Important Apple Health Signal Mining

Abstract

Apple Health is strongest for behavior-linked physiology: smoking relapse, alcohol/recovery, sleep, resting heart rate, HRV, and 30-day experiment response. It cannot diagnose plaque, SIBO, or bleeding, but it can show whether exposures are visibly moving recovery/autonomic markers.

apple-health · wearable · resting-heart-rate · HRV · sleep · smoking · alcohol · ferritin · lpa

Apple Health Signal Mining

What the imported data can and cannot answer

The import is already useful, but uneven.

Coverage in the most relevant windows
Window Sleep Resting HR Walking HR HRV Respiratory rate SpO2
2025-07-01 -> 2025-10-31 13.0% 17.1% 14.6% 18.7% 13.0% 17.1%
2025-11-01 -> 2026-04-15 64.5% 83.7% 60.2% 84.3% 67.5% 83.1%

Implication: the archive is strong enough for smoking/alcohol/recovery analysis from Nov 2025 onward, but weaker for reconstructing the immediate physiologic response to the Aug 2025 bleed.

Strongest personal signal already visible: smoking/autonomic stress

The cleanest new pattern is the shift from the late-2025 smoke-free window to the spring-2026 relapse window.

Late-2025 smoke-free vs spring-2026 smoking-relapse window
Metric 2025-11-15 -> 2025-12-31 2026-03-01 -> 2026-04-15 Change
Steps/day 8505 5088 -3416
Sleep minutes/night 324 371 +47
Resting HR 54.0 bpm 60.8 bpm +6.8 bpm
Walking HR average 79.3 bpm 89.5 bpm +10.2 bpm
HRV (SDNN) 38.7 ms 30.0 ms -8.7 ms
Respiratory rate 15.2 16.0 +0.8
SpO2 96.0% 95.6% -0.4

This is exactly the kind of within-person pattern that matters more than population averages:
- higher resting HR
- higher walking HR
- lower HRV
- lower activity

That pattern fits the smoking literature well. Smoking heaviness has a causal link to higher resting HR (PMID: 26538566), and smoking cessation improves HRV (PMID: 23397454).

Practical interpretation: Apple Watch is already giving a personal physiologic penalty score for smoking, not just a future-risk lecture.

Alcohol/recovery signature: likely present, but it needs deliberate tagging

The imported data support the alcohol-recovery framework more than they prove it retrospectively.

The clearest candidate cluster is around the Feb 2026 re-exposure period:
- 2026-02-13: respiratory rate 29, SpO2 93.5%, HRV 24.2 ms, resting HR 66 bpm
- 2026-02-15: resting HR 66 bpm, HRV 31.7 ms, sleep 300 min after the big-party window noted in the cloud doc
- 2026-02-27 -> 2026-03-01: resting HR 75 -> 69 bpm, HRV 12.6 -> 17.6 ms, walking HR 94 -> 99 bpm, low activity
- 2026-03-21 -> 2026-03-27: repeated low-HRV / high-RHR / poor-sleep days with one SpO2 nadir at 91%

These are real physiologic disturbances, but the archive alone cannot say whether each was driven by alcohol, smoking, infection, travel disruption, poor sleep, or gut symptoms. The fix is simple: tag exposure days.

Best use going forward

Create explicit day tags for:
- alcohol: 0 / 1-2 / 3-5 / 6+ small beers
- smoking: none / light / heavy day
- gut status: stable / bloated / pain / loose stool / blood
- illness: yes/no

Then compare next-morning resting HR, HRV, respiratory rate, sleep, and same-day walking HR against a 28-day rolling baseline.

The 2026 alcohol paper is directly relevant here: acute alcohol exposure raised nocturnal resting HR, lowered HRV, shortened sleep, and reduced next-day activity in a dose-dependent fashion (PLOS Digital Health 2026, doi 10.1371/journal.pdig.0001284).

Iron question: Apple data do not yet prove an iron-depletion performance pattern

The original high-yield hypothesis was: falling ferritin should eventually show up as worse exercise efficiency or poorer recovery.

The current import does not yet prove that, mainly because the best physiologic signals are sparse during late Aug to Oct 2025. The archive does show:
- very high pre-bleed activity: ~11.7k steps/day from 2025-07-01 -> 2025-08-18
- major restriction after the bleed: ~3.9k steps/day from 2025-08-20 -> 2025-10-31
- recovery by Jan-Feb 2026: ~10.5k steps/day
- VO2 max estimate improved from a single Aug 2025 point of 25.4 to ~32.7 in Nov-Dec 2025 and ~33.5 in Mar-Apr 2026

That does not rule out iron-related exercise inefficiency. It only means the current summary layer is too crude. VO2 max from Apple Watch is directionally interesting but not accurate enough to carry the analysis alone; validation work found underestimation overall, overestimation in lower-fitness users, underestimation in higher-fitness users, and poor overall reliability (PMID: 39083800).

