Source: Heart Rhythm
Abstract
Background
Early detection of atrial fibrillation (AF) is key for preventing strokes. Blood pressure monitors (BPMs) with built-in AF screening features have the potential for early detection at home. Recently, two BPMs (HEM-7371T1-AZ and HEM-7372T1-AZAZ, OMRON HEALTHCARE Co., Ltd.) which share a novel AF screening feature have been developed. Their AF screening feature utilizes an algorithm which incorporates machine learning, with the potential to improve diagnostic accuracy.
Objective
To evaluate the performance of this AF screening feature we performed a multicenter, prospective clinical study at five sites in the United States.
Methods
A total of 559 subjects were enrolled for this study – 267 in AF (AF cohort) and 292 in the Non-AF cohort. AF screening was carried out on all subjects by the two Omron BPMs and by one Microlife BPM (BP 3MX1-3, WatchBP Home A, Microlife Corp), and a simultaneous 12-lead ECG was recorded for comparison. All 12-lead ECGs were interpreted by a board-certified cardiologist who was blind to the BPM results. Sensitivity, specificity, and accuracy for the diagnosis of AF were calculated.
Results
Omron HEM-7371T1-AZ BPM had a sensitivity of 95.1% [95% CI 91.8-97.4], specificity of 98.6% [95% CI 96.6-99.7] and accuracy of 97.0% [95% CI 95.2-98.2]. Equivalent results were obtained with the Omron HEM-7371T1-AZAZ BPM. This compared favorably to the Microlife BPM (sensitivity 78.5% [95% CI 73.1-83.3], specificity 97.6% [95% CI 95.1-99.0], and accuracy 88.4% [95% CI 85.5-91.0]).
Conclusion
These data support both home and professional use of these novel Omron BPMs for the detection of AF.
Keywords
Abbreviations:
AI (artificial intelligence), AF (atrial fibrillation), BPM (blood pressure monitor), ECG (electrocardiogram), NPV (negative predictive value), PPV (positive predictive value), PACs (premature atrial contractions), PVCs (premature ventricular contractions), PPI (pulse-to-pulse intervals), PPW (pulse pressure waves)
Introduction
Atrial fibrillation (AF) is a common cause of ischemic strokes – where 25% of patients presenting with their first ischemic stroke are in AF [1], and as much as 30% of patients with cryptogenic stroke (stroke of unclear etiology) develop AF during 3-year follow up [2]. Early detection of AF prior to stroke allows for anticoagulation which is highly effective at reducing stroke risk. Early detection of AF also makes long-term maintenance of sinus rhythm more feasible. To this end, healthcare providers have long utilized medical devices in symptomatic patients to detect AF including: pacemakers and defibrillators, implantable loop recorders, externally applied heart monitors, in-office ECG, and in-hospital telemetry. More recently, commercially available devices have adopted AF detection algorithms with the potential to screen a larger, asymptomatic population. These include “smart watches”, home ECG monitors such as KardiaMobile®, and home blood pressure monitors (BPMs). Of these, BPMs have great potential for population-based screening for AF due to their low cost, ease-of-use, and public availability. In order to be used as a screening tool for the population, BPM’s detection algorithm for AF must have both a high sensitivity and specificity. Low sensitivity results in under-detection and a missed opportunity for prevention. Low specificity results in a mis-diagnosis of AF leading to negative emotional impact, needless subsequent health care costs and unnecessary treatments with potential for harm. Previous studies on Microlife and Omron BPM products have demonstrated variable sensitivity (Microlife 76-100%, Omron 30-100%) and specificity (Microlife 81-99%, Omron 78-97%) [3, 4, 5, 6, 7], with conflicting conclusions on their comparative efficacy [8,9]. Recently, two Novel BPMs with AF screening (HEM-7371T1-AZ with cuff HEM-FL31, and HEM-7372T1-AZAZ with cuff HEM-RML31, Omron Healthcare Co. Ltd., Kyoto, Japan) have been developed. Importantly, these Omron BPMs utilize novel technology in their shared AF detection algorithm – incorporating over three hundred mathematical indices into a machine learning algorithm (AdaBoost, [10], Figure 1). Machine learning algorithms are a sub-set of artificial intelligence (AI) and have previously been shown to improve AF detection utilizing electrocardiograms [11,12]. To investigate whether a machine learning algorithm can improve the diagnostic accuracy of these novel Omron BPMs intended for home use, a multicenter prospective clinical trial was conducted at five sites in the United States. The present trial evaluated the statistical accuracy of two Omron BPM/cuff combinations for the detection of AF, and compared them to a conventional BPM with an AF screening feature (BP 3MX1-3, WatchBP Home A, Microlife Corp, Taipei, Taiwan). All three BPMs analyze pulse-to-pulse intervals (PPIs) that are calculated based on pulse pressure waves (PPW) detected in their AF detection software to differentiate an ultimate “AF” versus “Non-AF” binary output. While the Microlife product utilizes an established “irregularity index” [13], the novel Omron BPMs utilize the machine learning algorithm.
