AFib screening on consumer wearables (Apple/Fitbit heart studies)

"We shipped a heart monitor to half a million wrists in a weekend. The hard part was not the algorithm; it was mailing back the ECG patches."

A Logistically Humbled AI Agent

From pulse timing to a rhythm decision

A wrist PPG has no P wave and no crisp QRS complex, so the morphology tricks of ECG are unavailable. What it does have is the arrival time of each blood-volume pulse, which lets you build a pulse-to-pulse interval series, the optical analogue of the R-to-R tachogram. In normal sinus rhythm those intervals are regular and drift slowly. In AFib the ventricles fire at irregular intervals, so the pulse series scatters. Detecting AFib on a watch is therefore, at its core, the same irregularity measurement introduced for the clinical case in Section 29.2: quantify how disordered the interval sequence is, and flag windows that cross a threshold. This is the anomaly-and-change framing of Chapter 12 aimed at a physiological target through a much noisier lens.

Why does the noise change the design so completely? Because motion corrupts a PPG far more than it corrupts a chest ECG. A single irregular-looking window is almost always an artifact, not an arrhythmia. The engineering answer, used by both major studies, is opportunistic sampling plus temporal persistence: only score short "tachograms" collected when the wearer is still (Apple sampled during periods of stillness; Fitbit ran its check largely during inactivity and sleep), and only raise a notification when several consecutive tachograms are independently classified as irregular. Apple's rule required 5 of 6 consecutive tachograms to read irregular before notifying. That persistence gate is not a detail; it is the primary tool for converting a jittery per-window classifier into a specific per-person alert.

The listing below builds an opportunistic PPG screener in miniature. It scores each still-window tachogram for irregularity, then applies a "5 of the last 6" persistence rule before it will notify, mirroring the Apple Heart Study logic on synthetic data.

import numpy as np

def tachogram_irregular(ppi_ms, thresh=0.12):
    """Flag one PPG tachogram as irregular from pulse-to-pulse intervals."""
    ppi = np.asarray(ppi_ms, float)
    rmssd = np.sqrt(np.mean(np.diff(ppi) ** 2))   # beat-to-beat scatter
    return (rmssd / ppi.mean()) > thresh           # normalize out heart rate

def screener(windows, k=5, n=6):
    """Notify only when k of the last n still-window tachograms are irregular."""
    flags, notified = [], False
    for w in windows:
        flags.append(tachogram_irregular(w))
        recent = flags[-n:]
        if len(recent) == n and sum(recent) >= k:
            notified = True
    return notified, flags

rng = np.random.default_rng(0)
sinus = [1000 + rng.normal(0, 8, 30) for _ in range(6)]          # regular
afib  = [1000 + rng.normal(0, 130, 30) for _ in range(6)]        # irregular
print("sinus ->", screener(sinus)[0])   # False: no notification
print("afib  ->", screener(afib)[0])    # True: persistence rule fires

As Listing 30.2 shows, the per-window test is almost trivial; the intelligence lives in the sampling policy and the persistence gate wrapped around it. Real products add a signal-quality classifier upstream (Section 30.6) so that only trustworthy tachograms ever reach the scorer, and they calibrate the threshold on the deployment population rather than on tidy clinical data.

import heartpy as hp
wd, m = hp.process(ppg, sample_rate=64.0)   # peaks + cleaning
intervals = wd["RR_list_cor"]               # corrected pulse-to-pulse intervals (ms)

The two studies that made the case

The Apple Heart Study (Perez and colleagues, New England Journal of Medicine, 2019) enrolled about 419,000 participants entirely through an iPhone app, with no in-person visit. The watch algorithm sampled tachograms during stillness and issued an irregular-pulse notification under the 5-of-6 rule. Roughly half a percent of participants (and a larger fraction of those over 65) were notified. Anyone notified was mailed an ambulatory ECG patch to wear for about a week as the reference standard. Among participants who returned a patch with simultaneous data, the wrist notification's positive predictive value against concurrent ECG was about 0.84, yet only around a third of notified participants showed AFib on the patch over its wear window, because the intermittent arrhythmia often did not recur during the days-later recording. The Fitbit Heart Study (Lubitz and colleagues, 2022) followed the same siteless template at even larger scale, about 455,000 participants, with a photoplethysmography algorithm that ran mainly during inactivity and sleep; its per-detection positive predictive value against a concurrent one-week ECG patch was reported near 0.98, reflecting a conservative operating point.

Why these are screening, not diagnostic, devices

The regulatory framing follows directly from the funnel. Apple's irregular-rhythm notification feature and its on-demand single-lead ECG app were cleared by the FDA in 2018 through the De Novo pathway as over-the-counter screening tools, explicitly not intended to replace clinical diagnosis, and Fitbit's photoplethysmography algorithm received comparable clearance. A screening claim is much weaker than a diagnostic one: the device must reliably raise a flag worth confirming, and it must not alarm so often that users mute it, but it is never the final word. This is why the studies report the notification-to-confirmation funnel rather than a lone accuracy number, and why every claim is anchored to a concurrent ECG reference. The broader machinery of prospective clinical validation, regulatory pathways, and post-market surveillance that these studies exemplify is the subject of Chapter 34.

The scale that makes wearable screening attractive also concentrates its risks. When a single algorithm runs on tens of millions of wrists, a small bias becomes a population-level harm. Two recurring threads matter here. First, the alert is only as trustworthy as its calibration on the deployment distribution, so a probability that means the same thing across ages, activities, and devices requires the calibration discipline of Chapter 18. Second, green-light PPG is absorbed differently by different skin tones, so signal quality, and therefore both sensitivity and the false-alarm rate, can vary systematically across users; that fairness-and-calibration problem is important enough to earn its own treatment in Section 30.7. A screener that works beautifully on the enrolled cohort but degrades on darker skin or during exercise is not finished; it is dangerous at scale.

What's Next

In Section 30.3, we broaden the same wrist sensor from rhythm to the everyday vital signs it is more commonly asked to report: blood-oxygen saturation, resting and continuous heart rate, and heart-rate variability, examining what each measurement really estimates from reflected light and where the optics quietly break down.