Atmospheric & environmental triggers of atrial fibrillation

• Earth's static field at the surface is 25–65 µT. A severe geomagnetic storm (Dst ≈ −589 nT, March 1989) perturbs it by about 1%; a typical "storm day" by 0.2–0.5%. Walking past steel reinforcement changes your local field by more.
• Induced tissue electric field, Faraday, torso loop r = 0.1 m, aggressive substorm dB/dt = 500 nT/min: E ≈ 4 × 10⁻¹⁰ V/m. The ICNIRP basic restriction for CNS tissue is 0.1 V/m — roughly 10⁹ times larger. Myocardial capture needs ~1–10 V/m. Membrane thermal noise is many orders above the induced signal.
• Magnetohydrodynamic potentials scale linearly with B. The 1–10 mV MHD artefact seen at 7 T scales to about 0.7 nanovolts at 500 nT. That is not a mechanism.
• Radical-pair/cryptochrome transduction requires either far higher intensity or radical-pair lifetimes >1 µs, "far longer than lifetimes typically observed in condensed matter" (Krylov 2026, Biol Rev, DOI 10.1111/brv.70108). ELF fields with periods >3.33 ms are quasi-static over a radical pair's lifetime, so a slow storm signal offers nothing to rectify.

1. The exposure is planetary. Kp/Ap/Dst is identical for everyone on Earth that day. There is no exchangeable unexposed group; even nominally individual-level case-crossover designs (Feigin 2014, PMID 24757102) carry a purely ecological exposure inside them.
2. Adverse seasonal confounding. Geomagnetic activity peaks at the equinoxes (Russell–McPherron) and cycles at 27 days and 11 years. Cardiovascular events peak in winter and follow temperature, infection, pollution and daylight. The exposure signal lives in the same frequency band as the confounders being smoothed out.
3. Multiplicity of 10³–10⁴. ~6–10 indices × ~5–10 lags × 4–8 outcomes × 4–8 strata, with no correction anywhere in the field.
4. Uncorrected autocorrelation. Mattoni 2020 (Eur J Appl Physiol, PMID 32306151) repeated the HRV analysis properly: "the loss of most significant effects after this correction suggests that previous findings may be a result of autocorrelation… we question the validity of the previous studies."
5. Sign incoherence. The same index is claimed to increase stroke, decrease ICD shocks, increase MI, decrease HRV and increase parasympathetic tone, depending on the group analysing it.
6. Author concentration. Four clusters (Stoupel; Halberg–Cornélissen; Vencloviene/Kaunas; HeartMath/Global Coherence) produce most positive reports, sharing data infrastructure and analytic conventions. Counting their papers as independent replication is a category error.
7. Publication bias. A 2025 scoping review found 28 of 36 studies (78%) positive — an implausible hit rate for a field with no mechanism and mostly ecological designs. The 2025 meta-analysis (PMID 40256184) included only 6 of 644 screened studies and reports ranges rather than a pooled estimate with I².
8. Effect sizes incompatible with the exposure. A claimed ~3-fold MI increase in women from a 0.2% field perturbation should reduce credence, not raise it.

1. Pre-register one primary triple — one exposure, one lag, one outcome. Everything else FDR-controlled against a published grid.
2. Use local measured fields, not planetary indices: nearest INTERMAGNET observatory dB/dt and local K.
3. Exploit the magnetic-latitude gradient. A real effect must be larger at high magnetic latitude, where perturbations are 5–10× greater. The northern-Sweden null (PMID 12135204) is a direct failure of that prediction and any positive claim must account for it.
4. Time-stratified case-crossover with same-month, same-weekday referents, plus explicit adjustment for temperature, PM2.5 and influenza activity.
5. Negative controls, pre-specified. Exposure: Kp shifted six months, or future exposure. Outcome: hip-fracture admissions or another endpoint sharing the care-seeking pathway with no plausible geomagnetic route.
6. Autocorrelation-aware inference — pre-whitening, AR error structure, or block-bootstrap. Naive correlation of two time series is inadmissible.
7. Power for a defensible effect. At OR ≤ 1.05 with ~2% exposed days, you need on the order of 10⁵–10⁶ events. Every study except Ebrille is underpowered by 2–3 orders of magnitude for any effect that is not biophysically incredible.
8. A sham-controlled human exposure experiment at realistic amplitudes (≤500 nT with a realistic dB/dt spectrum on a normal 50 µT background) does not currently exist. Every "simulated storm" study uses supraphysiological amplitudes, unstated N, or no sham arm.