02520-7&id=gr1.jpg){kind=link}
Methods
This study was a prospective, parallel-cohort, open-label, non-randomized study involving subjects with atrial fibrillation (AF) and without AF (Non-AF) (IRB #20223072). The AF cohort consisted of subjects with a known history of AF, in AF as confirmed by ECG at the time of enrollment. This was a multicenter study with subjects recruited from 5 independent clinical practices in the United States. The Non-AF cohort consisted of subjects without a clinical history of AF, and not in AF at the time of enrollment. Subjects with a pacemaker or defibrillator were excluded. Age, gender and racial diversity were ensured through prespecified subgroup recruitment requirements. Subjects with Non-AF arrhythmias (e.g. PACs, PVCs, or sinus arrhythmia) were not excluded from participating, but rather included in the Non-AF cohort.
Three device combinations were tested against a resting ECG for their diagnostic accuracy for detecting AF: (1) Omron BPM HEM-7371T1-AZ with cuff HEM-FL31, (2) Omron BPM HEM-7372T1-AZAZ with cuff HEM-RML31, and (3) Microlife WatchBP Home A, BP3MX1-3 BPM. Blood pressures were taken once and sequentially with the three combinations on all subjects, just after medical history was obtained and ECG performed. Possible discrete outputs from all three devices were: “AF”, “Non-AF”, or “Error” – as displayed on the monitor screen (presence vs. absence of “Possible AFib Detected” displayed for Omron BPM, “AFib” icon displayed for Microlife BPM). The first “AF” or “Non-AF” output was used for analysis – with “Error” accepted as the ultimate categorization only when 3 sequential “Error” responses were displayed. “Error” categorizations were considered as “Non-AF” responses for all analyses. The ECG was interpreted by a board-certified cardiologist who was blind to the BPM results.
Sensitivity, specificity, accuracy, positive predictive value (PPV) and negative predictive value (NPV) of all 3 device combinations were calculated using simple statistics, with 95% confidence intervals determined using Clopper-Pearson [14] method for sensitivity, specificity and accuracy; while standard logit confidence intervals [15] were used for NPV and PPV. Subgroup analyses were performed across genders (male vs. female), age (>65 vs. <65 years-old) and races (Caucasian vs. non-Caucasian) for both Omron BPM / cuff combinations. The sensitivity, specificity, accuracy, PPV and NPV for both Omron BPM/cuff combinations were compared to the Microlife BPM using the Wald test.
The research reported in this paper adhered to the CONSORT guidelines and the Helsinki Declaration; with the study protocol approved by institutional review boards at each of the five study locations. The sponsor, Omron Healthcare, participated in trial design, centralized data collection, and provided input during revision of the manuscript; while authors participated directly in data collection and had direct access to the complete dataset of results for preparing this manuscript.
Results
Informed consent was obtained from 654 potential subjects. 81 subjects (12%) screen-failed due to a history of AF but without AF on ECG, leaving 573 subjects initially enrolled in the study. 14 enrolled subjects (2%) were excluded when ECG could not confidently diagnose or exclude AF. The remaining 559 subjects were used for all analyses, comprised of 267 in the AF group and 292 in the Non-AF group (Figure 2). Demographic data (Table 1) demonstrated balanced gender representation (49.6% male) of a typical age distribution for AF; whereas racial minorities (especially Hispanic subjects) were somewhat under-represented compared to the United States’ population. This was likely at least partially due to higher prevalence of clinically detected atrial fibrillation amongst Caucasians [16].
