Accessibility settings

This paper is in the following e-collection/theme issue:

Theme Issue: AI and ChatGPT in Dermatology (53) Research Instruments, Questionnaires, and Tools (1247) Tools, Programs and Algorithms (611) Imaging Informatics (517) Digital Photography and Imaging in Dermatology (66) Digital Biomarkers and Digital Phenotyping (479) Artificial Intelligence (5069)

Published on in Vol 9 (2026)

Preprints (earlier versions) of this paper are available at https://preprints.jmir.org/preprint/93236, first published .
Dermatologist examines woman's face for anti-aging treatment

AI-Based Grading of Facial Aging Signs and Prediction of Visible Age From Standardized Facial Images: Algorithm Development and Internal Evaluation Study

AI-Based Grading of Facial Aging Signs and Prediction of Visible Age From Standardized Facial Images: Algorithm Development and Internal Evaluation Study

Authors of this article:

Frederic Flament1 Author Orcid Image ;   Panagiotis-Alexandros Bokaris1 Author Orcid Image ;   Tao Li1 Author Orcid Image ;   Aurelie Abric2 ;   Sileye Ba1 ;   Julien Despois1 Author Orcid Image ;   Hussein Jouni1 Author Orcid Image ;   Laudine Bertrand1 ;   Paul Michel1 ;   Thomas Dauxerre1 Author Orcid Image ;   Diala Haykal3 ;   Hughes Cartier4
Article Authors Cited by Tweetations Metrics

    1L'Oréal Research and Innovation, 9 rue Pierre Dreyfus, Clichy, France

    2Eurosyn, Villebon-sur-Yvette, France

    3Centre Médical Laser Palaiseau, private practice, Palaiseau, France

    4Centre Médical Saint Jean, Arras, France

    *all authors contributed equally

    Corresponding Author:

    Frederic Flament, PhD


    Background: Facial appearance influences visible age (VA), defined here as the age perceived by others; this perception-based outcome differs from chronological age and biological age. Human evaluation of facial aging signs can be time-consuming and subject to interobserver variability. A standardized AI-based approach might offer a more reproducible assessment of visible facial aging in digital dermatology research.

    Objective: This study aimed to develop and evaluate an AI-based process for automatically grading facial aging signs and predicting the VA of participants from standardized full-face images.

    Methods: Evaluation of full-face photographs of 468 North American women aged 18 to 80 years recruited in Dallas and Phoenix, Texas. Fifteen trained experts, including dermatologists, graded 13 aging signs using Skin Aging Atlas scales. A separate, naïve group of 253 North Americans aged 20 to 60 years, recruited in New York City, estimated the VA of each participant. For each sign, facial areas were pulled out using facial landmarks and fed into ResNet-based regression models, which were trained to match the averaged expert scores. The automated facial sign grading models were tested on 90% of the image set, and the remaining 10% was used for evaluation. Results were assessed using the mean absolute error, the proportions of estimates within 0.5, 1.0, and 1.5 grading units of the expert reference, and Spearman correlations between chronological age, human-assessed VA, and predicted visible age (PVA).

    Results: The mean absolute error was below 0.3 grading units for 11 of the 13 facial aging signs. The average proportion of automatic gradings within 1 grading unit of the expert reference was about 98% for participants of East Asian ancestry group, 96% for participants of African ancestry group, and 97% for participants of European ancestry group. Human-assessed VA correlated strongly with real age (r=0.94, 95% CI 0.928‐0.950; P<.001), although participants were perceived as around 10 years older than their real age, on average. The AI-PVA also correlated strongly with the human-assessed VA (r=0.93, 95% CI 0.917‐0.941; P<.001), with approximately 70% of estimates within 5 years.