If you were hoping to do this in a public ICU dataset: date-shifting destroys weather and air-quality linkage in every one of them.

• MIMIC-III / MIMIC-II — dates shifted per patient, but documentation explicitly states "the day of the week and season of the year were preserved." So season, day-of-week and hour-of-day analyses are legitimate; daily weather is not.
• MIMIC-IV — a single shift per subject_id, into 2100–2200; distinct patients are "not temporally comparable." Seasonality preservation is not asserted in the documentation (an open question in the MIT-LCP issue tracker). Treat season as unrecoverable until confirmed. anchor_year_group gives only a 3-year window.
• HiRID — the standout: documentation explicitly guarantees "we made sure to preserve the seasonality, time of day and the day of the week." Best free external validation cohort for season/hour/day-of-week hypotheses.
• eICU-CRD — hospital and unit identifiers removed; region only. AmsterdamUMCdb — millisecond offsets from admission, no dates.

Consequence: a daily weather or air-quality analysis is executable only on your own institutional registry with intact admission dates and a hospital geocode, or on a national registry (STS ACSD, EACTS, ANZSCTS, SWEDEHEART, NICOR) under a data agreement that retains dates. The proof of concept for ICU-to-monitor linkage already exists: Groves 2020 (PMID 32355989) linked 46,965 emergency ICU episodes across 87 ANZ ICUs to EPA monitors — PM2.5 per 10 µg/m³ → 30-day mortality RR 1.18 — but had no arrhythmia endpoint.

• Sound. Bedside L_Aeq,1min, L_Amax, number of night-time events >45 dB(A), alarm counts by type and source, conversation vs equipment fraction. Use event counts, not L_den-style averages — the randomised human data show the vascular effect scales with the number of events, not with equivalent level.
• Light. Bedside illuminance and melanopic EDI at eye level, full 24-h profile, day/night ratio, lights-on event count, window proximity, blind position.
• Circadian and sleep. Actigraphy or EEG-derived sleep fragmentation, night-time care-interaction density, urinary 6-sulfatoxymelatonin and cortisol rhythm, RASS trajectory, CAM-ICU delirium as a shared-cause co-outcome.
• Thermal. Bedside room temperature and RH; patient core-temperature trajectory, rewarming rate after hypothermic CPB, time to normothermia, fever burden, hypothermia overshoot.
• Indoor air. Indoor PM2.5 and PM1 (optical with gravimetric calibration), CO₂ as an air-exchange proxy, NO₂, VOCs; compute and report the infiltration factor F_inf for your unit empirically.
• Respiratory / oxygen. FiO₂ and SpO₂ trajectory, hyperoxia hours, desaturation burden (a T90 analogue), ventilation mode and PEEP as an intrathoracic-pressure and atrial-stretch term.
• Built environment. Single room vs open bay, bed position relative to door and nursing station, distance to the busiest alarm source.
• Excursions. Transport episodes to CT/cath lab/OR — flag or model separately; these are the only times the patient leaves the box.

• Air quality: PM2.5, PM10, PM_coarse, ultrafine particle number, black carbon, NO₂, NOx, O₃, SO₂, CO; PM speciation (sulfate, nitrate, ammonium, organic matter) where available; AQI category; wildfire-smoke days (satellite smoke-specific PM2.5, not total); dust-event flags. Windows: pre-operative 7/30/365 d and postoperative 0–5 d, modelled as a lag surface rather than a chosen lag.
• Meteorology: mean/min/max temperature, apparent temperature, wet-bulb globe temperature, heat-wave and cold-spell indicators, diurnal temperature range, day-to-day change, absolute and relative humidity, barometric pressure and its 24-h change, wind speed, frontal passage, precipitation, sunshine hours, photoperiod at latitude.
• Residential exposures (pre-admission): modelled L_den and L_night at the home address, artificial light at night, greenspace/NDVI, area deprivation, distance to major road.
• Biological aerosols: pollen counts by taxon, mould spore counts.
• Infection activity — mandatory, not optional: sentinel ILI rate, influenza/RSV/COVID laboratory positivity or wastewater surveillance. This is the dominant time-varying confounder for every winter-peaking exposure and almost no AF study adjusts for it.
• Site-level: hospital elevation, urban heat island index, climate zone (multicentre only).

