Journal of the American Medical Informatics Association Volume 13 Number 3 May / Jun 2006
`
`309
`
`Application of Information Technology n
`A Method for Automatic Identification of Reliable Heart Rates
`Calculated from ECG and PPG Waveforms
`
`CHENGGANG YU, PHD, ZHENQIU LIU, PHD, THOMAS MCKENNA, PHD, ANDREW T. REISNER, MD,
`JAQUES REIFMAN, PHD
`
`A b s t r a c t Objective: The development and application of data-driven decision-support systems for medical
`triage, diagnostics, and prognostics pose special requirements on physiologic data. In particular, that data are reliable in
`order to produce meaningful results. The authors describe a method that automatically estimates the reliability of ref-
`erence heart rates (HRr) derived from electrocardiogram (ECG) waveforms and photoplethysmogram (PPG) waveforms
`recorded by vital-signs monitors. The reliability is quantitatively expressed through a quality index (QI) for each HRr.
`
`Design: The proposed method estimates the reliability of heart rates from vital-signs monitors by (1) assessing the
`quality of the ECG and PPG waveforms, (2) separately computing heart rates from these waveforms, and (3) concisely
`combining this information into a QI that considers the physical redundancy of the signal sources and independence of
`heart rate calculations. The assessment of the waveforms is performed by a Support Vector Machine classifier and the
`independent computation of heart rate from the waveforms is performed by an adaptive peak identification technique,
`termed ADAPIT, which is designed to filter out motion-induced noise.
`
`Results: The authors evaluated the method against 158 randomly selected data samples of trauma patients collected
`during helicopter transport, each sample consisting of 7-second ECG and PPG waveform segments and their associated
`HRr. They compared the results of the algorithm against manual analysis performed by human experts and found that
`in 92% of the cases, the algorithm either matches or is more conservative than the human’s QI qualification. In the
`remaining 8% of the cases, the algorithm infers a less conservative QI, though in most cases this was because of
`algorithm/human disagreement over ambiguous waveform quality. If these ambiguous waveforms were relabeled,
`the misclassification rate would drop from 8% to 3%.
`
`Conclusion: This method provides a robust approach for automatically assessing the reliability of large quantities of
`heart rate data and the waveforms from which they are derived.
`
`j J Am Med Inform Assoc. 2006;13:309–320. DOI 10.1197/jamia.M1925.
`
`Decision-support algorithms that automatically interpret stream-
`ing physiologic time-series data are valuable tools for a broad
`range of medical surveillance applications. Examples of such
`applications include acute monitoring of patients in intensive
`
`Affiliations of the authors: Bioinformatics Cell, Telemedicine and Ad-
`vanced Technology Research Center, US Army Medical Research and
`Materiel Command, Fort Detrick, MD (CY, ZL, TM, JR); Department
`of Emergency Medicine, Massachusetts General Hospital, Boston,
`MA (ATR); Dr. Liu is currently with the Division of Biostatistics,
`Greenebaum Cancer Center and Department of Epidemiology and
`Preventive Medicine, University of Maryland Medical Center.
`
`The work presented here was supported by the U.S. Army Medical
`Research and Materiel Command, Fort Detrick, MD.
`
`The authors express their gratitude to Col. John Holcomb and Dr. Jose
`Salinas of the U.S. Army Institute of Surgical Research, Fort Sam
`Houston, San Antonio, TX, who provided the trauma patient data.
`
`The opinions or assertions contained herein are the private views of
`the authors and are not to be construed as official or as reflecting the
`views of the U.S. Army or the U.S. Department of Defense.
`
`Correspondence and reprints: Jaques Reifman, PhD, U.S. Army
`Medical Research and Materiel Command MRMC/TATRC, 504
`Scott Street, Ft. Detrick, MD 21702-5012; e-mail: <jaques.reifman@
`us.army.mil>.
`
`Received for review: 08/01/05; accepted for publication: 01/16/06.