    Conclusions: The AI-based system showed low grading errors across most facial aging signs and also a fairly strong relationship between the PVA and human-assessed VA under standardized image acquisition conditions. Consequently, the system could provide a reproducible digital outcome for facial aging research and future longitudinal monitoring. However, external validation across separate populations, in men, and across image acquisition devices is required before clinical and consumer-facing implementation. As the predicted visible age is derived from quantified clinical facial signs, it reflects measurable features of visible facial aging, while its relationship with biological age and overall health remains to be established through future studies integrating biological biomarkers.

    JMIR Dermatol 2026;9:e93236

    doi:10.2196/93236

    Keywords

    AI; visible age; biological markers; longevity; image analysis; clinical grading; facial signs 


    The lifespan of humans, otherwise called life expectancy, has shown a rather constant and significant progression in recent decades. According to the World Health Organization, global life expectancy increased from 66.8 years in 2000 to 73.1 years in 2019, and healthy life expectancy increased from 58.1 years to 63.5 years throughout the same time frame [1]. These figures vary, however, widely among adults with regard to genetic inheritance, exposure to life-threatening diseases, and geography. This large diversity introduces the notion of health span, that is, the years lived in good health, free from chronic disease and functional decline, including mental health [2-4].

    This has reshaped the frontiers of the science of longevity, that is, the use of integrative strategies involving cosmetics, devices, digital services, coaching, oral supplements, and so on, far beyond a mere topical antiaging cosmetic solution. Facial skin, as a marker of the aging process, whether chronological and/or photoinduced, displays different aspects over decades and is an important element of health span, including well-being [5-8]. Previous studies have shown that these changes vary substantially across ethnic ancestries and genders [9-12]. Especially as a mirror of inner and invisible changes, facial skin could pave the way to exploring new holistic skin “clocks.” Yet, in the absence of biological measurements, visible facial aging should not be deemed equivalent to biological aging.

    In this framework, different notions of age need to be differentiated. Real age (RA) refers to chronological age, whereas visible age (VA) refers to more or less the age perceived by others from facial appearance. Predicted visible age (PVA), as used in the current study, refers to an algorithm-derived estimate of the same VA. Meanwhile, biological age represents physiological or molecular aging and typically depends on biological markers. Although these dimensions seem related, they are not quite interchangeable. Previous work has already shown that perceived age can be linked with age-related phenotypes, functional status, molecular markers, and survival, supporting its interest as a perception-based aging outcome [13,14]. Recently, deep learning methods have likewise suggested the possibility of predicting perceived facial age from facial images [15].

    Within this evolving paradigm, cosmetic dermatology is thus emerging as more than a discipline of aesthetic refinement, and it has become a meaningful participant in promoting health span and contributing to broader integrative longevity strategies. VA (eg, age as perceived by others) may, therefore, represent a digital health and perception-based outcome. The extent to which VA mirrors biological age and could be compared with RA (chronological age, as presented on an ID card) is one facet, among others, of understanding individual aging and relationships with patients and consumers for all actors [16-27]. The present study does not investigate VA as a measure of biological age. In short, fitting the VA of a participant with some biological markers of the skin is an ambitious project that could ideally be performed in real-life conditions, such as points of sale (POS) for cosmetic products or clinics with doctors and practitioners. From a technical viewpoint, recent years have shown fruitful help from smartphones as image collectors, thanks to their efficient cameras and high resolution (≥5 megapixels). Smartphones equipped with adequate algorithms are increasingly used and can allow the prompt analysis of images from very large cohorts of participants in real life or at home [28-31]. Such image-based approaches are of special interest in digital dermatology, where more standardized and reproducible evaluation of visible facial features could complement conventional clinical or expert grading.

    As for the biological aspect, the recent introduction of systems for fast assessments of biomarkers will hopefully yield valuable data related to the root causes of skin aging, viewed as longevity “clocks.” Proteomic patterns, for example, can be extracted from adhesive patch sampling of the skin surface [32] and link invisible and visible aging manifestations. As these kinds of approaches are currently used to establish biological age, fitting their future results with a PVA remains instrumental in developing the most coherent ecosystem to accompany patients and consumers in engagement, compliance, and long-term monitoring. Hence, such integration represents a future perspective, as no biological markers were evaluated in the current study.

