The Respiratory Signature: A Novel Concept to Leverage Continuous Positive Airway Pressure Therapy as an Early Warning System for Exacerbations of Common Diseases such as Heart Failure.

ABSTRACT
Each night millions of patients use continuous positive airway pressure (CPAP) to treat obstructive sleep apnea (OSA). To facilitate monitoring of treatment success, modern CPAP machines routinely record and analyze the respiratory signal in near real-time and submit some of these data to the manufacturer's centralized cloud server. Some of the conditions frequently associated with OSA such as heart failure or chronic obstructive pulmonary disease result in characteristic changes of the respiratory signal ("signatures"), especially during exacerbations. Thus, this infrastructure could be leveraged to detect changes in patients' health status facilitating early interventions. To illustrate this concept, we present and discuss the case of a patient with OSA who showed abrupt changes in his breathing pattern (increase in periodic breathing and machine-detected obstructive apneas) from 10 days prior until 8 days after a hospitalization for acute heart failure exacerbation.


I NTRO DUCTI O N
Acute exacerbation of chronic congestive heart failure (CHF) is one of the most common causes of hospital admissions 1 thereby posing a major burden on the health care system. 2 For individual patients, CHF-related admissions herald an increased risk of future re-admissions and increased mortality. 3,4 Thus to break this vicious cycle, reduce health care costs and improve patients' long-term trajectory, early detection of worsening heart failure is critical to allow for interventions aiming to prevent full decompensation and need for inpatient management. Current practice relies largely on patients' self-monitoring of daily weights and alerting physicians of excessive changes, 5 which is suboptimal, 6 thus better detection methods (ideally based on physiologic measures) are warranted. Continuous positive airway pressure (CPAP) may provide such an opportunity in the subset of patients with heart failure who use CPAP for sleep apnea.
It has been well established that the pathophysiology of sleep apnea is generally multifactorial due to varying contributions from four different mechanisms or endotypes: compromised upper airway anatomy, dysfunction of pharyngeal dilator muscles, unstable ventilatory control (high loop gain), and low arousal threshold. 7 Coexistent sleep apnea is present in 50% to 70% of patients with CHF, 8,9 which is predominantly due to high loop gain manifesting as increased central apneas and periodic breathing. 10 This pattern is thought to be accentuated during CHF exacerbations but the time course and magnitude of these changes are unclear. 8 is CPAP, which is also used for central sleep apnea. 10 Because of insurance regulations and to guide clinical management, CPAP devices routinely record users' respiratory signal and analyze it in near real-time. The analysis results (and-in the case of Respironics machines-additionally some of the raw respiratory data) are then automatically uploaded to a central database (accessible online by physicians via proprietary systems including EncoreAnywhere for Philips-Respironics devices and AirView for ResMed, Inc. machines). Independent of data upload, most devices save months or years worth of raw respiratory signal data locally, overwriting older entries with newer ones once internal storage is full (accessible manually through SD cards).
We present a case illustrating how this infrastructure could be leveraged to develop an automated early warning system notifying health care providers with sufficient time to prevent full CHF decompensation (Figure 1).
Report of this case is permitted by Beth Israel Deaconess Medical Center IRB protocol 2016P000058.

R EPO RT O F CASE
A 69-year-old man with a complex history including remote bypass graft for coronary artery disease, chronic systolic heart failure (left ventricular ejection fraction 30% to 40%) on guideline directed therapy including beta blocker, spironolactone, Sacubitril-Valsartan and cardiac resynchronization therapy (fully paced due to prior complete heart block), and persistent atrial fibrillation on anticoagulation underwent overnight polysomnography in July 2017 revealing mild-moderate OSA with an apnea-hypopnea index of 12 events/h, a respiratory disturbance index of 25 events/h, and an oxygen saturation nadir of 82% (47% of total sleep time spent at ≤ 88%). Of note, no central events or periodic breathing were noted during this study and based on a successful titration he was started on auto-CPAP set at 7-10 cmH 2 O on August 3, 2017.
CPAP therapy was well tolerated with an average nightly use of about 4.5 hours. However, after a brief prodrome he was admitted to the hospital (cardiac critical care unit initially) from September 4-9, 2017 for acute respiratory failure secondary to decompensated heart failure (pulmonary edema on chest X-ray, N-terminal prohormone of brain natriuretic peptide [NT pro-BNP] 4,840 pg/mL) possibly related to dietary indiscretion. He initially required noninvasive ventilation but rapidly improved with IV diuresis leading to approximately 15 lbs of weight loss. In the weeks following discharge weight remained stable within 3 lbs of discharge weight, and NT-pro-BNP gradually decreased to levels in the 500s.