`
`care, home care, and ad hoc monitoring to continuously assess
`the health status of personnel, such as firefighters and
`soldiers, who are at risk of sudden injury.1 Advances in
`vital-signs monitoring software/hardware, miniaturization,
`storage capacity, wireless transmission, and computational
`power now allow recording and analysis of large quantities
`of physiologic data in a timely fashion. These data are invalu-
`able for the development of triage, diagnostic, and prognostic
`algorithms. However, collection of time-series vital-signs data
`is subject to many factors that affect the quality of the data. In
`particular, because vital-signs data are mostly collected in a
`noninvasive fashion, sensor motion artifact is of significant
`concern when the subject is moving or being transported.
`Other factors that may degrade data quality include electrical
`interference, sensor/monitor malfunction, and poor sensor
`placement on the subject. If valid decision-support algorithms
`are to be developed, and subsequently used to monitor
`patients, it is critical that reliable data be distinguished from
`artifact. Moreover, the process of distinguishing reliable from
`unreliable data must be automated since the sheer volume of
`collected time-series vital-signs data makes post hoc manual
`assessment an overwhelming task, while real-time streaming
`data cannot be manually evaluated at all.
`
`Heart rate (HR) is a critical vital sign that is continuously
`monitored during transport of trauma patients from the scene
`
`1
`
`APPLE 1080
`Apple v. AliveCor
`IPR2021-00972
`
`

`

`310
`
`YU ET AL., Identification of Reliable Heart Rates
`
`of injury to the hospital. It is used as an input for existing pre-
`hospital trauma severity scores, such as the prehospital in-
`dex,2,3 and may be used for future triage scoring systems.
`Also, studies of heart rate variability (HRV) suggest that de-
`creasing HRV may be associated with worsening patient sta-
`tus. Unfortunately, we have observed that randomly imposed
`noise spikes are sometimes counted as heart beats by a vital-
`signs monitor. These sorts of data corruption can mislead di-
`agnosis and compromise the development and application of
`inductive algorithms based on the synthesis of time-series
`physiologic data. Therefore, it is imperative that validated
`HRs be available for clinical use and development of
`advanced automated monitoring systems.
`
`Automated HR calculation is usually based on the identifica-
`tion of heart beat signals, which could be taken from the QRS
`complex or simply the R waves in electrocardiogram (ECG)
`waveforms, or the pulse waves in photoplethysmogram
`(PPG) waveforms,4–6 and dependent on the count of heart
`beats over a period of time. Given noisy waveforms, however,
`true heart beat signals may be masked or noise artifacts may
`resemble and be counted as true heart beats. Therefore, the
`quality of the HR calculated from the waveform depends
`on the quality of the waveform, making the qualification of
`waveforms a necessary step in validating HRs provided by
`a vital-signs monitor. Here, we refer to the monitor-calculated
`HRs as reference HRs (HRr). Accordingly, such HRr can be
`categorized as unreliable when the associated waveform is
`determined to be of suboptimal quality. For a conservative
`validation method, a high standard for good-quality wave-
`forms is preferred to minimize the possibility that bad-quality
`HRs are falsely categorized as good. However, an overly
`stringent threshold is not advisable since it will increase the
`chance that good-quality HRs are falsely categorized as bad
`and, for post hoc data analysis, will considerably reduce the
`amount of available good-quality HR for the development
`of data-driven, decision-support algorithms.
`
`In this paper, we present an approach to automatically and
`systematically qualify ECG HRr and PPG HRr provided by
`a vital-signs monitor. We assume that the monitor also pro-
`vides the corresponding waveforms from which they are de-
`rived and that the monitored individuals are alive and have
`been subject to a trauma injury, where arrhythmia is seldom
`observed. The approach numerically qualifies each sampled
`HRr by assigning to it a quality index (QI) that concisely ex-
`presses its reliability. The approach exploits the physical
`redundancy provided by ECG HRr and PPG HRr and em-
`ploys an independent method for recomputing HRs from
`the provided waveforms. This work addresses the first and
`key step of automatic and systematic qualification of large
`amounts of time-series data of our trauma database, so that
`we can next address our ultimate goal: mining these data to
`find predictive information for some clinical outcome.
