IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 53, NO. 11, NOVEMBER 2006 2211
Adaptive Change Detection in Heart Rate Trend
Monitoring in Anesthetized Children
Ping Yang*, Guy Dumont, Fellow, IEEE, and J. Mark Ansermino
Abstract—The proposed algorithm is designed to detect changes
in the heart rate trend signal which fits the dynamic linear model
description. Based on this model, the interpatient and intraoper-
ative variations are handled by estimating the noise covariances
via an adaptive Kalman filter. An exponentially weighted moving
average predictor switches between two different forgetting coef-
ficients to allow the historical data to have a varying influence in
prediction. The cumulative sum testing of the residuals identifies
the change points online. The algorithm was tested on a substan-
tial volume of real clinical data. Comparison of the proposed algo-
rithm with Trigg’s approach revealed that the algorithm performs
more favorably with a shorter delay. The receiver operating char-
acteristic curve analysis indicates that the algorithm outperformed
the change detection by clinicians in real time.
Index Terms—Adaptive Kalman filter, change point detection,
CUSUM, EWMA.
I. INTRODUCTION
HEART rate is one of the vital signs indicating disturbancesto the cardiovascular system in patients undergoing anes-
thesia and surgery. Astute pediatric anesthesiologists often rely
on subtle changes in the heart rate trend to guide changes in
depth of anesthesia as well as fluid administration. A seasoned
clinician can perceive changes in heart rate of three to four beats
per minute from the auditory signal [1], but may become dis-
tracted by simultaneous tasks. The monitoring strategy in cur-
rent devices only evaluates the heart rate amplitude against a
predetermined threshold without considering historical trends,
and therefore tends to be easily corrupted by artifacts, or irrele-
vant transients, causing a high false positive rate.
Various algorithms have been proposed for the trend detec-
tion problem in physiological variables [2]–[5]. In [4], the signal
is decomposed into successive straight lines using cumulative
sum (CUSUM) testing to determine where a new linear segment
needs to be fitted. Approaches based on fuzzy logic, Trigg’s
tracking index, wavelets and temporal polynomial fitting were
investigated in [5], and the performance of each method for heart
rate trend monitoring was compared in the aspects of detec-
tion capability, accuracy of pattern description and computation
speed.
Manuscript received June 26, 2005; revised April 9, 2006. Asterisk indicates
corresponding author.
*P. Yang is with the Department of Electrical and Computer Engineering,
the University of British Columbia, Vancouver, BC V6T 1Z4, Canada (e-mail:
pingy@ece.ubc.ca).
G. Dumont is with the Department of Electrical and Computer Engineering,
the University of British Columbia, Vancouver, BC V6T 1Z4, Canada.
J. M. Ansermino is with the Department of Anesthesia, British Columbia
Children’s Hospital, the University of British Columbia, Vancouver, BC V6H
3V4, Canada.
Digital Object Identifier 10.1109/TBME.2006.877107
The typical clinically significant change in the heart rate trend
signal is sustained over a period of time or has a large ampli-
tude. The required amplitude and duration vary between pa-
tients and between the types of surgery. Generally, a small am-
plitude and short duration change can be important in a smooth
and steady series, while in an eventful series, clinically signifi-
cant changes usually have a larger amplitude or longer duration.
Since most heart rate trend signals fluctuate in an unpredictable
pattern during surgery, in order to improve the online perfor-
mance, a real-time change detection algorithm needs to be able
to estimate the variability and adjust itself accordingly.
The CUSUM provides an obvious output for both duration
and amplitude, making it suitable for application in patient
monitoring. The challenge in using CUSUM testing lies in
determining the target values and thresholds adaptively. The
Kalman filter provides an efficient method of signal estimation
and change detection for a signal described by the dynamic
linear model (DLM). The conventional Kalman filter forecasts
the signal one step ahead, resulting in short-lived changes in
the residue. When used in heart rate trend monitoring, this
technique may not differentiate large momentary fluctuations
from true abrupt changes or miss some long duration changes
with small but relevant amplitudes. The exponentially weighted
moving average (EWMA) method is well suited for change
detection. However, its performance is limited by the fixed
forgetting parameter [6].
In this paper, we present a change detection scheme based
on an adaptive Kalman filter, an EWMA predictor and CUSUM
testing, and show its ability to adapt to both interpatient and in-
traoperative variations. Initially, we describe the proposed algo-
rithm and Trigg’s approach. The testing procedure is described
in Section III. Following the test steps, the proposed algorithm
is tested on a substantial volume of real data. The results of each
step are reported in Section IV. Limitations and extended appli-
cation to other parameters are discussed in the final section.
II. METHODS
A. Signal Modeling
Since the heart rate trend signal demonstrates piecewise linear
characteristics, the state-space linear growth DLM proposed by
Harrison and Stevens [7] is used to represent the heart rate as
(1)
where .
0018-9294/$20.00 © 2006 IEEE