Performance of Atrial Fibrillation Detection in a New Single-Chamber ICD

Transcription

1 DEVICES Performance of Atrial Fibrillation Detection in a New Single-Chamber ICD ABHISHEK DESHMUKH, M.D.,* MARK L. BROWN, PH.D., ELISE HIGGINS, B.S., BRIAN SCHOUSEK, B.S., ATHULA ABEYRATNE, PH.D., GIOVANNI ROVARIS, M.D., and PAUL A. FRIEDMAN, M.D.* From the *Mayo Clinic, Rochester, Minnesota; Medtronic plc, Mounds View, Minnesota; and Azienda Ospedaliera San Gerardo, Monza, Italy Background: Patients with implantable cardioverter defibrillators (ICDs) often have a history of atrial fibrillation (AF) or will develop AF after device implant. Optimal management of ICD patients includes early diagnosis of AF and monitoring of AF burden. We evaluated the performance of an algorithm for monitoring AF in single-chamber ICDs. Methods: The RR interval variability-based detection algorithm determines RR variability by creating a Lorenz plot of the change in RR intervals for the most recent interval pair versus the previous interval pair. A new plot is created every 2 minutes and the AF evidence score of the plot is computed. Patient RR interval data from several Holter databases were pooled to test the performance of the AF detection algorithm. Results: In total, 187 recordings from 186 patients were used to assess the performance of the AF detection algorithm integrated into a single-chamber ICD by comparing the ICD detection results to the Holter annotated truth. Thirty-five of 186 patients had a total of 94 AF episodes in their Holter recordings lasting a total of hours (mean episode duration 7.2 hours). The generalized estimating equations-adjusted estimate of episode sensitivity was 94.8% with 95% lower confidence limit of 87.2%. Gross duration sensitivity was 95.0% for AF episodes of at least 6 minutes duration with gross duration specificity of 99.6%. Conclusion: This RR interval-based AF detection algorithm performs well with high sensitivity and specificity. Integration of this algorithm into single-chamber ICDs will help monitor and detect AF, thus facilitating optimal therapy to prevent AF-related sequelae. (PACE 2016; 39: ) atrial fibrillation, single-chamber ICD, algorithm Introduction The worldwide prevalence of atrial fibrillation (AF) is 33 million and growing with Disclosures: Mark L. Brown, Elise Higgins, Brian Schousek, Athula Abeyratne are employees of and receive compensation from Medtronic. Abhishek Deshmukh: none. Giovanni Rovaris: Medtronic. Paul A. Friedman: Research Support Medtronic, Biotronik; Intellectual Property Rights Bard EP, Medical Positioning, Inc., Aegis Medical, NeoChord, Preventice, Sorin; Speaker or Consultant Bard, Biotronik, Leadex, Sorin, Boston Scientific, Helical Solutions. Address for reprints: Abhishek Deshmukh, M.D., Mayo Clinic, 200 First Street SW, Rochester, MN 55901; deshmukh.abhishek@mayo.edu This is an open access article under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs License, which permits use and distribution in any medium, provided the original work is properly cited, the use is noncommercial and no modifications or adaptations are made. Received April 4, 2016; revised June 13, 2016; accepted June 26, doi: /pace significant implications for healthcare utilization due to the increased risk of stroke, worsening heart failure, and recurrent hospitalizations. 1,2 Most new onset AF is discovered when patients report symptoms such as palpitations, fatigue, or syncope; however, many may be asymptomatic. Silent AF is found in approximately 30% of cryptogenic stroke patients when monitored with a cardiac implantable electronic device (CIED) that allows for continuous AF monitoring. 3 Electrocardiogram (ECG) confirmation of potential AF symptoms often depend on snapshot recording methods such as ECG examination, Holter monitoring, transtelephonic-ecg, or event monitoring. These snapshot techniques are inadequate and often underreport the true frequency of AF. 4 Approximately, 25% of implantable cardioverter defibrillator (ICD) recipients have a history of AF at implantation 5 and 20% of single-chamber ICD patients will have new-onset AF within 24 months of implantation. Typical ICD recipients may have multiple stroke risk factors in the presence of AF, including 2016 The Authors. Pacing and Clinical Electrophysiology published by Wiley Periodicals, Inc. PACE, Vol. 39 October