`
`Figure 1 illustrates the three components of the approach. In
`the first component, we use the newly developed adaptive
`peak identification technique, termed ADAPIT, to indepen-
`dently compute HRs (HRc) from both ECG and PPG wave-
`form segments corresponding to the HRr we wish to
`validate. ADAPIT is a computationally simple peak detection
`algorithm, yet robust in the presence of random, motion-
`induced noise spikes that are often observed in waveforms
`collected during transport of
`trauma patients. Unless
`
`F i g u r e 1 . The three elements of the algorithm used to
`infer a quality index for reference heart rates provided by a
`vital-signs monitor.
`
`accounted for, these noise spikes are likely to be counted as
`heart beats by the vital-signs monitor. Next, we separately
`qualify ECG waveform segments and PPG waveform seg-
`ments as either good (excellent quality) or bad (suboptimal
`quality) through the use of a machine-learning algorithm in
`the form of support vector machines (SVMs).7 In the third
`and final component, through a decision-logic algorithm,
`we combine the results of
`the two previous steps,
`the
`ADAPIT-computed ECG HRc and PPG HRc and the quality
`of their corresponding waveform segments, and compare
`them against ECG HRr and PPG HRr provided by a vital-signs
`monitor to infer a QI for the two HRr. A QI is inferred each time
`a HRr is provided by a vital-signs monitor and ranges from
`zero to three, with three representing the best-possible quality.
`In the absence of one of the waveforms, the decision–logic al-
`gorithm still provides a QI by assuming that the absent signal
`is present but possesses poor quality. Should additional HR
`sources be available, the approach could be extended by prop-
`erly accounting for the quality of the new signal information
`and modifying the QI decision rules.
`
`The approach is modular, self-contained, and independent of
`the data collection hardware. The waveform qualification al-
`gorithm (SVM), the HR recomputation algorithm (ADAPIT),
`and the QI decision rules are developed independently of
`each other and can be separately exchanged by functionally
`equivalent modules based on other methods. The three
`components form an effective, stand-alone system to validate
`reference HRs. Our approach is simply based on recorded
`time-series data from a vital-signs monitor, which is taken
`as a black box. From this point of view, the approach is inde-
`pendent of the data collection hardware.
`
`Methods
`In this section, we briefly describe the three components
`depicted in Figure 1: the HR estimation via the ADAPIT
`algorithm, the waveform qualification via an SVM algorithm,
`and the QI determination. We start by describing the data that
`precipitated the development of these components and that
`are used for the synthesis and testing of our algorithms.
`
`Data
`This study is based on physiologic time-series data collected
`during transport of trauma patients from the scene of injury
`by helicopter service to the Level I unit at the Memorial
`Hermann Hospital in Houston, TX.8,9 The data were collected
`by ProPaq 206EL vital-signs monitors10 on the helicopters
`and downloaded to an attached personal digital assistant.
`The data include, among other time-series data, ECG and
`
`2
`
`

`

`Journal of the American Medical Informatics Association Volume 13 Number 3 May / Jun 2006
`
`311
`
`PPG waveform signals and their corresponding monitor-
`calculated HRr. The time series sampling rates are approxi-
`mately 182 Hz for the ECG waveform, 91 Hz for the PPG
`waveform, and 1 Hz for the HRr. Complete vital-signs data
`for a total of 726 patients were deposited into our Physiology
`Analysis System,11 which provides curated data and the abil-
`ity to query and analyze discrete and time-series data over
`the Internet with a Web browser. The patient population is
`composed of 538 males and 186 females (two genders not
`noted), with a mean age of 37.7 years. The predominant type
`of injury is blunt trauma (641 patients), followed by penetrat-
`ing trauma (78 patients).
