Live raw ECG, on-device math, paper-referenced
RytmoECG streams the 130 Hz single-lead ECG from a Polar H10, runs every detector on your iPhone, and shows the formula, paper, and failure mode behind every value. No cloud. No diagnosis.
Wellness app, not a medical device. Requires a Polar H10. iOS 17 or later.
Try it
Pick a sample rhythm and watch the same algorithms that ship in the app run on it in your browser. No cloud round-trip; the analysis is computed from the strip you see.
The demo runs simplified versions of the same detectors that ship in the app. Real signal from your strap gives more nuance; here we use synthetic but physiologically plausible rhythms.
The pipeline
The Polar H10 streams 130 samples per second of single-lead ECG. RytmoECG takes that buffer end-to-end on your iPhone. Nothing leaves the device unless you tap Share.
Anatomy of one heartbeat
Every cardiac cycle is a five-wave story; atria depolarise, ventricles depolarise, ventricles repolarise. The intervals between them are what separate a healthy rhythm from one worth investigating. RytmoECG measures each one and labels them honestly.
Atrial depolarisation. Roughly 80 to 120 ms wide, under 0.25 mV. Single-lead chest-strap geometry makes the P-wave unreliable; RytmoECG does not separate atrial flutter from AF for that reason.
Ventricular depolarisation. Normal width 60 to 100 ms; wide complex above 120 ms suggests ventricular origin or bundle branch block. RytmoECG flags wide-complex beats as suspect PVCs.
Ventricular repolarisation. We locate T-end with the tangent method (Lepeschkin & Surawicz 1952); the slope of the descending limb is extrapolated to the isoelectric line.
Atrioventricular conduction time. Outside 120 to 200 ms suggests block or pre-excitation. RytmoECG does not target PR programmatically (single-lead noise on P-wave); we surface it inside Advanced Analysis as a measurement, not a flag.
Total ventricular electrical event. Strongly HR-dependent. RytmoECG reports the running median of the last two hundred beats; corrected via Fridericia (QT / RR^1/3), which holds up better than Bazett at the extremes of HR.
The time between successive R-peaks; the rolling RR-series is the input to every HRV metric we publish. PVC and SQI rejection ensure a single artefact does not corrupt the trend.
What RytmoECG catches
Each strip below is the exact shape the detector looks for. We capture fifteen seconds of ECG before the event from a ring buffer, then keep recording for thirty seconds of clean rhythm. If another event fires in that window, the recording extends; you get one clip of the whole burst, with a tally like "bigeminy x 9 · couplet x 2".
What it is. Regular rhythm originating in the SA node. RR intervals tight, narrow QRS, T-waves upright. RytmoECG shows "Steady" when this holds.
How we detect it. RR coefficient of variation under 6 %, no wide complexes for thirty consecutive beats, irregular detector silent.
What it is. A beat that fires early from a ventricular focus. Wide ( > 120 ms ), bizarre morphology, often followed by a compensatory pause.
How we detect it. Beat classified wide-complex by the morphology classifier, prematurity flagged ( RR shorter than 85 % of running median ), then pause classifier checks whether pre-RR + post-RR sums to ~2 x baseline ( PVC ) or ~1.85x ( PAC ).
What it is. Two consecutive wide-complex beats. More electrically irritable than a singleton PVC.
How we detect it. Two wide-complex beats back-to-back, both premature relative to baseline.
What it is. Every other beat is ventricular. Normal-PVC-normal-PVC pattern sustained for many cycles.
How we detect it. Pattern detector sees the N-V-N-V cadence sustained over at least eight beats with stable RR intervals between the singletons.
What it is. Every third beat is a PVC. Common in caffeine, stress, electrolyte shifts.
How we detect it. N-N-V cadence sustained for two full cycles ( six beats minimum ); cooldown twenty seconds.
What it is. Four or more consecutive ventricular beats but lasting under thirty seconds. Triggers a cardiology referral on a Holter; we flag it the same way.
How we detect it. Four or more wide-complex beats in a row with HR > 100 bpm; capture window automatically extends until thirty seconds of clean rhythm follow.
