Noise-resistant pulse oximetry using a synthetic reference signal.

We describe a noise-resistant pulse oximetry algorithm suited to both signal reconstruction and oxygen saturation estimation. The algorithm first detects relatively clean signal sections from which the heart rate is estimated. The heart rate is used to construct a synthetic reference signal that matches an idealized pulse signal. An adaptive filter continuously processes the sensor signals, reconstructing signals in a linear subspace defined by the reference signal. A projective subspace algorithm is then applied to find the oxygenation level of the blood. We show that under specific circumstances this algorithm solves the sufficiency condition for signal reconstruction in linear saturation estimators. The core principle of using a frequency modulated synthetic reference signal can be applied to adaptive filtering of other physiological signals controlled by the heartbeat, such as blood pressure and electrocardiogram.

Critique of arrhythmia detectors based on heuristic rules.

Every arrhythmia detector employs a beat classifier to discriminate between normal (N) and ventricular (V) beats. In most of these beat-classification algorithms, a set of rules is employed to distinguish between N and V beats using a common set of features extracted from the real-time ECG signal and/or correlation of QRS complexes with the dominant QRS template. A common set of these features includes: beat area, beat width, beat amplitude, beat polarity, and R-to-R interval. Heuristic methods are commonly used to adapt the rules to particular databases. These classifiers are rule-based classifiers that employ AND-OR binary structures and hand-tuned thresholds for making decisions in the feature space. The complexity of the feature space increases as the number of features increases. For k features, a k-dimensional space is required. Thus, the separation between N and V space distributions becomes more difficult, especially since these distributions overlap. When AND-OR binary structures with hand-tuned thresholds or linear-separation techniques are used to separate N and V distributions in a k-dimensional feature space, errors are guaranteed, because these distributions are not linearly separable. As a results, these algorithms have limited dynamic ranges. This means that the sensitivity for a certain class of beats (N or V) will grow only at the expense of positive predictivity for that class, and vice versa.

Algorithm to identify components of arterial blood pressure signals during use of an intra-aortic balloon pump.

Existing bedside cardiovascular monitors often inaccurately measure arterial blood pressure during intra-aortic balloon pump (IABP) assist. We have developed an algorithm that correctly identifies features of arterial pressure waveforms in the presence of IABP. The algorithm is adaptive, functions in real-time, and uses information from the electrocardiographic (ECG) and arterial blood pressure signals to extract features and numeric values from the arterial blood pressure waveform. In its current form, it requires reliable ECG beat detection and was not intended to operate under conditions of extremely poor balloon timing. The algorithm was evaluated by an expert (P.F-C.) on a limited data set, which consisted of 12 1-minute epochs of data recorded from 6 intensive care unit patients. A criterion for selection of patients was that the ECG beat detector could detect ECG beats correctly from the waveforms. The overall sensitivity and positive predictivity for beat detection were 94.04% and 100%, respectively. For feature identification, the overall sensitivity was greater than 89%, positive predictivity was 100%, and the false-positive rate was 0%. The performance measures may be biased by the criteria for patient selection. This approach to identifying waveform features during IABP improves the accuracy of measurements. The utility of using 2 sources of information to improve measurement accuracy has been demonstrated and should be applicable to other physiologic signal-processing applications.

Automated testing of arrhythmia monitors using annotated databases.

Arrhythmia-algorithm performance is typically tested using the AHA and MIT/BIH databases. The tools for this test are simulation software programs. While these simulations provide rapid results, they neglect hardware and software effects in the monitor. To provide a more accurate measure of performance in the actual monitor, a system has been developed for automated arrhythmia testing. The testing system incorporates an IBM-compatible personal computer, a digital-to-analog converter, an RS232 board, a patient-simulator interface to the monitor, and a multi-tasking software package for data conversion and communication with the monitor. This system "plays" patient data files into the monitor and saves beat classifications in detection files. Tests were performed using the MIT/BIH and AHA databases. Statistics were generated by comparing the detection files with the annotation files. These statistics were marginally different from those that resulted from the simulation. Differences were then examined. As expected, the differences were related to monitor hardware effects.