`
`Heart Rate Estimation with the ADAPIT Algorithm
`The first component of our approach is the independent esti-
`mation of ECG and PPG HRs from their corresponding high-
`frequency waveforms. While we acknowledge that a large
`body of work has been developed over the past
`two
`
`decades,4–6 most of the approaches are rather involved be-
`cause they are designed to accommodate irregular morphol-
`ogies and irregular rhythms, even though such phenomena
`are rarely observed in our data set of trauma victims. Due
`to the ambulatory nature and dynamic environment in which
`trauma data are collected, the major challenge is the filtering
`of noise and artifacts in the waveforms. Furthermore, most
`approaches are limited to the estimation of ECG-derived
`HRs through the detection and analysis of the QRS complex,6
`while we also need to estimate PPG-derived HRs. To achieve
`these objectives, we developed the ADAPIT algorithm.
`ADAPIT is a generic algorithm that, through changes in pa-
`rameter settings and one computational step, is equally appli-
`cable to the estimation of HRs from both ECG and PPG
`waveforms and is designed to filter out noise and artifacts
`so they are not counted as heart beats. ADAPIT, however,
`may have limited ability to compute HRs in settings of highly
`irregular rhythms.
`
`F i g u r e 2 .
`Illustration of the identification of heart beats by the ADAPIT algorithm. (a) Original 7-second ECG waveform seg-
`ment. (b) Waveform after application of a median filter. (c) Difference of the original waveform in a minus the median-filtered
`waveform in b. The threshold T1 defines the segment’s baseline range [2T1, T1] and the threshold T2 provides a first cut on
`the lower limit of the peaks’ magnitude. (d) The first estimates of the actual peaks and threshold T3 (horizontal line) are used to
`eliminate small-magnitude spikes that clearly are not actual peaks. (e) String of markers with constant period P. (f ) Best alignment
`between the actual peaks and markers, which is used to estimate heart rates. (g) The heart beats found by the ADAPIT algorithm
`are marked on the original electrocardiogram waveform.
`
`3
`
`

`

`312
`
`YU ET AL., Identification of Reliable Heart Rates
`
`Estimation from Electrocardiogram Waveforms
`The ADAPIT algorithm computes an HRc at each time point
`(i.e., each second) t that a HRr is provided by the vital-signs
`monitor. This computation is performed based on a 7-second
`ECG waveform from time t-7 to t, which is approximately the
`same waveform length used by the vital-signs monitor,10 to
`estimate one HRr. Figure 2 illustrates the four major steps
`of the algorithm to compute HRc at t 5 0 (see Appendix 1
`for additional technical details).
`
`Step 1. ADAPIT applies a median filter (with a 55-ms window
`size) to the original 7-second waveform (Fig. 2a) and then
`subtracts the filtered signal (Fig. 2b) from the original one
`to yield the waveform in Figure 2c. This step de-trends the
`waveform, retains the amplitude of sharp R waves, and atten-
`uates broad waves, such as the P wave and T wave.
`
`Step 2. This step provides a first estimate of the actual peaks of
`the waveform through the sequential computation of two
`thresholds, T1 and T2. T1, illustrated in Figure 2c, is taken as
`2s1, where s1 denotes the standard deviation of all data point
`values of the 7-second waveform and defines the segment’s
`baseline range [2T1, T1], from which the baseline standard
`deviation s2 is calculated. T2, set to 3s2, is used as a lower
`limit of the waveform amplitude for considering potential
`peaks. Peaks greater than T2 are taken as the first estimate
`of the actual peaks (Fig. 2d).
`
`Step 3. To eliminate small-amplitude spikes that clearly are
`not R waves, a threshold T3 is defined as one half of the me-
`dian amplitude of all peaks identified in Step 2 (Fig. 2d). All
`peaks less than T3 are eliminated, as illustrated in Figure 2e.