    Although there is growing interest in AI-based facial assessment, approaches to combining standardized facial image capture, expert grading, and independent human assessment of VA are still needed. The question remains: do we really get more reliable and less variable results when using automatic grading? To increase diversity within the development set, the study included North American women spanning various ages and ethnic backgrounds, and covered a broad range of facial skin tones and visible aging indicators. Basically, the work aimed to design and test an AI-driven pipeline for automatically scoring facial aging signs and estimating VA using standardized full-face photographs. It also intended to compare this automatic grading with expert assessments and examine the relationships among RA, human-assessed VA, and AI-PVA across the ancestry groups represented in the study [33-35].


    Study Design

    The study followed a 5-fold structure: the acquisition of standardized full-face images from a cohort of North American women; the grading of visible facial signs by a panel of trained experts; the building of an AI-based automatic grading system; the evaluation of VA by an independent, naïve panel; and, finally, the calculation of the PVA.

    Ethical Considerations

    The study was approved by L'Agence nationale de sécurité du médicament et des produits de santé (ANSM, France) under reference number 2025-A02830-49. Oral and written consent, in accordance with the Declaration of Helsinki, was obtained from all participants related to the study objectives and procedures. The consent also included that images would be taken under anonymized conditions, blind-coded, and deleted after study completion. Participants also consented to the use of their images for publication.

    Panel of Full-Face Photography

    A total of 468 North American women (aged 18–80 y) were selected through a local agency and photographed in local facilities (Dallas, Texas, and Phoenix, Arizona, USA) using the Nexa device (Canfield Beauty) for standardization (position, lighting, etc). Figure 1 illustrates the device and 3 examples of the photographs provided, showing ethnic backgrounds and skin tones. Two inclusion criteria were used: (1) age between 18 and 80 years and (2) absence of any facial disorder (scars, acne, melasma, seborrheic dermatitis, telangiectasia, etc). The panel consisted of participants of different ages and ethnic ancestries, as presented in Table 1.

    ‎
    Figure 1. The Nexa device and 3 examples of full-face images from different skin tones.
    Table 1. Distribution of the 468 North American women with clinical standardized pictures dataset according to ethnic ancestries and age groups.
    Ethnic ancestry18-30 y, n31-40 y, n41-50 y, n51-60 y, n61-70 y, n71-80 y, nTotal, n
    East Asian16161426208100
    African ancestry28252315286125
    European525648294216243
    Total969785709030468

    Panel of Trained Experts for Grading Facial Signs

    A panel of 15 trained experts (including dermatologists), previously used on different occasions for comparable tasks, was asked to grade the severity of 13 facial signs (Table 2) according to reference scoring ranges obtainable from ethnic-specific skin atlases [36,37] under standardized conditions. Each expert grader was seated about 1 m away from a 24-inch high-resolution display with a resolution of 1920 × 1200 pixels. The screen was calibrated daily using a colorimeter. To ensure robustness of this process, several images were presented twice during the evaluation process without being identified as duplicates. The high number of experts allowed for the exclusion of graders in cases of erratic scores.

    Table 2. Thirteen clinical facial signs and their respective photographic scale’s ranges.
    Facial signsDefinitionScale range
    Forehead wrinklesDepth of the transverse wrinkles on the forehead0‐8
    Glabellar wrinklesDepth of vertical wrinkles between eyebrows0‐6
    Crow’s feet wrinklesDepth of wrinkles at the area of the outer eye corner0‐6
    Periorbital wrinkles (upper cheek area)Depth of folds at malar area below Crow’s feet, eye orbit excepted0‐9
    Nasolabial foldDepth of the fold present between the base of the nose and lips0‐7
    Marionette linesDepth of folds at the corner of lips0‐8
    Pigmentary spots (density)Number of spots per area unit on the cheek0‐7
    Ptosis of lower part of the faceSagging severity of the lower parts on each side of the chin0‐6
    Diffuse rednessDiffuse redness and microvessels visible, especially on cheeks0‐4
    Cheek skin poresSize of visible pores on the cheek irrespective of their densities0‐5
    Interocular wrinklesDepth of horizontal folds between inner eye corners0‐7
    Perioral lip wrinklesDepth of wrinkles on superior labial area0‐6
    Eye bagsIntensity of the ocular bulge0‐6