2 DESHMUKH, ET AL. heart failure, hypertension, advanced age, and diabetes. Single-chamber ICDs only detect AF if the ventricular response exceeds the programmed ventricular arrhythmia detection rate. Given the high-stroke risk in most patients with an ICD, the underdetection of episodic AF with ventricular rates below the detection cutoff by single-chamber ICDs is a major limitation. Integrating floating atrial electrodes into a ventricular lead may overcome this limitation, but at the expense of added lead complexity and uncertain atrial sensing. The addition of a separate atrial lead (dual-chamber ICD) reliably permits AF detection, but dual-chamber ICDs are only indicated for patients with a pacing indication. Implantable cardiac monitoring (ICM) devices, such as Reveal R XT and Reveal LINQ TM (Medtronic, Minneapolis, MN, USA), detect slowly conducting AF without an atrial lead based on the signature of RR interval variability. 6 This RR interval variability-based detection algorithm could plausibly be integrated into a single-chamber ICD to enable AF detection without the need for an atrial lead or floating atrial electrodes. We hypothesized that integration of an interval-based algorithm would permit reliable detection of AF with a controlled ventricular response in single-chamber ICDs using a standard ventricular lead. Methods Databases Patient data from several Holter databases were pooled to test the performance of an ICD AF detection algorithm, described below. Two publically available Holter databases were used: MIT- BIH Arrhythmia Database and MIT-BIH Atrial Fibrillation Database. The MIT-BIH Arrhythmia Database contains 48 half-hour excerpts of twochannel ambulatory ECG recordings, obtained from 47 subjects studied by the BIH Arrhythmia Laboratory between 1975 and These half-hour excerpts were selected at random from 4, hour ambulatory ECG recordings. The MIT-BIH Atrial Fibrillation Database contains hour two-channel ambulatory ECG recordings from 25 patients. Together the MIT databases provide 73 records from 72 patients. The Medtronic databases used consist of 131 Holter recordings from 129 patients. All patients were implanted with a single- or dualchamber ICD and standard ICD leads. The Medtronic data were collected from patients using DR220/DR190/DR180 Holter recorders (North- East Monitoring, Maynard, MA, USA) modified to collect telemetered data from Medtronic ICDs. Along with surface ECG, these Holters are able to capture the intracardiac electrogram (EGM) and RR intervals recorded by the ICD. The recorded RR intervals were used for algorithm evaluation. Combining data from the Medtronic databases and the MIT Arrhythmia Databases provided 204 recordings from 201 patients for the analysis. Algorithm Description An AF detection algorithm was previously developed for use in an ICM (Reveal XT). 6,8 This algorithm was incorporated, with modifications described below, into single-chamber ICD firmware and installed in a system emulator for this study. Every 2 minutes the algorithm uses a Lorenz Plot to analyze RR intervals sensed from a near-field EGM (RVtip to RVring or RVtip to RVCoil) and calculates an AF evidence score. The Lorenz Plot graphs the current RR (the difference in adjacent RR intervals) against the previous RR. The AF evidence score is based on the scattering of events in the Lorenz Plot. 6 The AF evidence score is low if the RR interval differences are clustered around the origin and the AF evidence score is high when RR interval differences are highly scattered and variable (Fig. 1). Events classified as ectopic reduce the AF evidence score, limiting AF false positives due to persistent ectopy. (Appendix) At the end of each 2-minute period, the AF evidence score is compared to a threshold and the rhythm is classified as AF, NO AF, or UNCLASSIFIED. An UNCLASSIFIED classification arises when there is a high level of ventricular pacing or a high rate episode (e.g., ventricular tachycardia [VT]/ventricular fibrillation, supraventricular tachycardia [SVT], or oversensing) is detected. Unlike the Reveal XT, the ICD-based AF detection algorithm requires several 2-minute blocks of AF to classify the rhythm as AF to improve the positive predictive value (PPV). Typically, AF will be detected when there are three 2-minute periods classified as AF as shown in Figure 2. Another key difference between the Reveal XT and ICD-based AF detection algorithms is the signal source subcutaneous ECG for Reveal and intracardiac EGM for ICD. RR timing from intracardiac EGM is inherently more accurate than ECG-based timing resulting in improved PPV. Adjudication Method For the MIT-BIH database, the data are presented as a sequential list of RR intervals along with an annotation file showing the location of AF episodes. For the Medtronic data, a Holter technician initially identified recordings that contained AF episodes greater than or equal to 6 minutes. Rhythm experts used one channel of surface ECGs (e.g., Lead II) and up to two channels 1032 October 2016 PACE, Vol. 39

