A robust sensor fusion method for heart rate estimation

Physiologic data measured in the clinical environment is frequently corrupted causing erroneous data to be displayed, periods of missing information or nuisance alarms to be triggered. To date, the possibility of combining sensors with similar information to improve the quality of the extracted data has not been developed. The objective of this work is to develop a method for combining heart rate measurements from multiple sensors to obtain: (i) an estimate of heart rate that is free of artifact; (ii) a confidence value associated with every heart rate estimate which indicates the likelihood that an estimate is correct; (iii) a more accurate estimate of heart rate than is available from any individual sensor.

The essence of the method is to discriminate between good and bad sensor measurements and combine only the good readings to derive an optimal heart rate estimate. Past estimates of heart rate are used to derive a predicted value for the current heart rate that is also fused along with the sensor measurements. Consensus between sensor measurements, the predicted value and physiologic credibility of the readings are used to distinguish between good and bad readings. Three sensor measurements and the predicted value are evaluated yielding 16 possible hypotheses for the current state of the available data. A Kalman filter uses the most likely hypothesis to derive the fused estimate. Statistical measures of the sensor error and rate of change of heart rate are adaptively estimated when data are sufficiently reliable and used to enhance the hypothesis selection process.

The method of sensor fusion presented has been documented to perform well using clinical data. Limitations of the technique and the assumptions employed are discussed as well as directions for future research.