What it is. Atria fibrillate; ventricles respond irregularly. No P-waves, chaotic baseline, RR intervals "irregularly irregular".
How we detect it. CNN-BiLSTM with attention pooling trained on PhysioNet CinC 2017 ( F1 0.81 holdout ); runs every fifteen seconds on the current ECG buffer, ensembled with the rules-based irregular detector.
What it is. A gap between heartbeats exceeding two seconds. Can be benign ( vagal tone ) or significant ( sinus arrest, AV block ).
How we detect it. Any RR interval > 2.0 s after PVC exclusion. Captured as a discrete event so you can see what came before and after.
What it is. RR intervals vary widely without a fixed pattern, but morphology stays narrow. A pre-clinical heads-up before AF or just respiratory sinus arrhythmia at a high amplitude.
How we detect it. RR coefficient of variation > 12 % sustained over twenty beats with narrow QRS throughout.
What it is. An early beat from an atrial focus. Narrow QRS ( unlike a PVC ), often morphologically distinct, no full compensatory pause.
How we detect it. Premature beat with narrow QRS, then compensatory-pause classifier rules out PVC; ratio of pre-RR + post-RR to 2 x baseline below 0.95.
QT interval, in detail
Finding the end of the T-wave by eye is notoriously fiddly; the canonical workaround is to extrapolate the tangent of the descending T-wave limb down to the isoelectric line. RytmoECG does this per beat, rejects beats with SQI below 0.70, and reports the running median of the last two hundred beats.
QT is highly heart-rate dependent. RytmoECG shows the Fridericia correction by default ( QT / RR ^ 1/3 ); the Bazett correction is available in Advanced Analysis. On a noisy single-lead 130 Hz signal we treat QTc as a trend marker only; not a clinical QTc.
Poincare plot
The Poincare plot is the most honest HRV visual; each beat is a point on a scatter. SD1 ( perpendicular spread ) reflects short-term parasympathetic variation. SD2 ( spread along the identity line ) reflects long-term tone.
RytmoECG renders this live as a Canvas on the dashboard; you see ectopic beats fly off the diagonal in real time. AF turns the cluster into a starburst; sinus arrhythmia stretches the ellipse along SD2 without widening SD1.
All algorithms, all references
Each measurement is computed on-device. We show what it means, how RytmoECG computes it, the reference paper, and where it can fail. The same content lives inside the app on the Methodology page.
A CNN-BiLSTM with attention pooling trained on PhysioNet CinC 2017. Runs on-device every fifteen seconds, ensembled with the rules-based irregular detector.
How. Thirty-second window normalised, fed through the model. Threshold from model metadata (`optimal_threshold`); we display a five-state pill: dormant, learning, sinus, watching, AF likely.
Limits. CinC 2017 holdout F1 of 0.81 is good for single-lead; field validation across users is ongoing. RytmoECG never says "you have atrial fibrillation"; it says "rhythm consistent with AF, show a clinician".
Reference. PhysioNet / CinC Challenge 2017; Clifford et al., Computing in Cardiology 2017.
Two cardiac autonomic biomarkers that sit between consumer fitness metrics and clinical risk stratification.
How. Phase-Rectified Signal Averaging (Bauer 2006) computes deceleration and acceleration capacity per their original weighting. HRT (Schmidt 1999) averages the RR pattern around each confirmed PVC for turbulence onset and slope.
Limits. DC / AC need 400+ beats. HRT needs five or more clean PVCs in the session.
Reference. Bauer et al., Lancet 2006; Schmidt et al., Lancet 1999.
Per-beat QT measured via the tangent-of-the-T-wave-descending-limb method, plus a +40 ms QRS-onset offset, plus a SQI gate.
How. T-peak located in 150-400 ms after R; tangent extrapolated to baseline. QT = (T-end - R-peak) + 40 ms. Reported as the running median over the last 200 beats, both Bazett and Fridericia.
Limits. Single-lead 130 Hz makes the T-wave noisy. We treat QTc as a trend marker, not a clinical QTc. Beats below SQI 0.70 excluded.
Reference. Lepeschkin & Surawicz, Circulation 1952; Fridericia, Acta Med Scand 1920.