3 ATRIAL FIBRILLATION DETECTION IN ICD Figure 1. ECG/EGM and corresponding Lorenz plots. The top ECG figure is of normal sinus rhythm with PVCs. The corresponding Lorenz plot on the left shows that most pairs of RRs are located at the origin since RR intervals are very consistent. The PVCs show up as symmetric pairs away from the origin because of the compensatory pause after a PVC. In the bottom ECG, the RR interval varies on each beat consistent with AF. The corresponding Lorenz plot on the right illustrates the scattered filling of the plot which results in a high AF evidence score. AF = atrial fibrillation; ECG = electrocardiogram; EGM = electrogram; PVC = premature ventricular contraction. of ICD EGMs to adjudicate the cardiac rhythm for the entire duration of the recording. If there were discrepancies between the Holter technician and rhythm expert, a second rhythm expert determined the classification for the discrepant episode. Testing Configuration Recorded intervals were played into an emulator running released product ICD firmware enhanced with the ICD AF detection algorithm. All AF detection calculations were run at nominal settings (balanced AF sensitivity, ectopy rejection ON, and record EGM for episodes 6 minutes). The AF Sensitivity threshold determines the amount of AF evidence required to classify a 2-minute period as AF. Balanced sensitivity sets the threshold in the middle of the range. With ectopy rejection programmed ON, the AF evidence score will be reduced if the presence of ectopy is detected. As the record EGM for parameter was programmed to episodes 6 minutes, the ICD stored only AF episodes lasting 6 minutes or more. Each 2-minute segment of the Medtronic Holters was processed and excluded if any of the following criteria were met: 1 Significant telemetry loss (as annotated by Holter technician). 2 RV pacing >5%. 3 Atrial paced events 2. 4 Bad telemetry markers 2. PACE,Vol.39 October

4 DESHMUKH, ET AL. Figure 2. Episode detection. The episode start and end are indicated above. Each 2-minute block is labeled with the result from evaluation of a new Lorenz plot. Three 2-minute blocks determined to be AF results in an AF episode detection. A single 2-minute block determined to be no AF results in episode termination or no AF episode detection if it occurs before three AF blocks. When an episode is detected, 30 seconds of EGM are stored from the beginning of the 2-minute block that completed detection. The arrows below the figure indicate the diagnostics that are collected for the AF episode. AF = atrial fibrillation; EGM = electrogram. Figure 3. Definitions for Holter analysis by episode. The hash-filled boxes represent true AF episodes on the Holter; the dark boxes represent AF episodes detected by the AF detection algorithm. TPs are true-positive detections for evaluating sensitivity. These are the true AF episodes that are detected (at least once) by the detection algorithm. TPps are true-positive detections for positive predictive value (PPV). These are the AF detections that correspond to a true AF episode. Note that the relationship of TPp and TPs need not be 1:1. AF = atrial fibrillation. 5 Number of possible bad telemetry segments (RR interval >2 seconds) 2. Performance Analysis Figures 3 and 4 illustrate the definitions used to classify various combinations of AF rhythm truth on the Holter recording and the episode reporting by the AF detection algorithm. Definitions for analysis of continuous recording and detection are necessarily a little more complicated than for discrete event detection. Results A total of 204 patient recordings from 201 patients were selected for inclusion in this analysis from the combined MIT and Medtronic databases. Five of the MIT recordings were excluded due to the presence of ventricular pacing. After applying the exclusion criteria and removing any Medtronic Holters whose remaining recording duration was less than 6 minutes, 187 recordings from 186 patients remained with statistics as shown in Table I. These 187 recordings were used to assess the performance of the 1034 October 2016 PACE, Vol. 39

5 ATRIAL FIBRILLATION DETECTION IN ICD Figure 4. Definitions for analysis of duration statistics. It is common for the start and stop time for a true episode and a detected episode do not precisely align. This figure provides the definitions for true-positive duration, false-positive duration, false-negative duration, and truenegative duration for the various scenarios that can occur when processing the Holter recordings. Table I. Statistics for Holter Recordings: 187 Recordings from 186 Patients Lasting 2,710 Hours Contained 94 Episodes of AF with Average Duration of 7.2 Hours from 35 Patients Recording Statistics Recordings 187 Patient count 186 Total Holter time (hours) 2,710 Patients with AF 35 AF episode count 94 Total AF episode time (hours) Mean AF episode time (hours) 7.2 Median AF episode time (hours) 5.4 AF = atrial fibrillation. AF detection algorithm integrated into a singlechamber ICD by comparing the ICD