02520-7&id=gr2.jpg){kind=link}
Table 1Baseline Characteristics
| All Patients (N=559) | |
|---|---|
| Age in years (S.D.) | 64.9 (15.6) |
| Male | 277 (49.6%) |
| Race | |
| Caucasian | 424 (75.8%) |
| Black | 67 (12%) |
| Hispanic | 38 (6.8%) |
| Asian | 19 (3.4%) |
| Other | 11 (2.0%) |
| Heart Rate | |
| AF Cohort | 73 beats per minute |
| Non-AF Cohort | 71 beats per minute |
| Blood Pressure | |
| AF Cohort | 124 / 78 mmHg |
| Non-AF Cohort | 123 / 77 mmHg |
Categorization provided by the three BPM/cuff combinations (Table 2) was most notable for a higher error rate seen from the Microlife BPM (3.6% vs. 0.2% for both Omron BPMs), despite 3 attempts; and a higher false negative rate for the Microlife BPM or failure to detect AF when present. This resulted in a significantly lower sensitivity, accuracy and NPV for the Microlife BPM compared to both Omron BPM/cuff combinations (Table 3). Both Omron BPM/cuff combinations which incorporate the machine learning algorithm showed high sensitivity (95.1% [95% CI 91.8-97.4] in HEM-7371T1-AZ and 94.7% [95% CI 91.3-97.1] in HEM-7372T1-AZAZ), specificity (98.6% [95% CI 96.6-99.7] in HEM-7371T1-AZ and 98.3% [95% CI 96.1-99.5] in HEM-7372T1-AZAZ) and accuracy (97.0% [95% CI 95.2-98.2] in HEM-7371T1-AZ and 96.6% [95% CI 94.7-97.9] in HEM-7372T1-AZAZ) (Table 3). Subgroup analyses demonstrated no statistical difference in performance across genders (male vs. female), age (>65 vs. <65 years-old) and races (Caucasian vs. non-Caucasian) for both Omron BPMs – where across all subgroups, both devices’ sensitivity (93.3%-100.0%) and specificity (97.3%-99.3%) varied little (Table 4).
Table 2Categorization for AF Detection
| BPM | True Positive | True Negative | False Positive | False Negative | Error |
|---|---|---|---|---|---|
| Omron HEM-7371T1-AZ | 252 | 290 | 4 | 12 | 1 (0.2%) |
| Omron HEM-7372T1-AZAZ | 251 | 289 | 5 | 13 | 1 (0.2%) |
| Microlife WatchBP | 208 | 281 | 7 | 42 | 20 (3.6%) |
BPM = blood pressure monitor, True Positive = AF ECG + BPM output AF, True Negative = Non-AF ECG + BPM output Non-AF, False Positive = Non-AF ECG + BPM output AF, False Negative = AF ECG + BPM output Non-AF, Error = 3 consecutive Error outputs from BPM.
Table 3Overall Statistical Test Characteristics
| Microlife WatchBP | Omron HEM-7371T1-AZ | P Value∗ | Omron HEM-7372T1-AZAZ -AZAZ WatchBP | P Value∗ | |
|---|---|---|---|---|---|
| Sensitivity (%) | 78.5 (73.1-83.3) | 95.1 (91.8-97.4) | <0.0001 | 94.7 (91.3-97.1) | <0.0001 |
| Specificity (%) | 97.6 (95.1-99.0) | 98.6 (96.6-99.7) | 0.40 | 98.3 (96.1-99.5) | 0.58 |
| Accuracy (%) | 88.4 (85.5-91.0) | 97.0 (95.2-98.2) | <0.0001 | 96.6 (94.7-97.9) | <0.0001 |
| PPV (%) | 96.7 (93.5-98.4) | 98.4 (96.0-99.4) | 0.26 | 98.0 (95.5-99.2) | 0.39 |
| NPV (%) | 83.1 (79.7-86.1) | 95.7 (92.9-97.4) | <0.0001 | 95.4 (92.5-97.2) | <0.0001 |
∗ P Value is for each Omron BPM vs. Microlife BPM, BPM = blood pressure monitor, PPV = positive predictive value, NPV = negative predictive value, 95% confidence intervals in parentheses
Table 4Subgroup Statistical Test Characteristics
| HEM-7371T1-AZ Sensitivity | HEM-7371T1-AZ Specificity | HEM-7372T1-AZAZ Sensitivity | HEM-7372T1-AZAZ Specificity | |
|---|---|---|---|---|
| Male | 93.9 (90-97) | 97.9 (93-100) | 94.5 (91-98) | 97.9 (93-100) |
| Female | 97.6 (92-100) | 99.0 (96-100) | 95.2 (91-100) | 98.5 (96-100) |
| Age <65 | 93.3 (84-100) | 98.7 (95-100) | 96.7 (83-100) | 99.3 (96-100) |
| Age >65 | 95.3 (93-98) | 98.7 (95-100) | 94.5 (92-97) | 97.3 (95-100) |
| Caucasian | 94.6 (92-97) | 98.4 (95-100) | 94.6 (92-97) | 98.4 (95-100) |
| Non-Caucasian | 100 (87-100) | 99.1 (95-100) | 96.2 (80-100) | 98.2 (94-100) |
Both are Omron BPMs. Data as percentages (%). 95% confidence intervals in parentheses.
Discussion
The two novel Omron BPMs (HEM-7371T1-AZ and HEM-7372T1-AZAZ) which incorporate machine learning to detect AF were highly accurate (accuracy 97% with both cuffs, sensitivity 95% with both, and specificity over 98% with both). This compared favorably to a Microlife BPM (WatchBP Home A) which demonstrated a significantly lower sensitivity (79%, p<0.0001) driving a significantly lower diagnostic accuracy (88%, p<0.0001) for the detection of AF. Both a high sensitivity and specificity are vital if a BPM is to be utilized for AF screening, and the novel Omron BPMs used in this study achieve that requirement. Many people will use their BPM regularly to monitor blood pressure, and one false reading of AF could lead to unnecessary testing and turmoil (high specificity needed). However, an office-based BPM or one used in a public health campaign may be used thousands of times to screen a population of people just once, and under-detection of AF could be a missed opportunity to prevent a stroke (high sensitivity needed).
