Building a Recovery Score with Apple Health Data

Oura gives you a readiness score. Whoop gives you a recovery score. These are composite metrics that combine several physiological signals into a single number. You can build the equivalent from your own HealthKit data, using health4ai tools and a scoring formula defined in Claude.

This isn’t trying to replicate Oura’s proprietary algorithm. It’s building a recovery score that reflects your specific data, your personal baselines, and your definition of what recovery means.

The Inputs

A useful recovery score needs at least three data streams:

HRV (SDNN) — the primary autonomic recovery signal
Resting HR — secondary cardiovascular indicator, moves more slowly than HRV
Sleep quality — the mechanism of recovery (total hours, stage distribution)

Optional additions:

Training load (prior 48h) — contextualizes why recovery might be low
Training load (prior 7-30d) — cumulative fatigue indicator

Getting the Inputs

get_hrv_trend(days=14)          # HRV with trend context
query_metric(                   # Resting HR
    metric_type="HKQuantityTypeIdentifierRestingHeartRate", 
    days=14
)
get_sleep(days=3)               # Last 3 nights of sleep
get_workouts(days=7)            # Recent training load
get_metric_stats(               # Personal baselines
    metric_type="HKQuantityTypeIdentifierHeartRateVariabilitySDNN",
    days=90
)

Or use get_coaching_brief() which combines most of these in a single call.

A Simple Scoring Formula

Define a scoring approach in your Claude prompt:

Using today's health data, compute a recovery score from 0-100 using this formula:

HRV component (40 points):
- 40 points if today's HRV is above my 90-day p75
- 30 points if between p50 and p75
- 20 points if between p25 and p50
- 10 points if below p25

Resting HR component (30 points):
- 30 points if today's resting HR is at or below my 90-day p25 (low = good)
- 20 points if between p25 and p50
- 10 points if between p50 and p75
- 5 points if above p75 (elevated)

Sleep component (30 points):
- 30 points if sleep was 7.5+ hours
- 20 points if 6.5-7.5 hours
- 10 points if 6-6.5 hours
- 5 points if under 6 hours

Call get_coaching_brief() and get_metric_stats() for my HRV baseline. 
Calculate my score and explain what's driving it.

Claude runs the tools, applies the formula, and returns something like:

Recovery Score: 76/100

HRV: 58ms — above your p75 threshold of 55ms → 40/40 Resting HR: 57 bpm — between your p50 (55) and p75 (62) → 20/30 Sleep: 6.8 hours — borderline range → 20/30

Recovery is good overall. HRV is strong, which is the most reliable signal. The 76 score is held back by modest sleep last night. If you have intensity training planned, it’s appropriate — but get 7.5+ hours tonight.

Why Build Your Own vs Use a Wearable’s Score

Wearable recovery scores (Oura, Whoop, Garmin Body Battery) are proprietary algorithms trained on population data. They may not reflect your specific physiology or training patterns. A few reasons to build your own:

Customizable thresholds. The formula above uses your personal p25/p75 as thresholds, not population averages. A reading of 52ms HRV means different things for someone whose baseline is 45ms vs someone whose baseline is 65ms. Population-based scoring misses this.

Transparent inputs. You know exactly which metrics went into the score and how they were weighted. When the score is low, you can see why — it’s not a black box.

Custom weighting. If sleep consistently matters more for your recovery than HRV, weight sleep at 50% instead of 30%. If you find resting HR is noisy and HRV is more predictive for you, reduce the resting HR weight.

Extensible. You can add inputs over time — nutrition data if you track it, stress indicators, temperature if your wearable captures it.

Building It Into a Daily Script

#!/bin/bash
# ~/scripts/recovery-score.sh
DATE=$(date +%Y-%m-%d)

PROMPT="Compute my recovery score for $DATE using the formula below.
Call: get_coaching_brief(), and get_metric_stats(metric_type='HKQuantityTypeIdentifierHeartRateVariabilitySDNN', days=90)

Scoring:
HRV (40pts): above p75=40, p50-p75=30, p25-p50=20, below p25=10
Resting HR (30pts): below p25=30, p25-p50=20, p50-p75=10, above p75=5
Sleep (30pts): 7.5h+=30, 6.5-7.5h=20, 6-6.5h=10, under 6h=5

Return: Score X/100, bullet points for each component, 1-sentence guidance."

claude --print "$PROMPT"

Run it at 7:00 AM via LaunchAgent and you have a personalized recovery score every morning without a subscription.

Caveats

A few things to keep in mind:

The formula is arbitrary. The weights (40/30/30) and thresholds (p25/p75) reflect a reasonable starting point, not a validated clinical score. Oura’s and Whoop’s algorithms are trained on large datasets and validated against subjective recovery ratings. Yours is a heuristic.

Single-day HRV is noisy. HRV varies day-to-day for reasons unrelated to recovery (coffee, time of measurement, position). The trend (from get_hrv_trend) is more reliable than a single morning reading. Consider using the 3-day average HRV rather than today’s single reading.

Resting HR lags. Resting HR responds to stress and training more slowly than HRV — it might stay elevated for 2-3 days after a hard session while HRV rebounds. The two metrics together tell a more complete story than either alone.

Despite these caveats, a simple composite score that you understand and that uses your personal baselines is often more actionable than a sophisticated score you can’t inspect.

health4ai: Free while in early access.
Download on the App Store →