonary embolism [4]. Farhan et al. [5] emphasize the need for an automatic analysis system, as the clinical use of capnography nowadays is even covering spontaneous breathing patients. Previous studies show, that parameters describing the shape of capnograms do correlate with a disease. The present work deals with the automated classification of capnograms using artificial neural networks. To generate training and test patterns for the artificial neural networks [6], capnograms from clinically collected and labeled CO 2 curves were isolated and classified. To generate basic classes a correlation based forerunner algorithm as published in [1] was used. In the following, we present software implementing both algorithms (correlation and ANN) on a PocketPC. The inherent mobility of the device is meant to increase the acceptance and motivation of future users to the system. II. MATERIALS AND METHODES A. Capnograms Data sets for the present study were 144 CO 2 curves, collected in a clinical setting with 32 patients using a side stream CO 2 monitor [MicroCap, Oridion Medical, Luebeck, Germany]. Each CO 2 -curve contains the capnograms [1] of a 15 minutes data recording period. There were CO 2 curves in the ventilation situations „controlled ventilation, intubated“ (34 curves), „spontaneous breathing, intubated“ (36 curves), „spontaneous breathing, extubated, sleepy“ (37 curves) and „spontaneous breathing, extubated, awake“ (37 curves) recorded. We had at least one CO 2 curve for each patient and ventilation situation. CO 2 curves were digitized with a sampling frequency of 10Hz. A single CO 2 curve was stored as a text file. An average file contained 8891 sample values, corresponding to a measurement period of 15 minutes. B. Cluster Algorithm The goal of the clustering-by-correlation algorithm is to isolate analyzable capnograms from noise samples. A