Skip to main content

m38 · Wearables and implantable monitors

Warning

Education only. Not for diagnosis, treatment or any clinical decision.

In 60 seconds

Smartwatches and fitness trackers watch the pulse with light, and implanted monitors watch the heart’s electrical activity for years. Both run detectors around the clock, so even a small false alarm rate per check adds up to many alerts. Wearables answer this by demanding several irregular windows in a row before they notify (Lubitz et al. 2022); clinics that follow implanted monitors answer it with better detection algorithms and with people who review every alert (Neiman et al. 2024). This module derives three things an engineer needs here: how pulse timing differs from heartbeat timing, what “k windows in a row” buys and costs, and how a triage filter trades review time against missed events.

From the field

  • Large wearable studies measured predictive value, not only accuracy. In the Stanford-led Apple Heart Study, the watch’s PPG sensor measured the pulse intermittently while the wearer was at rest. 2,161 of 419,297 participants (0.52%) received an irregular-pulse notification over a median of 117 days. Among those who returned an analyzable ECG patch, AF was present in 34% (Perez et al. 2019).
  • The rule matters as much as the model. The Fitbit Heart Study notified only after 11 consecutive irregular 5-minute pulse windows. Of 225 participants with another such detection while wearing an ECG patch, 221 had AF at the same time: a positive predictive value of 98.2% (Lubitz et al. 2022).
  • Implanted monitors generate review work. Historically about 75% of implantable loop recorder (ILR) alerts were false positives. In one surveillance program, two newer remotely reprogrammable ILR models with AI-based detection had 18.0% and 38.3% false-positive alerts (Neiman et al. 2024). A systematic review found false alarms in 59.7% of remote transmissions, most often for AF, mainly triggered by premature beats (Covino and Russo 2024).
  • Device electrograms can be classified too. A Stanford team trained deep networks on intracardiac electrograms to tell AF from atrial tachycardia, which devices tend to lump together as high-rate events, and compared them with rate and regularity rules (Rodrigo et al. 2022). A review covers AI across implanted devices: alert filtering, event prediction and programming (Antoun et al. 2025).
  • Agents are arriving. An agent that called signal-processing tools estimated heart rate from PPG better than language models reading the numbers directly (Feli et al. 2025). Another study used a language model to aggregate simulated pacemaker remote-monitoring data and estimate battery level (Iwai et al. 2025). Module m44 covers how such agents are built and kept safe.

Derivations

Symbols: \(RR_i\) the \(i\)-th interval between R peaks (s), \(\Delta_i\) the delay from R peak \(i\) to its pulse at the wrist (s), \(\mathrm{IBI}_i\) the interval between pulses \(i\) and \(i+1\) (s), \(q\) the chance that a window in regular rhythm is judged irregular, \(p\) the same chance in AF, \(k\) the number of irregular windows required in a row, \(n\) windows per day, \(L\) windows in an AF episode.

Pulse timing is not heartbeat timing

Photoplethysmography (PPG) shines light into the skin and measures how much returns. The pulsatile part of the signal follows the change in blood volume with each heartbeat, on top of a slowly varying baseline (Allen 2007). Heart rate comes from the mean interval:

\[ \mathrm{HR} = \frac{60}{\overline{\mathrm{IBI}}} \quad \text{beats per minute} \tag{1}\]

Each pulse reaches the wrist a delay \(\Delta_i\) after its R peak, and that delay varies from beat to beat. Subtracting successive pulse times gives an exact identity:

\[ \mathrm{IBI}_i = RR_i + \left(\Delta_{i+1} - \Delta_i\right) \tag{2}\]

If the delays are stationary, with variance \(\sigma_\Delta^2\) and lag-one correlation \(\rho\), and independent of the RR intervals, then

\[ \operatorname{Var}(\mathrm{IBI}) = \operatorname{Var}(RR) + 2\sigma_\Delta^2\,(1 - \rho) \tag{3}\]

Pulse intervals are always at least as variable as heartbeat intervals. Delay jitter makes a regular rhythm look a little irregular, and quick changes in delay (\(\rho\) near 0) are worse than slow ones (\(\rho\) near 1). Motion adds missed and extra pulses on top. A pulse-based irregularity detector must allow for both.

Key takeaways

  • PPG measures the pulse, not the heartbeat; its intervals add the change in pulse delay.
  • Delay jitter inflates variability by \(2\sigma_\Delta^2(1-\rho)\), so regular rhythm can look irregular.
  • Thresholds tuned on ECG intervals do not transfer unchanged to PPG.