What would make the iron signal visible

The next-level analysis is not more blood tests first. It is a standardized workout-efficiency metric.

Use one repeatable benchmark:
- same route
- same duration or distance
- same time of day
- preferably similar fed/fasted state
- avoid smoking/alcohol confounding the prior evening

Track:
- average heart rate for the route
- pace or distance for the route
- next-morning resting HR
- next-morning HRV
- ferritin / CBC windows

Why this matters: non-anemic iron deficiency can impair exercise tolerance before dramatic anemia develops, and the effect is easier to see in performance and recovery than in one-off resting physiology alone (PMC10608302; JACC Case Reports 2026 nonanemic iron-deficiency case report).

Gut-inflammation use case: promising as a flare discriminator, not disease-specific proof

The imported data create a realistic way to separate inflammatory bad days from functional bad days.

Wearable studies in IBD found that HRV, HR, RHR, steps, and oxygenation shift during flare periods and can change up to 7 weeks before flares (PMID: 39826619). Diverticular disease is not IBD, so this is an extrapolation, not a validated disease-specific rule.

Still, it gives a useful decision frame:
- symptoms + rising resting HR + falling HRV + lower steps = more likely systemic inflammatory burden
- symptoms without physiologic disruption = more likely meal intolerance / post-inflammatory sensitivity / local gut irritation

Candidate future workflow:
1. Repeat calprotectin when symptoms are active.
2. Pull the prior 14-day wearable window.
3. Compare against rolling baseline.
4. Ask: was there a whole-body physiologic shift, or mainly symptoms?

That is likely to be more informative than CRP alone for this profile.

Cardiovascular use case: helpful for rhythm and recovery, not for plaque burden

For very high Lp(a), Apple Watch is useful in a narrow way:
- tracking resting HR / HRV trends
- catching symptom-time ECGs
- prompting BP monitoring discipline
- flagging unexplained reductions in exercise tolerance

It is not a structural risk monitor.

What the watch can meaningfully do
What it cannot do

Consumer wearables are strongest for AF screening, not broader arrhythmia rule-out; positive findings still need clinical confirmation with medical-grade evaluation (Cleveland Clinic Journal of Medicine 2024 review). For this profile, the one missing home metric with the highest ROI is still blood pressure, not more passive watch-derived novelty. AHA/ACC guidance and AHA patient guidance support home BP monitoring with an upper-arm automatic cuff.

Which Apple metrics deserve trust here

High-value
Medium-value
Low-value / overread risk

Sleep-device validation work supports this hierarchy: total sleep duration is decent, but stage estimates are much less reliable, and Apple Watch specifically tended to overestimate light sleep and underestimate deep sleep compared with polysomnography (PMID: 39460013).

Data already present but not fully exploited

The import includes:
- 94 GPX workout routes
- 16 ECG sidecar files
- 140 workouts total

A quick route check shows that some 2026 sessions are already usable as benchmarks:
- 2026-02-14 running route: ~3.36 km in 24.6 min (~7.32 min/km)
- 2026-03-10 running route: ~1.91 km in 14.2 min (~7.44 min/km)

The current summary artifact does not yet join route pace with heart-rate efficiency. That is the single best upgrade if the goal is to detect iron-related or smoking-related performance drift early.

Key Takeaways for This Profile

  1. The strongest personal signal already visible is a smoking/autonomic signature: higher resting HR, higher walking HR, lower HRV, and lower activity in the relapse window.
  2. The alcohol-recovery hypothesis is plausible in the data, but it needs explicit day tagging to separate alcohol from smoking, illness, and bad sleep.
  3. The iron question is still open because the most interesting Aug-Oct 2025 period has poor wearable coverage; the answer will likely come from standardized route-efficiency tracking, not raw VO2 max.
  4. Apple Health is better for flare discrimination and recovery tracking than for direct disease diagnosis.
  5. For Lp(a), the most important missing home metric remains upper-arm blood pressure, not another blood panel or repeated Lp(a).
  6. ECG recordings matter only when tied to symptoms. Silent background ECG collection is low-yield.
  7. The current archive already contains enough route and ECG data to justify a second-pass analytic upgrade if needed.

Research Trace