Importantly, the methodology utilized in the present study aimed to ensure external validity, by calculating BPM performance statistics using the first BPM output, “AF” or “No-AF”, on a per-subject basis. Additionally, when 3 sequential error messages occurred that subject was considered a failed attempt to detect AF in the AF cohort (false negative). This mimics in-home, in-office and public use; and this methodology differs from prior studies investigating BPM accuracy for AF detection. Specifically, some prior studies on a Microlife BPM utilized the “2-out-of-3 rule” – requiring 3 separate readings with at least 2 indicating AF to be considered positive for AF, bolstering sensitivity and diagnostic accuracy [17]. Also, some prior studies quoting high sensitivity and accuracy of a Microlife BPM performed statistics using multiple measurements from single subjects in their analyses rather than reporting on a per-subject basis [4,6], and one additionally excluded mis-diagnosis of Non-AF arrhythmias as AF from calculations [17]. In fact, when analyzed on a per-subject basis in the primary care setting the reported sensitivity for detection of AF for the Microlife BPM was similar to the current study – 79% current study, vs. 80% and 83% in two prior studies [6,18]. This low sensitivity seen in the general population led the UK to remove the Microlife Watch BP Home A BPM from its National Institute for Health and Care Excellence (NICE) guidelines on the detection and diagnosis of AF in 2022; instead recommending pulse palpation, 12-lead ECG and ambulatory ECG monitors [19]. The sensitivity for detecting AF for prior Omron BPMs without machine learning algorithms has also been called into question [20].
While the accuracy reported here for BPMs incorporating AI into AF detection is quite high – there remains room for improvement as false positive results were noted (“AF” resulted when no AF present). False positive AF results predominantly occurred in the setting of marked sinus arrhythmia or frequent atrial and ventricular ectopy. False negatives were rare, and resulted when atrial fibrillation lacked marked beat-to-beat heart rate variability. Also, further prospective studies validating the Omron BPMs’ performance for the detection of AF in both a general clinical population and an at-home settings, where the prevalence of AF would be much lower, should be considered. Additionally, the AF cohort studied here had an established diagnosis and therefore well-controlled ventricular rates. Future studies could confirm the Omron BPMs’ accuracy at higher ventricular rates which often occur when AF is first diagnosed.
BPMs remain only one method for population screening being utilized, as commercially-available “smart watches” and at-home ECG recording devices are becoming more commonplace. One comparative limitation of BPMs for screening is that they do not produce an ECG reviewable by a clinician. Therefore an “AF” alarm at home may be accurate (true positive), but still not lead to an immediate diagnosis of AF if the individual returns to normal sinus rhythm before an EKG or heart monitor can be obtained; as may occur with paroxysmal AF. However, BPMs are far more prevalent in homes around the world compared with ECG-based devices due to their affordability and their routine use for tracking blood pressure. Also, BPMs are already utilized at most visits with a healthcare provider, whereby utilizing a BPM with an accurate AF detection rhythm can offer AF screening without any added time or cost. Also, ECG-based screening at home has its own unique challenges with false positive results due to baseline artifact from motion or poor connectivity; whereas measuring pulse-to-pulse intervals (PPIs) with BPMs does not have these limitations.
It Is important to note that the machine learning algorithm incorporated into these novel Omron BPMs contributed to achieving their high diagnostic accuracy. This is supported by the consistent performance seen with two differing BPM/cuff combinations which share this algorithm. Machine learning is just one subset of the broader category of artificial intelligence (AI), where a device can independently improve its performance without additional programming. Defining atrial fibrillation by its irregular and unpredictable PPIs can be challenging. The Microlife BPM comparator uses a single mathematical index, the “irregularity index” – which is equal to the standard deviation divided by the mean of the measured pulse-to-pulse intervals (PPIs). The Omron BPMs’ AF detection algorithm utilizes this and over three hundred additional parameters through the machine-learning algorithm AdaBoost [11]. Individual parameters can draw from PPI intervals, but also consider heart rate and pulse amplitude in varying equations. AdaBoost incorporates all of these mathematical indices and adjusts their diagnostic weight in an iterative “decision-tree” model [Figure 1].
Conclusion
The ability of the studied Omron BPMs to incorporate AI technology to improve their diagnostic accuracy for detecting atrial fibrillation, with no external computing power or additional components, is a novel innovation.