`Step 4. To determine actual R waves from the peaks retained
`in Step 3, strings of markers with period P (Fig. 2e) are itera-
`tively generated and moved along the time line to align with
`the retained peaks. Through this iterative process, P is modi-
`fied to range from lengths equivalent to HRs between 25 and
`250 beats per minute (bpm). The string with the largest P
`aligned to the largest number of retained peaks is selected.
`Next, each unaligned marker of the selected string is allowed
`to move back and forth along the time line by as much as one
`half of P in an attempt to line up any unaligned peak (Fig. 2f).
`Finally, all aligned peaks, marked with circles on the original
`ECG waveform in Figure 2g, are assumed to be actual R
`waves. It should be noted that ADAPIT computes HRc based
`on all markers rather than the aligned peaks because an R
`wave could have been dropped during data collection or fil-
`tered out during the ADAPIT four-step process.
`
`To verify ADAPIT’s capability to filter out motion-induced
`artifacts and correctly compute HR of ambulatory trauma vic-
`tims, we had a human expert visually estimate the HR of
`80 seven-second, good-quality waveform samples from our
`database. Considering the human’s estimations as the gold
`standard, we compare them against ADAPIT, HRr, and a
`well-established QRS-based detection program termed
`ecgpuwave.12
`Figure 3 shows the difference between the algorithms’ and the
`human’s estimations for each of the 80 samples. The mean
`differences of ADAPIT, HRr, and ecgpuwave are, respec-
`tively, 20.62, 0.78, and 1.03 bpm, and the root mean square
`differences are 7.1, 5.1, and 7.1 bpm, respectively. These results
`indicate that in the process of filtering out noise, so as not
`to be counted as heart beats, ADAPIT tends to underestimate
`
`F i g u r e 3 . Difference in heart rates computed by three dif-
`ferent algorithms (ADAPIT, reference heart rate [HRr], and
`ecgpuwave) and a human expert.
`
`HRs, while the two other algorithms tend to overestimate
`them. This feature of ADAPIT is noticed, in particular, in wave-
`forms with highly irregular rhythms (samples 33 and 76) and
`provides a lower bound estimate for the HRs that allows for a
`conservative consistency check (larger delta) between HRr
`and HRc.
`Estimation from Photoplethysmogram Waveforms
`ADAPIT employs the same four-step process with two small
`modifications in the estimation of PPG-derived HRc. First, in
`Step 1, the median filter window size is extended to 550 ms to
`preserve broad pulse waves and attenuate sharp dicrotic
`notches. Second, after the identification of peaks in Step 3,
`each peak is smoothed with a moving-average filter of win-
`dow size equal to 110 ms. This additional filtering is needed
`to smooth out the broad and often distorted pulse waves
`and reduce the ambiguity in detecting the exact time of a
`heart beat, assumed to occur when the smoothed pulse
`wave reaches its maximum.
`Waveform Qualification
`This component of the approach implements our premise that
`the reliability of HRr is highly dependent on the quality of the
`underlying waveforms from which they are derived. A ma-
`chine learning classifier, implemented by an SVM, automates
`the categorization of waveforms by attempting to mimic the
`performance of human experts who rely on visual inspection
`and the application of some implicit or explicit rules of
`thumb. A classifier ‘‘learns’’ these rules by finding coefficients
`that optimize the ‘‘correlations’’ between a set of waveform-
`extracted features and waveform quality obtained from man-
`ually categorized waveform samples.
`
`Figure 4 illustrates the four steps in the development of a
`machine-learning classifier: (1) manually categorize sample
`waveform segments, (2) define candidate waveform features
`that distinguish good/bad waveforms, (3) select the most in-
`formative features, and (4) train and test the classifier. Once
`trained and given input features, the classifier categorizes
`waveform segments as being good or bad.