Review of his CPAP device data (Figure 2) is remarkable for a rise in detected periodic breathing from a baseline of < 5% to 10%, increasing to 10% to 20%, from August 25-28, and a further sharp rise to 60% to 70% during August 29-30, after which there was essentially no CPAP use until admission on September 4 perhaps due to periodic breathing-related CPAP intolerance. During admission from September 4-9 the patient used a hospital CPAP device rather than his own, thus no device data is available for this period. Following discharge periodic breathing gradually declined from 60% on September 10 to baseline levels after September 16. Similar patterns are visible in plots of machine-detected hypopnea and apnea indices. Notably, the obstructive rather than the central apnea index seems to track better with changes in periodic breathing. This is possibly because of nocturnal, rostral fluid shifts increasing upper airway compromise, 12 but likely also because of some misclassification of central as obstructive apneas by the proprietary machine algorithm, as periodic breathing is pathophysiologically linked to repeated central rather than obstructive events. Central apneas are defined by the absence of respiratory effort during the event; unable to measure this effort directly, CPAP devices use airway patency as a proxy (eg, by administrating a brief square-wave pressure pulse of 2 cmH 2 O after ~6 seconds of apneic flow with a biphasic flow response suggesting patency), 13 but many central apneas are associated with a (gradual) airway closure 14 probably explaining the poor correlation between central apneas based on polysomnography versus machine-detected clear airway apneas. 13

DISCUSSI ON
This real-world example demonstrates that CHF exacerbation may result in respiratory signal changes as early as 10 days prior to admission, suggesting that there may be time for (1) recognition of an impending change in clinical status and (2) intervention by clinicians to intervene. These data suggest two possible candidates for the "respiratory signature" of acute CHF exacerbations: the machine-detected periodic breathing or obstructive apnea index. In this case periodic breathing is elevated from 10 days prior until 8 days after the hospitalization and may thus allow somewhat earlier prediction of heart failure exacerbations than the obstructive apnea index, which only rises 5 days prior to admission; on the other hand, the obstructive apnea index normalizes earlier and seems to have less baseline fluctuations, thus it may be a more specific marker and also detect immediate recurrences earlier (in this case from day 3-8 after hospitalization while periodic breathing is still elevated and thus "refractory," although, theoretically changes in the rate of periodic breathing "decay" may possibly be similarly informative).
The traditional gold standard for heart failure outpatient monitoring is a daily weight. We note that during the recovery from hospitalization phase of this patient's illness there was no major change in weight, yet breathing metrics did change. Thus, these metrics might be more sensitive than gross changes in weight.
However, both of these machine-detected measures likely suffer from under-detection of true events, 15 potentially explaining the sudden rise in periodic breathing from day 6 to day 5 prior to hospitalization and possibly limiting their (1) The CPAP device analyzes the respiratory signal each night and sends the raw data as well as the analysis results to a central cloud server (2). If the "respiratory signature" (eg, sudden increase of periodic breathing) of an impending hospitalization for acute heart failure exacerbation is detected by an algorithm then the patient's physician (eg, cardiologist, sleep provider, or primary care provider) is notified (3), who can then then contact the patient (4) in order to intervene at a time when the exacerbation may still be reversible (eg, increase of diuretics). Note that the detection algorithm could be either executed in the device or on the cloud server. An alternative, simpler variant may be to send alerts directly to the patient with the recommendation to contact his/her treating physician. Element (1), "Máquina CPAP" originally created by "PruebasBMA", reused without changes under CC BY-SA 3.0 license. CPAP = continuous positive airway pressure.
predictive accuracy for impending heart failure exacerbations. Increased transparency regarding algorithms used to detect residual events could help maximize predictive ability, especially among various device manufacturers. Of course, larger scale studies are needed to assess the value of these metrics, but an interesting alternative may be to measure loop gain directly from the respiratory signal, 16,17 as it is the underlying driver of both periodic breathing and central apneas (as mentioned above the obstructive apnea index in this case likely reflects an increase in central rather than obstructive apneas) and is probably the metric that is most strongly correlated with changes in cardiac function.