    For each facial sign and photograph, the mean of the scores given by the experts was used as the reference score. Despite using integer grades on the photographic scales, the averaged expert scores provided continuous reference values and enabled the automatic system to account for subgrade variations.

    Naïve Panel Attributing VA to Photographs

    A total of 253 North American women of different ethnic backgrounds were recruited in New York City, USA, as a naïve panel by a local agency dedicated to part-time occupations. These participants were aged 20 to 60 years, with 48% (n=121) aged between 20 and 39 years and 52% (n=132) aged between 40 and 60 years, with good visual acuity as a unique inclusion criterion. They were informed about the objective of their task and signed an informed consent form, ensuring full anonymity. The whole panel, when successively viewing the 468 blind-coded photographs presented at random, was asked to answer a simple question: “How old is this woman?” Their answers then referred to VA and were linked to their RA.

    AI-Based Automatic Grading System

    Based on the expert’s grading, an AI-based automatic grading system was developed to automatically compute the severity of the 13 facial signs. To assess facial clinical signs with high accuracy, the AI-based automatic grading system is framed as a regression task, where a dedicated model is trained to estimate a continuous severity score for each targeted clinical sign from a cropped facial image.

    First, an off-the-shelf landmark detection algorithm is run to compute key points used as reference landmarks on the face, such as the tip of the nose, inner corners of the eyes, etc. Based on these landmarks, a cropping step uses handcrafted heuristics to extract the relevant zones of the face as defined by the clinical grading atlases.

    The chosen backbone was ResNet, a streamlined convolutional neural network using residual connections to improve prediction accuracy [38,39]. The architecture consists of 5 convolutional residual blocks, progressively reducing the spatial dimensions to a 1 × 1 × 512 feature map, followed by 3 fully connected layers with rectified linear unit (ReLU) activations that reduce the output dimensionality to a single scalar (the predicted clinical severity score). Each of the 13 facial signs was handled by a dedicated model instance, resulting in a holistic multistep diagnostic pipeline that takes a full-face image as input and yields a list of sign-specific predicted scores as output.

    Functionally, each model takes as input a crop of the image that corresponds to the target clinical sign (eg, forehead wrinkles, periorbital wrinkles, and marionette lines) and predicts a scalar value within skin aging atlas grading scales (Table 2) [36,37]. The average grades provided by 15 experts were used as reference labels. Although the scales are defined with integer values (grade 0, grade 1, etc), the averaged expert annotations allowed the models to account for subgrade variations. The pooled labeled data from 15 expert annotators, trained to evaluate with high precision, allow us to account for subgrade accuracy by targeting the average prediction.

    Model Training and Internal Evaluation

    Each model was trained using the following configuration:

    • Loss function: difference between the predicted score and the average of expert annotations
    • Optimizer: Adam [40] with β1=0, β2=0.99
    • Learning rate: 0.0005
    • Batch size: 32
    • Training duration: 300 epochs

    To prevent overfitting and maximize generalizability, two complementary strategies were used. First, early stopping was monitored using an internal validation subset, carved out from the training data (10% of the 90% training subset), with training halted if validation loss did not improve over a defined patience window. Second, data augmentation was systematically applied during training, including the addition of chromatic noise, shifts in hue and lighting intensity, and small random rotations of input crops. These augmentations simulated real-world variability in capture conditions (eg, minor lighting inconsistencies and slight head movements) and reduced the risk of overfitting to the controlled photographic conditions of the training set, thereby improving robustness when deployed in live settings.