`
`Manual Waveform Categorization
`To develop the SVM classifier, human experts visually exam-
`ined and categorized 7-second waveform segments for 362
`
`4
`
`

`

`Journal of the American Medical Informatics Association Volume 13 Number 3 May / Jun 2006
`
`313
`
`the discrete-time fast Fourier transform13 to the ECG time-
`series data. These features are designed to exclude ECG fre-
`quency components that are associated with a QRS complex,
`while capturing high- and low-frequency component charac-
`teristics that may be attributed to noise and baseline drifts
`and shifts.
`
`The first time-domain feature is the fraction of aligned waves
`FW, which provides a measure of temporal regularity of po-
`tential heart beat signals. The second time-domain feature is
`a specific signal-to-noise ratio SN, which provides a measure
`of the distinctiveness of potential heart beat signals above the
`baseline. The pulse-wave variability (PV), extracted from PPG
`waveform segments, is the third time-domain feature and
`provides a measure of the variability of the time interval
`between two adjacent pulse waves.
`
`Feature Selection
`The goal of automatic feature selection is to choose and retain
`a subset of salient features from the original list of candidate
`features such that the process of pattern discovery by the
`machine-learning classifier is implemented in a reduced space
`without degrading its performance. The underlying philoso-
`phy is to retain features that can clearly characterize or dis-
`criminate the quality of
`the waveforms and eliminate
`features that are redundant, and hence, do not contribute
`additional information. Here, we employ information en-
`tropy14,15 as a measure of discriminatory power of the fea-
`tures. The most discriminatory (informative) feature has the
`lowest entropy.
`Our previously developed Rule Generator (RG) program14,15
`is used to compute entropies of candidate ECG and PPG
`waveform features. The RG program also defines patterns
`formed by these features and populated by the previously
`characterized samples to discriminate good/bad waveforms.
`The features that characterize the most discriminatory pat-
`terns, defined as the patterns that discriminate the largest
`number of samples, are selected as the most informative.
`Through this procedure, we find that HFE, FW, and SN are
`the most discriminatory features for ECG waveform classifi-
`cation and that FW and PV are the most informative features
`for PPG waveform classification.
`
`Support Vector Machine Classifier
`In this study, we employ our previously developed version of
`an SVM algorithm16 to classify ECG and PPG waveforms.
`The SVM, a recently proposed supervised machine-learning
`algorithm,7 has been shown to be an effective classifier in a
`wide variety of applications, including the categorization of
`ECG data.17–20 As a supervised-learning algorithm, the devel-
`opment (or ‘‘training’’) of an SVM requires a set of input/out-
`put training samples, where the inputs consist of a list of
`discriminatory features, such as the three ECG features and
`two PPG features selected in the previous section, and the
`outputs consist of labeled binary classes, good and bad.
`Once trained to implicitly ‘‘learn’’ the ‘‘rules’’ embedded in
`the training samples, given the values of the input features,
`extracted from a waveform segment that we wish to classify,
`the SVM automatically categorizes the segment as good or
`bad. An in-depth description of SVMs can be found in
`Vapnik.7
`We trained and tested an SVM classifier through a cross-
`validation procedure employing the manually categorized
`
`F i g u r e 4 . The development of machine-learning classi-
`fiers requires (1) manual categorization of good/bad wave-
`form-segment samples,
`(2) definition and extraction of
`candidate waveform features, (3) selection of the most discrim-
`inatory features, and (4) training and testing of the machine-
`learning classifier. Once trained and given input features, the
`classifier categorizes waveform segments as being good or
`bad.
`
`ECG samples and 388 PPG samples randomly selected from
`different patients. Of these, 194/168 ECG samples and 180/
`208 PPG samples were categorized as good/bad based on
`the following rules:
`
`An ECG segment is ranked as bad (suboptimal) if more than one ex-
`pected R wave is not observed or if the R wave is indistinguishable
`from noisy peaks. Otherwise it is ranked as good. A PPG segment
`is ranked as bad (suboptimal) if more than one expected pulse
`wave is not observed or if any one pulse wave peak cannot be distin-
`guished from a dicrotic notch. Otherwise it is ranked as good.