Many windows a day: what “k in a row” buys

A wearable checks its data in windows, for example 5-minute windows, about 288 a day. Suppose each window in regular rhythm is wrongly judged irregular with probability \(q\), independently. If the device notifies after \(k\) irregular windows in a row, the expected number of false notifications in \(n\) windows (counting each run once) is

\[ \mathbb E[\text{runs}] = q^{k}\left[1 + (n - k)(1 - q)\right] \tag{4}\]

A run of at least \(k\) starts at the first window when the first \(k\) windows are positive, and at a later window when that window and the next \(k-1\) are positive and the window before it is not.

The same rule can miss a short AF episode. If an episode spans \(L\) windows, each flagged with probability \(p\), let \(a_m\) be the probability of no run of \(k\) in \(m\) windows. Then \(a_m = 1\) for \(m < k\), and

\[ a_m = \sum_{j=0}^{k-1} p^{j}(1 - p)\, a_{m-j-1}, \qquad P(\text{episode caught}) = 1 - a_L \tag{5}\]

since a sequence with no run of \(k\) must begin with \(j < k\) flagged windows followed by an unflagged one. Raising \(k\) cuts false notifications geometrically, as \(q^k\), but it also cuts the chance of catching episodes shorter than a few times \(k\) windows.

Exact

Assumptions

  • Model: Equation 4 and Equation 5, exact when windows are judged independently with fixed \(q\) and \(p\). Real windows are not independent: noise and motion come in bursts, and overlapping windows share data. Correlated false windows make false notifications more likely than Equation 4 predicts (lab part E measures this on Holter data).
  • All slider values are illustrative inputs. The 5-minute default window echoes the windows used in Lubitz et al. (2022); \(q\), \(p\) and the episode length are not taken from any study.
  • What would change the result: windows that overlap, a rule that is not “k in a row” (for example “k of the last m”), or episodes that start and stop mid-window.

Key takeaways

  • Over a day of windows, small per-window false rates become daily alerts.
  • “k in a row” divides false notifications by roughly \(1/q\) per extra window required.
  • It also misses short episodes; choose \(k\) knowing which episodes matter.

The review inbox

Every alert from an implanted monitor that reaches a clinic needs a person. With \(N\) patients, \(\lambda\) alerts per patient per day, a fraction PPV of them true and \(\tau\) minutes to review each, the daily load is

\[ A = N\lambda, \qquad A_{\text{false}} = N\lambda\,(1 - \mathrm{PPV}), \qquad H = \frac{N\lambda\,\tau}{60}\ \text{hours} \tag{6}\]

A triage filter, whether a better on-device algorithm or a model at the clinic, keeps a share \(\kappa\) of true alerts and removes a share \(r\) of false ones. The alerts that remain have

\[ \mathrm{PPV}' = \frac{\kappa\,\mathrm{PPV}}{\kappa\,\mathrm{PPV} + (1 - r)(1 - \mathrm{PPV})} \tag{7}\]

and \(N\lambda\,\mathrm{PPV}(1-\kappa)\) true alerts a day are no longer seen by anyone. This is Bayes’ rule again (m35), with the raw PPV in the role of prevalence, \(\kappa\) as sensitivity and \(r\) as specificity.

Exact

Assumptions

  • Model: Equation 6 and Equation 7, exact for the stated rates. Alerts are treated as independent; review time per alert is constant.
  • The default raw PPV of 25% corresponds to the historical figure of about 75% false-positive ILR alerts quoted by Neiman et al. (2024). Clinic size, alert rate, review minutes and the filter’s \(\kappa\) and \(r\) are illustrative inputs.
  • What would change the result: true alerts are not all equally important; a filter that hides a rare dangerous rhythm costs far more than one that hides a common benign one.

Key takeaways

  • Review time scales with alerts, not with patients who are ill.
  • A filter’s value is review time saved; its cost is true alerts no one sees.
  • Report both \(\kappa\) and \(r\), and which kinds of true alerts are lost.

Clinical correlate

NotePending clinical review

This section describes published findings only. It is not guidance for patient care, and it has not yet been reviewed by an EP physician.

False alerts from implanted loop recorders. An implantable loop recorder (ILR) is a small device placed under the skin that records the ECG for years and sends alerts when its detector fires. A systematic review of 12 reports covering 3,305 patients followed by remote monitoring found that false-positive alarms made up 59.7% of remote transmissions and affected about one patient in five. False AF alarms were the most common type, mainly due to premature atrial and ventricular beats (Covino and Russo 2024).