    An automatic grading system was trained and evaluated on a standardized dataset that had been scored by experts and dermatologists. The dataset was split into training and test subsets, with 90% of the data used for training the model and 10% used to evaluate the capacity of the system to generalize to images not encountered during training [22,33]. Within the 90% training subset, a further internal validation split (10% of the training data) was used exclusively for early stopping and model selection purposes, while the remaining training data were used for model fitting. The split was stratified by ethnic ancestry and age group to ensure balanced representation across subgroups. Participant-level separation was strictly enforced to prevent any form of data leakage.

    To ensure the resulting system is as robust and accurate as possible, the training procedure includes a data augmentation step, where input crops are randomly degraded or altered to teach the system to ignore such perturbations. Despite the use of controlled lighting and poses, data augmentation minimizes the risk that unforeseen circumstances degrade the performance of the system. The data augmentation includes the addition of chromatic noise, shifts in hue and lighting, as well as small rotations of the crops.

    The crops are fed into the corresponding models to obtain a predicted grade of clinical severity, thereby linking with PVA. This results in a holistic algorithm that functions as a multistep diagnosis tool, taking a full-face image as input and yielding a list of predicted scores as the output.

    PVA Derivation

    Based on the validated automatic grading outputs, PVA was computed from the 10 facial signs found to be significantly correlated with VA as assessed by the naïve panel (Table 3; cheek skin pores (CSP), pigmentary spots (PS) density, and diffuse redness (DR) were excluded due to low correlation with VA). For each retained sign i, a second-order polynomial regression of the form Y = Aixi2 + Bixi + Ci was fitted using ordinary least squares on the training subset exclusively, where xi was the automatically predicted facial sign grade, Y was the associated VA estimate, as assessed by the naïve panel, and Ai, Bi, and Ci were the regression coefficients. The PVA of each photographed participant was obtained by averaging the polynomial outputs across all 10 retained signs. Coefficients Ai, Bi, and Ci for each sign, as derived from the training set, are reported in Table 4. The generalization performance of the PVA model was then assessed on the held-out test subset by computing the Spearman correlation between the PVA and the VA assessed by the naïve panel (N=468), yielding r=0.93 (P<.001). This strict train/test separation ensures that the reported correlation reflects a genuine out-of-sample prediction and provides a robust safeguard against overfitting.

    Table 3. Overall agreement statistics for the extended facial sign/imaging condition analyses.
    Facial sign/imaging conditionMAEaRMSEbICCc (95% CI)Percentage within 1 grade (%)
    Crow’s feet wrinkles (n=173)0.3140.4780.949 (0.930-0.960)93
    Eyebags (n=173)0.1840.3650.906 (0.870-0.930)96
    Pigmentary spots (polarized; n=186)0.8791.0800.659 (0.130-0.840)66
    Pigmentary spots (nonpolarized; n=186)0.2260.3060.954 (0.940-0.970)98
    Cheek skin pores (n=186)0.2390.3390.895 (0.860-0.920)98
    Marionette lines (n=186)0.2180.3210.976 (0.970-0.980)98
    Forehead wrinkles (n=92)0.2780.3590.937 (0.910-0.960)98
    Glabellar wrinkles (n=92)0.2400.4680.912 (0.870-0.940)98
    Interocular wrinkles (n=92)0.4010.5440.909 (0.860-0.940)96
    Nasolabial fold (n=186)0.2150.3380.970 (0.960-0.980)96
    Ptosis (n=92)0.3030.4660.886 (0.830-0.920)93
    Periorbital wrinkles (n=186)0.2010.2830.967 (0.960-0.970)99
    Perioral lip wrinkles (n=92)0.3410.4930.878 (0.820-0.920)93
    Diffuse redness (nonpolarized; n=92)0.2040.2580.948 (0.920-0.970)100
    Diffuse redness (polarized; n=92)0.1770.2180.963 (0.940-0.980)100

    aMAE: mean absolute error.

    bRMSE: root mean squared error.

    cICC: intraclass correlation coefficient.