`
`These rules express the hypothesis that if more than one heart
`beat signal in a 7-second waveform segment is ambiguous,
`the HR calculated from such segment may be inaccurately ex-
`trapolated. The rules are conservative by design so that the in-
`ductively constructed classifiers are equally conservative and
`attempt to ensure that even if the classifier produces occa-
`sional false good waveform evaluations, those false good
`waveforms will still be of sufficient quality for estimation of
`HRs.
`
`Candidate Waveform Features
`A key phase in the development of machine-learning classifiers
`involves the definition and extraction of candidate features
`that can be used as class discriminators. For the characteriza-
`tion of waveforms as good or bad, we define three features in
`the frequency domain from ECG waveforms and three features
`in the time domain from ECG and PPG waveforms. Their def-
`initions are presented in Appendix 2.
`
`Similar to the ADAPIT algorithm, we extract features from
`7-second waveform segments that
`immediately precede
`each HRr we wish to qualify. The three frequency-domain fea-
`tures, high-frequency energy (HFE), low-frequency energy
`(LFE), and their ratio LFE/HFE, are obtained by applying
`
`5
`
`

`

`314
`
`YU ET AL., Identification of Reliable Heart Rates
`
`waveform samples (362 ECG and 388 PPG), where at each of
`200 cross-validation repetitions 70%–30% of the samples were
`used for training-testing the classifier. For all simulations, we
`used the same SVM model with a linear kernel function and
`at the end of the 200 simulations computed average perfor-
`mance measures, such as sensitivity and specificity, for the
`classifier. We did not attempt to optimize the SVM classifier.
`Classifier sensitivity provides a measure of the incorrectly
`classified (i.e., missed) bad waveform segments, whereas
`classifier specificity provides a measure of false hits, i.e., the
`fraction of good segments classified as bad.
`
`Averaged over the 200 cross-validation repetitions, the SVM
`yielded 93% sensitivity and 96% specificity for the ECG wave-
`forms, and 91% sensitivity and 88% specificity for the PPG
`waveforms. The slightly worse performance for the PPG
`waveforms reflects the increased difficulty in classifying this
`waveform due to a lack of more distinct characteristics of
`its profile. Considering that very conservative rules were
`used to categorize bad waveform segments, the 93% and
`91% sensitivity result can be taken as conservative estimates
`of the classifier’s ability to correctly categorize truly bad
`waveforms. Indeed, for those segments assigned bad quality
`by human experts that the SVM misclassified, our visual esti-
`mates of the HRs are compared and agree with those HRr
`provided by the vital-signs monitor. This indicates that mis-
`classification of bad waveforms by the classifier may still
`lead to correct estimation of HRc.
`
`Quality Index Determination
`The final component of the algorithm is the numerical quali-
`fication of the ECG HRr and PPG HRr provided by the vital-
`signs monitor. The qualification combines the independent
`estimation of ECG HRc and PPG HRc from redundant sour-
`ces, their reference values HRr, and the results of the wave-
`form SVM classifier to assign a QI that concisely expresses
`the reliability of each HRr provided by a vital-signs monitor.
`A QI of 3 indicates that both ECG HRr and the PPG HRr are
`‘‘highly’’ reliable, a QI of 2 indicates that ECG HRr is ‘‘fairly’’
`reliable, a QI of 1 indicates that PPG HRr is ‘‘somewhat’’ reli-
`able, and a QI of 0 indicates that neither HRr is reliable. The
`qualification algorithm assumes that good-quality HRr
`should come from high-quality waveforms and should be
`consistent with our independently calculated HRc. Another
`implicit assumption is that the data originate from live
`patients.