Every other algorithm that ships in the app. Same depth on the Methodology page inside; this is the index.
How. Rolling mean of 60 / RR over the last sixty beats; min and max gated by a five-beat warmup.
Reference. LITFL ECG rate interpretation.
How. Task Force 1996 standard on PVC-rejected, SQI-filtered RR series. RMSSD = root mean square of successive RR differences; SDNN = std of all NN; pNN50 / pNN200 = share of successive differences > 50 / 200 ms.
Reference. Task Force ESC / NASPE, Circulation 1996; Shaffer & Ginsberg PMC5624990.
How. Per beat, Q-onset is located by back-search from R-peak; S-end by forward-search; both stop when the signal returns within a 50 uV band of the baseline. Width = ( S-end - Q-onset ) / sampleRate. Rolling median over the SQI-filtered series.
Limits. Single lead at 130 Hz; we use baseline-return rather than slope-onset because the latter is noise-sensitive at this rate.
How. Voltage 20 ms after S-end, baseline-subtracted, converted to microvolts. Session mean over the SQI-filtered beats.
Limits. A wellness proxy for ST-segment level; not a diagnostic ST measurement on a single lead.
How. For each beat, classify the largest T-extremum ( 150-400 ms after R-peak ) as positive or inverted relative to baseline. meanTPositive = % of positive T-waves over the session.
How. Real-input FFT via Accelerate vDSP on a power-of-two slice of the 30 s ECG ring buffer; Hann window applied to reduce leakage. The metric is the ratio of in-band power to total power, refreshed every five seconds.
Limits. Non-specific; reflects rapid fluctuations of any origin ( sharp QRS energy, motion, noise ). Useful to compare conditions on the same subject; not clinical.
How. Exponential moving average of mean |sample| in microvolts. A simple "is the strap making good contact today?" marker.
How. Scatter of ( RR_n , RR_n+1 ); SD1 = std perpendicular to identity, SD2 = std parallel. Rendered live.
How. Smoothed instantaneous HR series ( window of five beats ), max minus min. Descriptive, requires no user input.
How. HR reserve = ( avgHR - restHR ) / ( maxHR - restHR ), clamped to [0, 1]. HREI = duration_minutes x HR_reserve x sex_factor ( 1.92 male, 1.67 female ). Settings live under Personal in Settings; no entry required for the metric to still surface a generic value.
Reference. Banister, E.W. Modeling Elite Athletic Performance, 1991.
How. Drop from peak HR to lowest HR in the sixty seconds that follow the peak. Standard fitness signal for vagal reactivation.
Reference. Cole et al., NEJM 1999.
How. Default-to-trust; score starts at 1.0, only deducts for clear-cut problems ( implausible amplitude, baseline drift, off-physiology RR ). Rolling mean of last thirty beats shown as the Signal pill.
How. Three-second std-dev of accelerometer magnitude. Thresholds: standstill < 0.04 g, light < 0.15 g, moving ≥ 0.15 g.
How. RR-series resampled to 4 Hz with linear interpolation, Hann-windowed, real-input FFT via Accelerate vDSP. Power integrated per Task Force 1996 bands: VLF 0.0033-0.04 Hz, LF 0.04-0.15 Hz, HF 0.15-0.40 Hz. LF/HF ratio published as the autonomic-balance metric.
Reference. Task Force ESC / NASPE, Circulation 1996.
How. Free byproduct of the HRV spectrum: the dominant frequency inside the HF band corresponds to respiratory sinus arrhythmia. Multiply by sixty for breaths per minute. Typical 10-20 /min at rest.
How. Detrended Fluctuation Analysis on the integrated mean-removed RR series. Linear fit slope of log(fluctuation) vs log(box size) for short scales (4-16 beats, alpha1) and long scales (16-64 beats, alpha2). The endurance community uses alpha1 = 0.75 as a heuristic for the first lactate threshold.
Reference. Peng et al., Chaos 1995; Rogero, Gronwald et al. 2020.
How. Non-linear complexity of the RR series. SampEn = -ln(matches at m+1 / matches at m), m=2, r=0.2 x SD. ApEn computed similarly. Lower values mean more regular rhythm; high entropy is associated with healthy variability.
