{Abstraction and Representation of Repeated Patterns in High-Frequency Data}

Monitoring devices in intensive care units deliver enormous streams of data in the form of snapshot measurements. In contrast, physicians use high-level abstractions in reasoning about the parameters observed. Standard data abstraction algorithms can only handle data which are more regular than many variables observed in medicine. A typical example is the ECG in intensive care, where electric currents are measured at the skin surface and displayed amplied in order to detect problems in the conduction system of the muscular contraction pattern. We developed an algorithm to transform a curve constituted by a series of data points into a set of bends and lines in between them. The resulting qualitative representation of the curve can be expressed as a list of objects each describing a bend. In this format, it can easily be utilized in a Knowledge-Based System (KBS). In addition, in the case of rhythmical data, comparing selected bends in all cycles of the oscillation yields new information. This comparison can be done by plotting derived data as separate graph beside the original one or by encoding the knowledge behind the reasoning in rules in the KBS. Our algorithm performs best on curves which are rhythmical but too irregular to be analyzed by Fast Fourier Transformation or other standard methods aiming at describing regular patterns. We demonstrate our approach by displaying heart rate and Q-S-distance graphically aside of ECG-data (to detect impeded conduction) and by showing example code for rules detecting pathological deviations from the standard based on the qualitative representation of the curve.