detection results to the Holter annotated truth. Episode detection sensitivity is the percentage of true AF episodes (longer than 6 minutes) in the Holter recordings that were detected by the ICD AF detection algorithm. Eighty-nine of the 94 AF episodes were detected for a gross sensitivity of 94.7%. The generalized estimating equations (GEE) adjusted estimate of sensitivity was 94.8% (n = 35 patients) with a 95% lower confidence limit of 87.2%. Episode PPV is the percentage of device detected AF episodes that are true AF episodes. Eighty-nine of 112 AF episode detections were true AF for a gross PPV of 79.5%. The GEE-adjusted estimate of PPV was 74.7% (n = 44 patients) with a 95% lower confidence limit of 63.0%. Duration sensitivity measures the percentage of time that true AF is detected by the AF detection algorithm. The ICD AF detection algorithm detected of hours of AF for a gross duration sensitivity of 95.0%. The patient average duration sensitivity was 93.6% in 35 patients with a 95% lower confidence limit of 88.4%. Duration specificity measures the percentage of the time that true non-af is classified as non- AF by the AF detection algorithm. The ICD AF detection algorithm detected 9.4 hours of AF during 2,459.5 hours of true non-af for a gross duration specificity of 99.6%. The patient average duration specificity was 96.3% in 178 patients with a 95% lower confidence limit of 94.3%. Duration PPV measures the percentage of time the ICD AF detection algorithm is detecting AF that is truly AF. Of hours detected as AF, hours were true AF for a gross duration PPV of 96.2%. The patient average duration PPV was 66.2% in 44 patients with 95% lower confidence limit of 55.5%. Results are summarized in Table II. Discussion The ability of a single-chamber ICD to detect episodic, asymptomatic AF, even when it is slowly conducted, is important given the relative frequency of new-onset AF in ICD recipients PACE,Vol.39 October

6 DESHMUKH, ET AL. Table II. AF Detection Performance Results: Episode Detection Sensitivity was 94.7% and Episode Detection PPV was 79.5%; Duration Sensitivity was 95.0% and Duration Specificity was 99.6% Gross Average (%) Patient Average (%), (N, 95% Lower CI) Episode (35, 87.2) detection sensitivity Episode (44, 63.0) detection PPV Duration (35, 88.4) sensitivity Duration (178, 94.3) specificity Duration PPV (44, 55.5) Patient averages of duration sensitivity, specificity, and PPV are unadjusted because there was only one observation per Holter. One-sided lower confidence limits of the mean are based on t-distribution. AF = atrial fibrillation; CI = confidence interval; GEE = generalized estimating equations; PPV = positive predictive value. GEE estimate assuming multiple episodes within a Holter are correlated according to compound symmetry correlation structure. with no prior history of AF. We tested the ability of an algorithm developed for an ICM and adapted for single-chamber ICDs to diagnose AF. The RR interval-based algorithm performed well when evaluated retrospectively using expertreviewed Holter recordings. We found gross episode sensitivity of 94.7% and gross duration sensitivity of 95.0% for AF episodes of at least 6 minutes duration with very high gross duration specificity (99.6%). AF is associated with no obvious symptoms and no noticeable short-term sequelae in at least one-third of patients. In some cases, asymptomatic AF is revealed only after an event such as stroke or congestive heart failure has occurred as initial manifestation. About 25 30% of patients presenting with strokes have AF that was not previously recognized. 9 Subclinical AF can be detected by various cardiac monitoring methods, including external surface monitoring (e.g., standard 12-lead electrocardiogram, ambulatory Holter monitors, event monitors) and by CIEDs (e.g., implantable cardiac monitors, dual-chamber pacemakers, dual-chamber ICDs, cardiac resynchronization therapy [CRT] devices), many of which enable remote monitoring. Currently most single-chamber ICDs do not provide the capability of detecting AF with a ventricular response below the VT detection rate. Given the increasing use of high ventricular detection rates, 10,11 it is likely that most episodes of AF are undetected. In this study, we evaluated an algorithm designed for use in a single-chamber ICD. The clinical goal of an AF detection algorithm in an ICM is to determine whether the patient has AF and quantify the AF burden. Systematic reviews assessing the detection of AF with external ECG monitoring in patients after cryptogenic stroke have shown a detection rate of newly diagnosed AF of 5 20%. 4 Observational studies using an ICM in similar populations have suggested a detection rate of approximately 25%. Continuous monitoring in patients at risk for developing AF is not realistic in the absence of another indication for an implantable device; ICD recipients have such an indication. 