• Hour of day of surgery start and of cross-clamp release; hour of day of AF onset; postoperative hour and day.
• Day of week; weekend vs weekday; public holiday; proximity to a DST transition; academic-calendar turnover.
• Month, season, calendar year (secular trend), and time since the start of the surgical programme.
• Lunar phase and illumination fraction — include it, but pre-specify it as a negative control rather than a hypothesis. If it comes up "significant" in your model, that is information about your multiplicity, not about the moon.

• Local ground-magnetometer dB/dt and local K from the nearest INTERMAGNET observatory — not planetary Kp.
• Dst minimum, Ap, storm-day classification, AE index.
• Solar wind speed, IMF B_z, F10.7, sunspot number; solar flare class and timing; solar proton events.
• Neutron-monitor count rate and Forbush-decrease flags (anticorrelated with solar activity).
• Pre-specified negative-control exposure (Kp shifted six months, or future-lagged exposure) and negative-control outcome (an endpoint sharing the detection pathway with no plausible geomagnetic route). A non-null negative control quantifies your residual confounding directly — which is the only genuinely useful thing this arm can produce.
• Magnetic-latitude gradient across sites, if multicentre.

Postoperative hour (the dominant driver — risk peaks at 48–72 h), beta-blocker and amiodarone administration timing and dose, beta-blocker withdrawal, magnesium and potassium, CRP/IL-6/white cell count, fluid balance, inotrope and vasopressor dose, haemoglobin, pain score, sedation depth, mechanical ventilation status, temporary epicardial pacing mode and rate (it both masks AF detection and is itself adjusted in response to rhythm), CPB and cross-clamp time, left atrial diameter, prior paroxysmal AF, and — critically — telemetry gap minutes and artefact fraction per hour as the ascertainment-intensity term.

• Device / telemetry interrogation studies are the reference design. Machine-adjudicated outcome with second-level timestamps, no care-seeking filter, genuine hour-level lag resolution, natural within-person design, exact person-time. Their own threats: highly selected populations on antiarrhythmics, device classifications that are not diagnoses (undersensing at high rates, far-field oversensing, manufacturer-specific mode-switch thresholds), and small N inflating winner's curse.
• Time-stratified case-crossover removes all time-invariant confounding by construction but is silent on anything varying over days to weeks — temperature, co-pollutants, pollen, influenza, weekend alcohol.
• Time-series with DLNM is the right tool for lag surfaces, but the degrees-of-freedom-for-time choice directly determines the estimate, and knot placement × max lag × centring multiplies the researcher's choice space. Demand a sensitivity sweep across 3–12 df/year.
• Ecological/correlational designs — the geomagnetic literature's mainstay — are disqualifying here, not merely weak.

Self-matched design, exposure standardised within stratum, 3 control periods per case, α=0.05, 80% power. Required events, not patients:

Three corrections make that worse. (1) Infiltration. Indoor PM2.5 in air-conditioned space runs F_inf ≈ 0.47 ± 0.18, and 0.09–0.27 in HVAC commercial buildings; a true OR of 1.20 per SD of indoor exposure becomes ≈1.06 per SD of outdoor exposure, moving you from 315 events to ~3,000. Measuring indoor exposure is worth more than a tenfold increase in sample size. (2) Within-stratum variance is small in a climate-controlled room — pilot it for three months and use the observed σ, not 1. (3) Adjudication attrition and telemetry gaps cost 10–25% of usable events.