A retrospective study of a national surveillance program looked at two remotely reprogrammable ILR models with two-stage, AI-based detection (Neiman et al. 2024). Every alert was reviewed by staff. False positives were 18.0% of alerts for one model (239 alerts, 117 patients) and 38.3% for the other (300 alerts, 105 patients). After remote reprogramming following false alerts, median alerts per day fell from 0.13 to 0.03 and from 0.15 to 0.01.

In the terms of this module:

  • The burden is in the inbox. Equation 6 turns these percentages into hours.
  • Premature beats break interval rules. Irregular intervals are not specific to AF, the same failure seen in m37’s ectopy test.
  • Fewer alerts is not the whole story. A study must also report the true events that were not sent (Equation 7), which needs a reference such as continuous recording.

Visual summary

Schematic

SensorPPG or ECG Window detectorq, p per window k in a rownotify Triagekeep κ, remove r Human reviewτ minutes each delay jitter,motion premature beatslook irregular fewer false alerts,short episodes missed PPV up, truealerts hidden hours per dayN λ τ / 60 Every stage trades false alerts against missed events
Figure 1: The chain from sensor to reviewed alert, with what each stage adds or removes (gray).

Assumptions

How experts think

Before trusting an alert rate or a detector’s accuracy, ask:

  1. What does the sensor actually time? A PPG times pulses, not beats (Equation 2). Was the detector tuned on PPG or on ECG intervals?
  2. What is the unit of the result? Per window, per notification or per person? Lubitz et al. (2022) reports PPV per detection during patch monitoring; Perez et al. (2019) also reports AF in 34% of notified participants on a later patch.
  3. What did the rule cost? Requiring \(k\) windows (Equation 4) lowers false notifications but misses short episodes (Equation 5).
  4. Who reviews, and how long does it take? Turn rates into hours with Equation 6.
  5. What did the filter hide? Ask for the share of true alerts kept, by rhythm type, measured against a reference (Equation 7).
  6. Are the windows independent? Bursts of noise and runs of premature beats break the assumption behind Equation 4.

One sentence to keep: an always-on detector is judged by its alerts per day and the events it misses, not by its accuracy on one window.

Cheat sheet

Quantity Result
Heart rate (Equation 1) \(\mathrm{HR} = 60/\overline{\mathrm{IBI}}\) bpm
Pulse intervals (Equation 2) \(\mathrm{IBI}_i = RR_i + \Delta_{i+1} - \Delta_i\)
Pulse-interval variance (Equation 3) \(\operatorname{Var}(\mathrm{IBI}) = \operatorname{Var}(RR) + 2\sigma_\Delta^2(1-\rho)\)
False notifications (Equation 4) \(\mathbb E[\text{runs}] = q^k[1 + (n-k)(1-q)]\)
Episode caught (Equation 5) \(1 - a_L\), with \(a_m = \sum_{j<k} p^j(1-p)a_{m-j-1}\)
Review load (Equation 6) \(A = N\lambda\), false \(N\lambda(1-\mathrm{PPV})\), hours \(N\lambda\tau/60\)
After a triage filter (Equation 7) \(\mathrm{PPV}' = \kappa\mathrm{PPV}/[\kappa\mathrm{PPV} + (1-r)(1-\mathrm{PPV})]\); lost \(N\lambda\mathrm{PPV}(1-\kappa)\) per day

Published numbers used in this module

Study Design Result
Perez et al. (2019) 419,297 smartwatch users 0.52% notified; AF in 34% of notified participants with an analyzable patch
Lubitz et al. (2022) 455,699 tracker users; 11 consecutive irregular 5-min windows Detection PPV 98.2% (95% CI 95.5–99.5%) for concurrent AF
Neiman et al. (2024) ILRs in a surveillance program Historical false alerts about 75%; 18.0% and 38.3% for two newer models
Covino and Russo (2024) Systematic review, 3,305 ILR patients False alarms 59.7% of remote transmissions

Rules of thumb

  • Multiply per-window rates by windows per day before judging them.
  • Each extra window in “k in a row” divides false notifications by about \(1/q\).
  • Quote a filter’s kept-true share next to its removed-false share.

Interview questions

Try each question before you open its model answer. In a real interview, say which assumptions you are making.

Q1. Why are intervals between PPG pulses more variable than RR intervals, even in perfectly regular rhythm?