12 AF diagnosis should be quite sensitive in order to avoid undesirable consequences. Continuous monitoring with an ICM has been shown to be times as sensitive for detecting AF compared to various snapshot methods. Integration of an interval-based algorithm into a single-chamber ICD for detection and diagnosis of AF should provide diagnostic yield similar to other continuous monitors. The AF detection algorithm under test only detected episodes of AF with duration of at least 6 minutes. The minimum duration of AF that is clinically important is not known. In the ASSERT study, patients with CIEDs with subclinical atrial arrhythmias defined as an atrial rate >190 beats/min for at least 6 minutes had a 2.5 times increase risk of stroke or systemic embolism, 13 suggesting that detection of 6-minute episodes is clinically meaningful. The implications of detection of even brief AF episodes with a single-chamber ICD include (1) increased detection of AF leading to identification of patients at risk for stroke who may require anticoagulation therapy or left atrial appendage closure, (2) AF detection may help in management of antiarrhythmic therapy and monitoring of recurrence post-nonpharmacological treatment, and (3) assessment of rate during AF episodes may help in titration of atrioventricular node blocking medications and prevent hemodynamic deterioration, the development of ventricular dysfunction, or inappropriate shocks should AF intermittently conduct rapidly. This, in turn, may help in preventing AF related resource utilization. Certain limitations need to be acknowledged. The algorithm is AF specific and will not account for atrial flutter and atrial tachycardia. Ventricular pacing less than 30% was the cutoff for data inclusion in this study, so results for V pacing 1036 October 2016 PACE, Vol. 39

7 ATRIAL FIBRILLATION DETECTION IN ICD >30% are not available. AF episodes less than 6 minutes were not included. In conclusion, an RR interval-based algorithm designed to be integrated into a single-chamber ICD performs well for detection of AF. Integration of this algorithm in standard ICD leads may enhance AF detection resulting in early intervention. References 1. Chugh SS, et al. Worldwide epidemiology of atrial fibrillation: A Global Burden of Disease 2010 Study. Circulation 2014; 129: Patel NJ, et al. Contemporary trends of hospitalization for atrial fibrillation in the United States, 2000 through 2010: Implications for healthcare planning. Circulation 2014; 129: Glotzer TV, Ziegler PD. Silent atrial fibrillation as a stroke risk factor and anticoagulation indication. Can J Cardiol 2013; 29:S14 S Seet RC, et al. Prolonged rhythm monitoring for the detection of occult paroxysmal atrial fibrillation in ischemic stroke of unknown cause. Circulation 2011; 124: Schloss EJ, et al. How common is new onset atrial fibrillation in single chamber ICD patients? Sub-analysis from the painfree SST Study. Circulation 2015; 132:A17946 A Sarkar S, et al. A detector for a chronic implantable atrial tachyarrhythmia monitor. IEEE Trans Biomed Eng 2008; 55: Goldberger AL, et al. Physiobank, physiotoolkit, and physionet components of a new research resource for complex physiologic signals. Circulation 2000; 101:e215 e Purerfellner H, et al. P-wave evidence as a method for improving algorithm to detect atrial fibrillation in insertable cardiac monitors. Heart Rhythm 2014; 11: Lin HJ, et al. Newly diagnosed atrial fibrillation and acute stroke. The Framingham Study. Stroke 1995; 26: Gasparini M, et al. Effect of long-detection interval vs standarddetection interval for implantable cardioverter-defibrillators on antitachycardia pacing and shock delivery: The ADVANCE III randomized clinical trial. JAMA 2013; 309: Moss AJ, et al. Reduction in inappropriate therapy and mortality through ICD programming. N Engl J Med 2012; 367: Cotter PE, et al. Incidence of atrial fibrillation detected by implantable loop recorders in unexplained stroke. Neurology 2013; 80: Healey JS, et al. Subclinical atrial fibrillation and the risk of stroke. N Engl J Med 2012; 366: Appendix Description of false-positive events. Description of record during false-positive detection Count Many short AF episodes in record, some 9 of which were detected as AF by algorithm Signal noise plus arrhythmias such as 4 bigeminy, trigeminy, PVC Entire record-adjudicated atrial flutter, 3 algorithm-detected AF episode Arrhythmias such as bigeminy trigeminy 2 couplets Atrial flutter interspersed with AF 2 Algorithm detected as AF Signal noise 1 Short AF episodes, plus intermittent 1 telemetry loss interspersed with AF episodes >6 minutes Short runs of SVT, PVC. Persistent ectopy 1 Total false positives 23 PACE,Vol.39 October