Search limitations

• The NCBI E-utilities API and PubMed/PMC pages were not reachable from this environment (403 on eutils; reCAPTCHA on PubMed/PMC; Europe PMC, Crossref, OpenAlex and Semantic Scholar likewise blocked or heavily rate-limited). Retrieval used web search, direct publisher sites (OUP, Springer/BMC, Nature, MDPI, Frontiers, LWW, J-Stage), DOAJ, institutional repositories, and a Wiley-weighted academic corpus.
• Consequence: this is a comprehensive structured review, not a PRISMA-reproducible systematic search. Study counts per exposure are indicative rather than de-duplicated screening outputs.
• PMIDs. No PMID here is invented. Where a PMID appeared verbatim in a resolved PubMed URL or fetched text it is given plainly; a number of citations were verifiable only by DOI. Verify every PMID and every effect estimate against the primary record before it goes into a manuscript, grant or protocol. Elsevier, AHA and JACC full texts were largely inaccessible, so several effect estimates came from abstracts or secondary sources.
• The web-search budget for this session was exhausted before an independent verification pass could be run.

Where the two independent appraisals disagreed

• Short-term PM2.5 → AF. The substantive review graded it MODERATE (consistent direction, meta-analytic support, plausible HRV mechanism). The methodological review graded it LOW (pooled AF estimate rests on 4 studies, the best-measured device subgroup is null for ventricular arrhythmia, E-value 1.15–1.46, no synthesis has ever applied GRADE, and the umbrella review states arrhythmia is the thinnest stratum in environmental cardiology). Rows where this applies are flagged CONTESTED.
• Cold → AF. Graded MODERATE on the strength of two very large Chinese case-crossover/DLNM studies; a methodological reading puts it at VERY LOW because there is no AF-specific meta-analysis, the parent syntheses run I² ≈ 93–99%, and the best-adjusted temperate-climate study is null.
• Noise → AF. MODERATE from the pooled Nordic cohort; LOW from the umbrella-review perspective (WHO never graded AF; 64% of the underlying meta-analyses were rated LOW or VERY LOW by the umbrella's own GRADE).

Where they disagreed, the table shows the higher grade with the contested flag, so you can see both readings rather than a laundered consensus.

Specific numbers not to use

• Zhang 2026, Heart (UK Biobank): HR 1.37 per 1 µg/m³ PM2.5 for AF, implying HR ≈ 25 per 10 µg/m³, with a CI width of 0.02 on 28,977 events. Not biologically credible; almost certainly a units or specification error. The downstream claim that 26–52% of AF cases are PM2.5-preventable should be treated as artefactual. Two other analyses of the same cohort report HR 1.044 per IQR and HR 1.26 per 10 µg/m³.
• Shuhaiber 2013 full-moon mortality OR 0.21 (0.05–0.81) — n=210, ≥16 uncorrected comparisons, no mechanism, no replication.
• Ho 2022's "8% OHCA reduction per 1 µg/m³" — internally inconsistent with the paper's own RR of 1.022 per 10 µg/m³.
• Any geomagnetic effect size implying a large biological response to a 0.2–1% field perturbation.

Priority full texts to pull before writing anything

1. PMID 35138681 — de Bont umbrella review; the load-bearing citation that AF/arrhythmia is the thinnest stratum in environmental cardiology.
2. DOI 10.1161/JAHA.123.030907 — POAF ascertainment meta-analysis (assessment method and definition). Essential for your methods section.
3. DOI 10.1161/CIRCRESAHA.123.323477 — "Environmental Exposome and Atrial Fibrillation," Circ Res 2024. Best single framing citation.
4. DOI 10.1016/j.jcrc.2021.09.006 — circadian variation in new-onset AF in ICU patients; as far as I can tell the only paper directly on circadian timing of ICU NOAF onset.
5. PMID 25659238 (Ebrille) and PMID 32306151 (Mattoni) — the two results that should anchor any space-weather section.
6. PMID 29470452 (WHO/van Kempen) — the only body that applied GRADE outcome-by-outcome to environmental noise.
7. PMID 28693043 — E-value methodology.
8. NCT03503812 (MEDEA) — randomised dust-avoidance intervention with device-measured AF burden as a primary outcome. Results were not locatable; check the registry directly. This is the closest existing trial to the design you would want.