One challenge would be that loop gain (and thus periodic breathing and central apneas) can increase due to other factors, most importantly due to atrial fibrillation. 18 But by focusing on relative rather than absolute changes in these metrics the issue of stably elevated loop gain from persistent atrial fibrillation could likely be addressed, and "false-positive" alerts due to acute or paroxysmal atrial fibrillation episodes (including those causing CHF exacerbation) are arguably events that could benefit from alert-triggered medical attention as well (ie, even if specificity for CHF exacerbation were to be limited, the majority of alerts would likely signal a significant clinical problem). Furthermore, due to "cardiopulmonary coupling," 19 it is likely that there are subtle differences in respiratory signal changes between a heart failure exacerbation and acute atrial fibrillation-something that could for example be explored with unsupervised machine learning techniques.
In the inpatient setting electronic surveillance to detect emerging illnesses is frequently employed (eg, acute respiratory distress syndrome or sepsis), 20,21 but in the outpatient world such approaches are largely being hampered by lack of reliable data. One may contest that there is an ever-expanding consumer market of wearables monitoring physiologic data, but validation data of devices are often sparse thus precluding meaningful interpretation by health care providers, and manufacturers of such devices generally avoid diagnostic claims for medicolegal reasons. 22 CPAP on the other hand is probably the most commonly used medical grade wearable device, generating individual level big data while treating millions of patients with OSA every night. For instance, assuming that 50% to 70% of the 5.8 million patients with CHF in the United States have sleep apnea, 23 of which perhaps 15% are diagnosed, then-even with a conservative Figure 2-Changes of machine-detected parameters derived from the respiratory signal before and after a hospitalization for acute heart failure.
Figure shows machine-detected periodic breathing (percent per night), apnea and hypopnea indices (events per hour) plotted over a 2-month period (from 8/1/2017-9/30/2017) in the upper, middle and lower box, respectively (download from Philips-Respironics EncoreAnywhere portal). Patient was admitted for acute heart failure exacerbation from 9/4-9/9 (red shaded); note how periodic breathing is elevated from 10 days prior until 8 days after the hospitalization (yellow shaded) and may thus serve as a "respiratory signature" allowing early prediction of heart failure exacerbations; the obstructed airway apnea index follows a similar pattern, but rises later and normalizes earlier with less baseline fluctuations and thus may reflect a less sensitive but more specific marker which may also detect recurrences more rapidly again (in this case from day 3-8 after hospitalization while periodic breathing is still elevated and thus "refractory").
adherance to CPAP of 50%-there are about 200,000 patients with CHF using CPAP on a regular basis. If positive pressure therapy becomes a clinical standard for heart failure relatedapnea, the number of users may increase by several fold.
There are several important limitations to consider. First, the patient in the present case had multiple reasons for elevated loop gain and thus periodic breathing/central apneas (eg, low ejection fraction, atrial fibrillation); in a patient with milder disease it may be harder to detect a change in the respiratory signal. Second, while many health care providers (including the authors) feel it is helpful to treat sleep-disordered breathing in patients with heart failure via CPAP, based on recent highprofile trials (eg, CANPAP 24 and SERVE-HF 25 ) there has been much debate whether certain CPAP devices could be harmful in this setting, which ultimately could limit availability of CPAP data in patients with heart failure. Third, while prevalence estimates in stable (~40%) 9 versus decompensated (75%) 8 patients with CHF suggest that heart failure decompensation results in an increase of periodic breathing, 11 direct observations to corroborate this relationship are lacking.
Ultimately, we believe this concept could potentially be used to develop an Early Warning System for Acute Heart Failure Exacerbation among Positive Airway Pressure Users (SAFE-PAP); as a next step towards this vision we propose an observational study to better characterize temporal changes in the respiratory signal in a cohort of patients with CHF with an acute decompensation. Once a clear "signature" has been defined, a randomized controlled trial could then test if signature-triggered notification as outlined in Figure 1 reduces patient-important outcomes such as admissions.
In conclusion, CPAP device data may provide an ideal opportunity to develop an early warning system for acute heart failure exacerbation in a large group of patients, but the approach outlined in Figure 1 is also scalable, as the same methodology could be applied to monitor patients with OSA who have other important comorbidities with "respiratory signatures," eg, chronic obstructive pulmonary disease/asthma exacerbations which may be heralded by gradually longer expiration times, or slow/ataxic breathing as a marker of toxicity in patients using narcotics chronically. 26 We speculate that the use of CPAP as a "health monitor" may improve outcomes of monitored symptoms, and also exert a positive effect on CPAP adherence through increased patient engagement, and thus may improve OSA treatment itself.