Model answer

Each pulse arrives a delay \(\Delta_i\) after its R peak, and that delay varies. Then \(\mathrm{IBI}_i = RR_i + \Delta_{i+1} - \Delta_i\) (Equation 2), so even constant RR intervals give varying pulse intervals. With stationary delays independent of RR, \(\operatorname{Var}(\mathrm{IBI}) = \operatorname{Var}(RR) + 2\sigma_\Delta^2(1 - \rho)\), which is Equation 3. Motion artifacts add missed and extra pulses on top.

Q2. Derive the expected number of false notifications for a “k windows in a row” rule.

Model answer

Count runs by where they start. A run of at least \(k\) starts at window 1 if windows 1 to \(k\) are all positive: probability \(q^k\). It starts at window \(i > 1\) if window \(i-1\) is negative and windows \(i\) to \(i+k-1\) are positive: probability \((1-q)q^k\), for \(n - k\) values of \(i\). Summing gives \(q^k[1 + (n-k)(1-q)]\) (Equation 4). Each run is counted once, at its start.

Q3. With 288 windows a day and \(q = 0.05\), roughly how many false notifications a day for \(k = 1\) and \(k = 3\)?

Model answer

\(k = 1\): \(0.05[1 + 287 \times 0.95] \approx 13.7\) a day. \(k = 3\): \(0.05^3[1 + 285 \times 0.95] \approx 1.25 \times 10^{-4} \times 271.75 \approx 0.034\) a day, about one a month (Equation 4). Each extra window divides the rate by roughly 20 here.

Q4. What does requiring more windows in a row cost?

Model answer

Short episodes. An episode of \(L\) windows is caught with probability \(1 - a_L\) (Equation 5), which falls as \(k\) rises, sharply once \(k\) approaches \(L\), and is zero for \(L < k\). If short episodes matter for the purpose of monitoring, a large \(k\) hides them. Demo m38-d1 shows both curves together.

Q5. A wearable study reports a positive predictive value of 98%. Another reports AF in 34% of notified people. Can both be right?

Model answer

Yes, because they answer different questions. Lubitz et al. (2022) reports the PPV of a detection that occurred while an ECG patch was worn, for AF at the same time. Perez et al. (2019) reports how many notified participants had AF anywhere on a patch applied on average 13 days later; paroxysmal AF may simply not occur during the patch. Always ask what the numerator and denominator are, and when the reference was recorded.

Q6. A clinic follows 500 ILR patients with 0.1 alerts per patient per day, 25% of them true, 3 minutes each. What is the daily load?

Model answer

50 alerts a day, 37.5 of them false, and \(50 \times 3/60 = 2.5\) hours of review (Equation 6). The load depends on the alert rate and review time, not on how many patients are ill.

Q7. A triage filter removes 70% of false alerts and keeps 98% of true ones. Starting from a PPV of 25%, what is the new PPV, and what is lost?

Model answer

\(\mathrm{PPV}' = 0.98 \times 0.25 / (0.98 \times 0.25 + 0.3 \times 0.75) = 0.245/0.47 \approx 0.52\), by Equation 7. Shown alerts fall to 47% of the original. 2% of true alerts are hidden: in the Q6 clinic, \(12.5 \times 0.02 = 0.25\) a day, about 7.5 a month, which no one reviews.

Q8. Why do premature beats cause false AF alerts, and what feature could help?

Model answer

Interval-based detectors see irregular RR intervals, and premature beats produce them; the review by Covino and Russo (2024) found premature beats the main cause of false AF alarms. Features that use atrial activity (P waves on an ECG, or electrogram morphology in devices) can separate them, as can patterns of intervals: a premature beat is usually followed by a longer interval, while AF intervals vary without that pattern.

Q9. The independence assumption behind Equation 4 fails in practice. In which direction is the error?

Model answer

False windows cluster: noise, motion and runs of premature beats last longer than one window. With positive correlation, runs of \(k\) false windows are more likely than \(q^k\) suggests, so the formula underestimates false notifications. Measure the lag-one correlation of window flags on real data, as lab part E does, before trusting the formula.

Q10. How would you evaluate an AI filter for ILR alerts before using it in a clinic?

Model answer

Run it silently on all alerts, with every alert still reviewed by staff, for long enough to see rare rhythms. Report \(\kappa\) (true alerts kept) and \(r\) (false alerts removed) with confidence intervals, by rhythm type. Count review hours saved (Equation 6, Equation 7). Look at every true alert the filter would have hidden. Decide in advance which losses are acceptable. Module m45 covers silent deployment and monitoring.