Reference. Richman & Moorman, AJP-Heart 2000; Pincus, PNAS 1991.
How. Linear regression of smoothed instantaneous HR over the session. Reported as drift rate plus R^2 so you know how steady the workload actually was. Smaller drift at sustained effort = better aerobic fitness.
How. Both computed from a 30-day baseline of your own sessions, never anyone else's. Stress = mean of three normalised inputs (RMSSD drop, resting HR rise, LF/HF rise). Readiness = RMSSD lift (60 %) + resting HR low (40 %). Every input is shown next to the score; no black box.
How. Pulls sleep windows from HealthKit, buckets each RytmoECG session as "night" or "day", and reports the median RMSSD and HR per bucket. Healthy autonomic recovery shows higher RMSSD overnight than during day; we make the gap visible.
How. For each HealthKit workout, locate the nearest RytmoECG session within 8 h before and 12 h after. Report the delta RMSSD so you can see whether the bout left your vagal tone elevated, depressed, or unchanged.
How. GitHub-style 53 x 7 cell grid; each cell is one day, colour-coded by your chosen metric (RMSSD, mean HR, or session count). A year of recording visible without scrolling. Tap to switch metric.
How. Five-zone breakdown by percent of personal maxHR (set in Settings). Each beat contributes its RR interval to the matching zone; per-session bars show time and percent for each. Standard endurance-coaching layout.
How. Every session can be exported as plain CSV: time_seconds, rr_seconds, hr_bpm, sqi per beat, plus a header comment block with sample rate, device, and timestamp. Drop straight into Python / R / MATLAB.
How. Re-runs every detector against a single thirty-to-sixty-second strip in isolation; lists all parameters with flag colours. Verdict: likely arrhythmia, ambiguous, likely normal, or poor signal.
Three personas, one app
RytmoECG computes everything for everyone, but a researcher, a patient and an athlete want very different things foregrounded. Pick a display mode in Settings; the dashboard re-arranges around what matters to you. Nothing is hidden from view, only re-prioritised.
Maximum density. The default.
Simplified Now. RytmoECG answers, you breathe.
Training emphasis. Zones up front.
The mode does not change what is computed. Every detector runs the same way regardless of skin; we are not hiding scary numbers from the patient mode, just not putting them where you would have to look at them.
How RytmoECG avoids false alarms
Most wellness apps that flag arrhythmias get noisy fast; poor signal, movement artefacts and respiratory sinus variation all fake textbook patterns. RytmoECG stacks three guards on top of every detector so you do not get spammed with "low signal · maybe PVC" alerts.
If on-strap accelerometer std-dev is above 0.15 g (active movement) for the last three seconds, all auto-event capture is suppressed outright. Light motion (0.04 to 0.15 g) reduces confidence by 25 %.
Below 0.50 mean SQI over the last thirty beats, no auto-event is emitted. The floor sits at 0.50, not higher, because a PVC's own odd shape drags the rolling SQI down; too strict a gate would block the very beats it should catch. Under 0.50 the RR series is unreliable enough that a detector would invent variability that is not there.
The PVC compensatory-pause detector needs a pause that sums with the early RR to roughly 2 x baseline, plus either a QRS shape unlike the dominant sinus template or a beat at least 20 % early. Requiring shape alone missed PVCs the classifier had merged into sinus; the timing criterion recovers them. Respiratory sinus arrhythmia (8-10 % RR swing at the breathing rate) still falls below the bar and is filtered out.
When the R-peak detector has not seen a beat in two seconds, the signal-peak estimate gently decays toward the noise estimate so the threshold relaxes. Plus a full threshold reset at the end of the ten-second warmup so pairing artefacts cannot lock the detector high.
Where RytmoECG fails
What RytmoECG will never do
Get it
iPhone, iOS 17 or later. Requires a Polar H10 chest strap.
App Store; coming soonRytmoECG is a wellness app; not a medical device. It does not diagnose, treat, or screen for any condition. If you have chest pain, shortness of breath, or persistent palpitations, contact emergency services or your physician.