הדוח המלא: אמינות, נכונות, תיקוף חיצוני וביקורת תקינה — עם
המספרים ורווחי הסמך. כתוב לקוראים מקצועיים. השרשרת שמייצרת כל מספר כאן נמצאת
ב-research/PROVENANCE.md במאגר.
| white | ICC 0.79 · 68 פנים |
| black | ICC 0.84 · 13 פנים |
| east_asian | ICC 0.68 · 9 פנים |
| west_asian | ICC 0.87 · 9 פנים |
| 1 | Identify whether interrater/intrarater reliability or agreement was investigated models/calibration.json records retest as "neutral vs smiling front shot, face moved between them" — test-retest of one automated rater, not interrater. There are no human raters to disagree. | עומד |
| 2 | Name and describe the measurement device explicitly analyzer/metrics.py: eleven named metrics, each computed in its own facial zone in CIELAB, on a lighting-normalized image. Version and scale recorded per scan. | עומד |
| 3 | Specify the subject population of interest 102 identities, mean age 27.7 (SD 7.1, range 18-54), 49 female / 53 male, 69 white / 13 black / 10 west Asian / 9 east Asian — recorded in models/calibration.json under calibration_set.demographics. | עומד |
| 4 | Specify the rater population (if applicable) It WAS not applicable — the rater is a deterministic program, so there is no rater population and no interrater variance. That stopped being true when a grading study ran: twelve faces graded blind on published photonumeric scales. The rater population is therefore ONE non-expert vision model, which is not a dermatologist and not a population. Every Spearman interval from it crossed zero, so it establishes nothing about agreement and is reported as a pilot that sized the real study rather than as a result. | חלקית |
| 5 | Describe what is already known about reliability and provide a rationale analyzer/evidence.py — 25 sources, including the smartphone-colorimeter benchmark (ICC 0.85-0.95) our own numbers are compared against, and Flament 2023, which found the same two weak spots we measured. | עומד |
| 6 | Explain how the sample size was chosen The calibration sample was fixed at 102 before this could be planned, so the honest answer is a PRECISION analysis of it rather than an a priori calculation, and both are reported. Delivered: ICC intervals with a median width of 0.156, from 0.027 (unevenness) to 0.333 (pores). Required: interval width scales as 1/sqrt(n), so the four-face file this replaced carried a width near 0.79 — the entire ICC scale, which is why its thresholds understated the noise by up to eightfold. For the expert-agreement study, which is not yet run and CAN be planned, models/study_design.json carries a simulated power calculation: n=30 for 80% power at rho=0.6, n=40 for an interval narrower than 0.50. | עומד |
| 7 | Describe the sampling method Convenience sample: the complete Face Research Lab London Set (CC-BY-4.0, DOI 10.6084/m9.figshare.5047666), every identity with both expressions used, none excluded. Convenience sampling is a real limit on generalisation and is recorded as one. | עומד |
| 8 | Describe the measurement process (time interval, blinding) Two front shots per identity, same session, different expression. Blinding is not applicable to a deterministic program. The interval is short, which makes the retest HARDER rather than easier: the difference contains a real facial change, so the noise estimate is conservative. | עומד |
| 9 | State whether measurements were conducted independently analyzer/metrics.py holds no state between calls: each image is processed independently and the analyzer cannot see the other measurement, which is the strongest form of independence available. | עומד |
| 10 | Describe the statistical analysis sigma_within = SD of paired differences / sqrt(2); sigma_between = sqrt(max(var(first) - var_within, 0)); ICC = var_between / (var_between + var_within); MDC95 = 1.96*sqrt(2)*sigma_within. Intervals by bootstrap over FACES, 2000 resamples, seed 7 — recorded in calibration.method so the analysis is reproducible from the file. | עומד |
| 11 | State the actual number of raters, subjects and replicate observations Per metric: n_faces and n_rows, in every entry of calibration.metrics. A metric measured on fewer faces than another cannot hide behind the headline count. | עומד |
| 12 | Describe the sample characteristics models/calibration.json, calibration_set.demographics: age mean/SD/range, gender split, ethnicity counts. Added because this item asked for it and the file did not carry it. | עומד |
| 13 | Report estimates of reliability and agreement including statistical uncertainty Every metric carries icc, icc_ci95 (bootstrap over faces), noise_std, mdc95 and a quality band judged on the LOWER bound of the interval per Koo & Li 2016 — never on the point estimate. | עומד |
| 14 | Discuss the practical relevance of the results MDC95 IS the practical relevance: it is the threshold the product must clear before it may call a change real, and it moved the bar on the overall score from 6.15 points to 8.52. models/validation.json states separately that clearing it means detectable, NOT important — de Vet & Terwee, and the distinction the product is forbidden from blurring. | עומד |
| 15 | Provide detailed results if possible models/calibration.json and models/validation.json are served in full at /v1/evidence and summarised at mirra.skin/sources.html — the numbers, the intervals, the sample, and the limitations. | עומד |
Colour is read in CIELAB following CIE recommendations, and erythema is separated from melanin by reflectance rather than taken from the red channel. EEMCO explicitly endorses image analysis by colour camera as an accurate route. It also warns that erythema is barely visible in deeply melanized skin — the limitation our own fairness work found independently, which is the strongest sign this is the right standard to be held to.
Every colour metric is computed in CIELAB, and skin tone is reported as ITA° = arctan((L*-50)/b*)·180/π, the accepted colorimetric definition.
The standard asks for control of individual, temporal and ENVIRONMENTAL variables. We record head pose, expression, capture quality and the weather at scan time, and the capture coach enforces framing, distance, tilt and expression before the shutter. What we cannot do is control the room: a clinic fixes the illuminant, and a phone in a bathroom does not. Lighting normalisation reduces this and does not remove it.
MIRRA is not a medical device and holds no CE mark or FDA clearance. It measures and tracks; it does not detect, diagnose or treat, and the moment a skin measurement claims to do any of those it is regulated as a device. Stated here so its absence reads as a decision rather than an omission.
| פצעונים | Automated facial acne lesion detecting and counting algorithm for acne severity evaluation |
| עיגולים כהים | Cutaneous colorimetry · O'Mahony MM, Sladen C, Crone M, et al · X-Rite ColorChecker Classic — CIE L*a*b* reference values · Axelsson J, Sundelin T, Ingre M, et al |
| יובש | Koseki K, Kawasaki H, Atsugi T, et al · Garg A, Chren MM, Sands LP, et al · Altemus M, Rao B, Dhabhar FS, et al |
| נקבוביות | Flament F, et al · Dissanayake B, Miyamoto K, Purwar A, et al |
| אדמומיות | Fink B, et al · Cutaneous colorimetry · Dawson JB, et al · Erythema image analysis in rosacea · VISIA redness comparison · X-Rite ColorChecker Classic — CIE L*a*b* reference values · Piérard GE · Fullerton A, et al |
| קמט השפה-אף | Narins RS, Carruthers J, Flynn TC, et al |
| ברק שומני | Kohli I, Kastner S, Thomas M, et al |
| כתמים | Fink B, Matts PJ · Cutaneous colorimetry · X-Rite ColorChecker Classic — CIE L*a*b* reference values · Retinol 0 |
| חוסר אחידות | Fink B, Matts PJ · Fink B, et al · Fink B, et al · Cutaneous colorimetry · Koseki K, Kawasaki H, Atsugi T, et al · X-Rite ColorChecker Classic — CIE L*a*b* reference values · Piérard GE · Retinol 0 |
| קווים דקים | Jiang LI, Stephens TJ, Goodman R · Retinol 0 |