Q11. A model trained on intracardiac electrograms separates AF from atrial tachycardia well. Why might that matter for implanted devices?

Model answer

Devices tend to group AF, flutter and atrial tachycardia as high-rate events, which can delay or misdirect therapy (Rodrigo et al. 2022). A classifier that separates them could make alerts more specific. In that study, rule-based classifiers using rate or regularity misclassified about 10–12% of cases. It would still need testing on device-recorded signals and in prospective use.

Q12. Where could an AI agent help a remote-monitoring clinic, and where must it not act?

Model answer

It could gather each transmission, run validated detectors as tools, pull the patient’s history and pre-fill the review form, so a person decides faster. Numbers must come from the tools, not from the language model (Feli et al. 2025). It must not dismiss alerts on its own or change device programming; every action that changes something goes through a person. Module m44 treats this design.

Code lab

The lab times pulses against beats, tests the consecutive-window rule, simulates a review inbox and then tries the rule on real Holter recordings. It runs in about a minute.

Part What you do Data
A Build ECG and PPG timing; check Equation 2 and Equation 3 synthetic
B A windowed irregularity detector on a day of rhythm: estimate \(q\) and \(p\) synthetic
C The k-in-a-row rule: simulate days and episodes; check Equation 4 and Equation 5 synthetic
D A clinic inbox with a triage model: review hours and hidden true alerts (Equation 6, Equation 7) synthetic
E The same detector on Holter beat annotations; how correlated are false windows? PhysioNet
F Optional: a small network on RR-interval windows, in PyTorch synthetic
  • Python notebook: labs/python/m38_lab.ipynb. Run it from the course repository after pip install -r requirements.txt, or open it in Google Colab.
  • Data for part E: beat (.qrs) and rhythm (.atr) annotations of five records of the MIT-BIH Atrial Fibrillation Database (PhysioNet 2000; Goldberger et al. 2000) (Open Data Commons Attribution v1.0), a few hundred kilobytes, downloaded from PhysioNet and checked against its published SHA-256 list. Without internet access, part E is skipped.
  • Part F runs only if PyTorch is installed (pip install torch).

Each part ends with checks: lines that pass with the code as given. If you change something and a check fails, that is information, not an error to hide.

Self-check

Short checks you can do in your head. Open each answer after you have tried it.

S1. Mean pulse interval 0.75 s. Heart rate?

Answer

\(60/0.75 = 80\) beats per minute (Equation 1).

S2. Pulse delays vary with SD 20 ms and no correlation between beats. By how much does the variance of pulse intervals exceed that of RR intervals?

Answer

\(2 \times (0.020)^2 = 0.0008\ \mathrm{s}^2\), an extra SD of about 28 ms (Equation 3).

S3. True or false: with \(k = 1\), the expected number of notifications is about \(nq\).

Answer

Almost: \(q[1 + (n-1)(1-q)]\), slightly less than \(nq\), because neighbouring positive windows merge into one run (Equation 4).

S4. An AF episode lasts 3 windows and the rule needs 4. Chance of catching it?

Answer

Zero: \(a_m = 1\) for \(m < k\) (Equation 5).

S5. A filter keeps every true alert and removes no false ones. What happens to PPV?

Answer

Nothing: with \(\kappa = 1\) and \(r = 0\), \(\mathrm{PPV}' = \mathrm{PPV}\) (Equation 7).

S6. Doubling the number of patients followed, with the same alert rate, does what to review hours?

Answer

Doubles them (Equation 6).

Literature figures

These figures are worth looking at alongside this module. We link to them instead of copying them: some licenses do not allow reuse here, and others we could not confirm. Each entry says what to look for and where this module explains it.

  • Notification, confirmation and PPV by subgroup. Lubitz and colleagues, Figure 2 [Lubitz et al. (2022); free full text]. Three panels: the share of participants notified, the share of notified participants with AF or flutter on a later patch, and the detection’s PPV against a patch worn at the same time. Each panel has a different denominator (interview Q5).
  • How long the AF lasted. Lubitz and colleagues, Figure 3 (Lubitz et al. 2022). AF burden on the patch among participants with confirmed arrhythmia, and the duration of the longest episode. Compare with what a “k in a row” rule can catch (Equation 5).
  • Alert transmissions per patient. Neiman and colleagues, Figure 1 [Neiman et al. (2024); free full text]. Mean alert transmissions per patient, among patients with at least one alert, for the two ILR models. Turn them into review hours with Equation 6.
  • The algorithm behind a smartwatch notification. Perez and colleagues describe the algorithm and how pulse intervals (tachograms) were sampled in their supplementary appendix [Perez et al. (2019); free full text]. Look for the windowing and the confirmation rule, and compare them with Equation 4.

