Device crosswalk: wearable fields to LOINC

​

What is in it

from mirobody.translate import devices

devices.lookup("apple", "HKQuantityTypeIdentifier", "stepCount")
# VendorField(vendor='apple', type='HKQuantityTypeIdentifier', field='stepCount',
#             metric='steps', loinc='55423-8', confidence='confident', note='...')

devices.by_code("112429-6").fields
# {'apple': 'heartRateVariabilitySDNN'}        one vendor; the others publish RMSSD

[u.metric for u in devices.unmappable() if u.reason == "VENDOR_SCORE"]

python scripts/device_crosswalk_report.py vendors | reasons | base

​

By vendor

​

What no code fits

​

Read this before normalising across vendors

1. HRV: one vendor publishes SDNN, five publish RMSSD, four publish nothing. Apple’s heartRateVariabilitySDNN is SDNN. Health Connect’s HeartRateVariabilityRmssdRecord, Huawei’s heartRateVariabilityRMSSD, Fitbit’s hrv.value.dailyRmssd, WHOOP’s recovery.score.hrv_rmssd_milli and Oura’s sleep.average_hrv are RMSSD. Samsung, vivo, Xiaomi and Zepp publish no HRV type. Garmin publishes beat-to-beat RR intervals, from which SDNN can be computed. LOINC 2.83 codes only the SDNN family (112429-6, 76643-6, 80404-7); RMSSD has zero hits in the axis table. The two are different definitions and cannot share a row: the catalogue keeps hrvSDNN (coded), hrvRMSSD (uncoded) and hrvDatas for a series whose statistic is not stated. Syncing an Oura ring into Apple Health writes RMSSD into the field Apple defines as SDNN; a downstream reader that trusts the field name is then wrong.
2. Skin temperature depends on the site. 61008-9 Body surface temperature (Temp/Qn/Surface) is the quantitative code for a wrist device, marked unverified because the wrist has no code of its own and the code’s IEEE origin is a monitor’s body temperature. 60830-7 Finger temperature is the confident code for a ring (Oura). 60833-1 is the toe. 39106-0 Temperature of Skin is a nominal finding and cannot hold a number; 60839-8 is the infant-incubator context. Health Connect’s SkinTemperatureRecord carries measurementLocation, the one source that lets a decoder pick the site automatically.
3. Deviation is not temperature. Fitbit’s tempSkin.value.nightlyRelative and Oura’s temperature_deviation are offsets from a personal baseline; WHOOP’s skin_temp_celsius and Health Connect’s baseline are absolute. An offset goes to temperatureDelta, which has no LOINC code and must never be filed under 61008-9 or 8310-5.
4. Weighted minutes are not minutes. Fitbit active-zone minutes count cardio and peak minutes twice, by the documentation’s own words; Garmin intensity minutes weight vigorous minutes the same way. Undo the weighting before any LOINC duration code, or do not code them.
5. Energy units. WHOOP’s cycle.score.kilojoule is kilojoules; every other vendor is kilocalories. Divide by 4.184.
6. Heart-rate zones align nowhere. Fitbit has four, WHOOP five, the catalogue five, LOINC three plus a maximum zone. Not coded; the only comparable quantity across vendors is total moderate-plus-vigorous time.
7. Activity intensity bands half align. Oura’s low, medium and high map onto LOINC’s light, moderate and vigorous durations (101688-0, 101689-8, 101690-6) and is the one confident case. Health Connect’s ActivityIntensityRecord has two bands and no light. Huawei’s exercise_intensity.v2 is not an intensity band at all; its field is exercise_type, the kind of exercise.
8. Four vocabularies for the sleep stages. Deep sleep is Apple asleepDeep, Health Connect DEEP, Huawei 深睡, Fitbit deep, WHOOP slow_wave_sleep and Oura deep_sleep_duration; light sleep is Huawei 浅睡, Health Connect LIGHT, Fitbit and Oura light, and Apple asleepCore. All of them land on 93831-6 and 93830-8, but Apple’s Core and WHOOP’s slow-wave sleep are marked unverified: synonymous, not the same name.
9. The sleep mode is a routing key. Fitbit has two log modes, classic without stages and stages with them. Huawei’s sleep_type has three values: 1 with every field, 2 with no REM, 3 with only a total. Oura’s sleep.type includes late_nap. Route on the mode before coding stages, or a stage the mode does not provide is stored as zero. This is the most dangerous line in this list.
10. A field name is not its meaning. vivo’s DATA_TYPES.WALKING_SPEED is in steps per minute: a cadence, not a speed. Read the unit column of the vendor’s documentation, never the name.
11. Statistical capability is a constraint. vivo’s watch API supports minimum and maximum for heart rate, blood oxygen and stress, and sums for standing, intensity, steps, distance and calories, and a mean for nothing. A mean code such as 103205-1 has no vivo source; do not build the mapping because it “should” exist.
12. Glucose depends on the unit. Honor’s BLOOD_GLUCOSE.measureValue is in mmol/L and codes to 15074-8, not to the catalogue’s 2339-0 (mg/dL). The catalogue alias names the analyte and the printed unit picks the variant; mirobody.translate.code does this for every alias.
13. Total is not active. Oura’s total_calories and Health Connect’s TotalCaloriesBurnedRecord include basal metabolism; the catalogue’s activeCalories does not. Only the active figure is coded confidently.

​

LOINC’s own axis defects

​

How the catalogue uses it

• A metric’s identity is ("loinc", code) only when the code is confident (Metric.canonical); otherwise it is ("mirobody-device", name), so an unverified code never becomes a series key.
• The device catalogue’s answer for a metric is handed to mirobody.translate.code as a catalogue alias. The alias names the analyte; when the printed unit fits a sibling of that code in the same specimen, the sibling is used (glucose in mmol/L), and when it fits none, the alias stands.
• Apple’s SDNN identifier is decoded to hrvSDNN; the Apple gait, perfusion and six-minute-walk identifiers, previously quarantined, now have rows.
• Five invariants in test_engine_coverage.py hold the tables and the catalogue together: every code is in the axis table with a confidence, no two metrics share a code unless registered as one quantity at two grains, every crosswalk row names a real code and a real catalogue row, the Apple table says what the Apple decoder does, and the unit gate on an alias.

​

The base table

​

Sources

​

Changing it