`
`Table 1 describes the rules used to generate the four QIs. The
`entries in the second and third columns indicate the quality of
`the two waveforms. The entries in the fourth and fifth
`columns indicate whether the ECG HRr and PPG HRr, re-
`spectively, are consistent with their corresponding HR com-
`puted by ADAPIT. HRr and HRc are consistent with each
`other when the discrepancy e1 , 5%, with
`
`e1 5
`
`jHRr 2 HRcj
`0:5 ðHRr 1 HRcÞ:
`
`ð1Þ
`
`The entries in the last column indicate whether all four HRs
`are consistent. Consistency is achieved when the discrepancy
`e2 , 10%, where e2 is defined as the ratio of the largest abso-
`lute difference among the six possible pairwise comparisons
`and the average HR over the four values. The table entries
`denoted with a dash indicate that consistency is not required.
`
`Table 1 j Rules Describing the Four Quality Indices
`
`Waveform Quality
`
`Heart Rate Consistency
`
`Quality Index
`
`ECG
`
`PPG
`
`ECG
`
`PPG
`
`All Four
`
`3
`2
`
`1
`
`0
`
`Yes
`Yes
`Yes
`Good
`Good
`–
`No
`Yes
`Good
`Good
`–
`–
`Yes
`Bad
`Good
`–
`Yes
`No
`Good
`Good
`–
`Yes
`–
`Good
`Bad
`All other cases that do not match conditions above
`
`– represents that consistency is not required.
`ECG 5 electrocardiogram; PPG 5 photoplethysmogram.
`
`For example, a QI of 3 is inferred when both ECG and PPG
`waveforms are classified by the SVM as having good quality,
`ECG HRr and PPG HRr are consistent with their correspond-
`ing ADAPIT-computed HR, and all four HRs are consistent
`with each other. This rule infers that the two HRr, which orig-
`inate from redundant sources (ECG vs. PPG) and are inde-
`pendently calculated (vital-signs monitor vs. ADAPIT), are
`in agreement and, with high confidence, correctly represent
`the actual HR. In this case, either ECG HRr or PPG HRr could
`be used to represent the actual HR. A QI of 2 is inferred in two
`possible scenarios. First, when both ECG and PPG wave-
`forms have good quality and ECG HRr is consistent with
`ECG HRc. Second, when the ECG waveform has good qual-
`ity, the PPG waveform has bad quality, and ECG HRr is con-
`sistent with ECG HRc. In essence, this rule expresses the
`situation where the ECG provides reliable information and
`ECG HRr alone should be used to represent the actual HR.
`Conversely, a QI of 1 indicates the situation where the PPG
`provides reliable information and PPG HRr alone should be
`used. Note that for equivalent requirements in the rules for a
`QI of 1 and QI of 2, we assign a higher confidence for the
`ECG HRr. This is to reflect that the ECG waveform is generally
`more reliable, possesses distinctive features that facilitate char-
`acterization, and is the one often used as the gold standard for
`HR computation. Finally, by exclusion, when none of the
`above conditions are satisfied, we assign a QI of 0 to indicate
`that neither HRr should be used to represent the actual HR.
`In the absence of one of the waveforms or their derived HRs,
`the decision–logic algorithm still provides a QI by assuming
`that the absent signal is present but possesses poor quality.
`
`In summary, the highest reliability (QI 5 3) is achieved only
`when the redundantly measured and independently com-
`puted HRs corroborate each other, while the lower reliability
`levels (QI 5 2 and QI 5 1) only require agreement between
`independent computations from the same source.
`
`Results and Discussion
`We demonstrate the performance of our algorithm through
`the analysis of two examples illustrated in Figures 5 and 6.
`Figure 5 shows an example where the ECG waveform is
`mostly noisy over a 40-second interval and the PPG wave-
`form is partially noisy. The thick horizontal bars on the top
`panel of the figure indicate segments of bad-quality wave-
`forms determined by the SVM classifier. The middle panel
`shows a deviance of the ECG HRr that is considerably larger
`than the other three HRs, which, in contrast, are more consis-
`tent. This suggests that noise spikes counted by the vital-signs
`monitor as heart beats are filtered out by ADAPIT in spite of
`