Further reading

  • PPG. Allen (2007): a review of photoplethysmography, its signal and its clinical uses.
  • Wearable studies. Perez et al. (2019) and Lubitz et al. (2022): two large prospective studies of notification algorithms.
  • Implanted monitors. Covino and Russo (2024) (false alarms, systematic review), Neiman et al. (2024) (AI-based detection in practice) and Antoun et al. (2025) (AI across implanted devices).
  • Device electrograms. Rodrigo et al. (2022).
  • Data. PhysioNet (2000), with Goldberger et al. (2000) for PhysioNet.
  • Agents. Feli et al. (2025) and Iwai et al. (2025), and module m44.
  • AI competency areas. Machine learning; Bayesian networks (chained predictive values); logic (alert rules); AI and society (alert burden). See the AI competency map.

References

Allen, John. 2007. “Photoplethysmography and Its Application in Clinical Physiological Measurement.” Physiological Measurement 28 (3): R1–39. https://doi.org/10.1088/0967-3334/28/3/R01.
Antoun, Ibrahim, Alkassem Alkhayer, Ahmed Abdelrazik, et al. 2025. “Artificial Intelligence and the Future of Cardiac Implantable Electronic Devices: Diagnostics, Monitoring, and Therapy.” Journal of Clinical Medicine 14 (24). https://doi.org/10.3390/jcm14248824.
Covino, Simona, and Vincenzo Russo. 2024. “False-Positive Alarms in Patients with Implantable Loop Recorder Followed by Remote Monitoring: A Systematic Review.” Pacing and Clinical Electrophysiology 47 (3): 406–16. https://doi.org/10.1111/pace.14941.
Feli, Mohammad, Iman Azimi, Pasi Liljeberg, et al. 2025. “An LLM-Powered Agent for Physiological Data Analysis: A Case Study on PPG-Based Heart Rate Estimation.” Annual International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE Engineering in Medicine and Biology Society. Annual International Conference 2025: 1–7. https://doi.org/10.1109/EMBC58623.2025.11254428.
Goldberger, A. L., L. A. Amaral, L. Glass, et al. 2000. “PhysioBank, PhysioToolkit, and PhysioNet: Components of a New Research Resource for Complex Physiologic Signals.” Circulation 101 (23): e215–20. https://doi.org/10.1161/01.cir.101.23.e215.
Iwai, Rumi, Takunori Shimazaki, Jaakko Hyry, et al. 2025. “Evaluation of Pacemaker Battery Level Estimation Using Large Language Model and Remote Monitoring Historical Data.” Annual International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE Engineering in Medicine and Biology Society. Annual International Conference 2025: 1–6. https://doi.org/10.1109/EMBC58623.2025.11251697.
Lubitz, Steven A, Anthony Z Faranesh, Caitlin Selvaggi, et al. 2022. “Detection of Atrial Fibrillation in a Large Population Using Wearable Devices: The Fitbit Heart Study.” Circulation 146 (19): 1415–24. https://doi.org/10.1161/CIRCULATIONAHA.122.060291.
Neiman, Zachary M, Merritt H Raitt, Gregory Rohrbach, et al. 2024. “Monitoring of Remotely Reprogrammable Implantable Loop Recorders with Algorithms to Reduce False-Positive Alerts.” Journal of the American Heart Association 13 (5): e032890. https://doi.org/10.1161/JAHA.123.032890.
Perez, Marco V, Kenneth W Mahaffey, Haley Hedlin, et al. 2019. “Large-Scale Assessment of a Smartwatch to Identify Atrial Fibrillation.” The New England Journal of Medicine 381 (20): 1909–17. https://doi.org/10.1056/NEJMoa1901183.
PhysioNet. 2000. MIT-BIH Atrial Fibrillation Database (Version 1.0.0). PhysioNet. https://doi.org/10.13026/C2MW2D.
Rodrigo, Miguel, Mahmood I Alhusseini, Albert J Rogers, et al. 2022. “Atrial Fibrillation Signatures on Intracardiac Electrograms Identified by Deep Learning.” Computers in Biology and Medicine 145: 105451. https://doi.org/10.1016/j.compbiomed.2022.105451.