    Table 4. Definition of coefficients for the equation for predicted visible age (PVA) based on facial clinical signs.
    Facial signs/coefficientsABC
    Crow’s feet wrinkles−1.4318.0714.13
    Eye bags−4.7024.2634.38
    Marionette lines−1.0712.1334.22
    Forehead wrinkles−1.7815.5933.33
    Glabellar wrinkles−2.7520.0632.19
    Interocular wrinkles−1.6016.0328.73
    Nasolabial fold−1.5514.9034.99
    Ptosis of the lower part of the face−2.2218.4531.89
    Periorbital wrinkles−2.1019.0427.57
    Perioral lip wrinkles−3.2722.8230.71

    Statistics

    The level of agreement between the grades provided by the AI-based system and the experts was analyzed globally and stratified by ethnic ancestry based on metrics including MAE, root mean squared error (RMSE), intraclass correlation coefficient (ICC), Bland-Altman analysis, and calibration statistics. The number of paired observations varied by sign, reflecting the imaging protocol (not all signs were captured for every participant at every session). MAE was defined as the mean of the absolute differences between the automatically predicted grades and the averaged expert reference grades. For each facial sign, the proportions of individual predictions with an absolute error below 0.5, 1.0, and 1.5 grading units were also generated.

    The correlations between PVA and VA, on the one hand, and PVA and RA, on the other hand (2 paired methods to be compared), were computed. Their significance was assessed using Spearman coefficient correlation tests overall, considering all photographed volunteers (N=468).


    Assessment Panel

    The standardized photographic dataset consisted of 468 North Americans aged between 18 and 80 years old, including 100 women of East Asian ancestry, 125 of African ancestry, and 243 of European ancestry, spread across the different age groups, as shown in Table 1.

    The expert grading panel comprised 15 trained experts, including dermatologists with previous experience in facial sign grading. The naïve panel included 253 North American women from different ethnic profiles, aged between 20 and 60 years old, with approximately 50% aged between 20 and 39 years and 50% aged between 40 and 60 years.

    Experts Panel Consensus and Variability

    Table 5 illustrates the scoring dataset established by the expert panel, including MAEs calculated at different grade thresholds according to the Skin Aging Atlas classification. Most facial skin signs appear to be correctly assessed, with an acceptable MAE of around 0.5, except for periorbital wrinkles and the density of pigmentary spots (approximately 0.6).

    Table 5. Values of the MAE in the scores of 13 facial signs recorded by 15 experts and dermatologists at different selected thresholds.
    Facial signMAEaPercent |error|<0.5Percent |error|<1.0Percent |error|<1.5
    Panel A: East Asian ancestry
    Forehead wrinkles0.44669399
    Glabellar wrinkles0.37749699
    Crow’s feet wrinkles0.52588897
    Periorbital wrinkles0.59548294
    Nasolabial fold0.44679097
    Marionette lines0.52598896
    Pigmentary spots (density)0.62518093
    Ptosis of the lower part of the face0.49579199
    Diffuse redness0.377297100
    Cheek skin pores0.52578797
    Interocular wrinkles0.55558696
    Perioral lip wrinkles0.43689299
    Eye bags0.45669098
    Panel B: African ancestry
    Forehead wrinkles0.46639098
    Glabellar wrinkles0.39739599
    Crow’s feet wrinkles0.53578796
    Periorbital wrinkles0.59528495
    Nasolabial fold0.49658896
    Marionette lines0.51628896
    Pigmentary spots (density)0.66467992
    Ptosis of the lower part of the face0.49639198
    Diffuse redness0.37799497
    Cheek skin pores0.59538294
    Interocular wrinkles0.51578897
    Perioral lip wrinkles0.47629199
    Eye bags0.43699297
    Panel C: European ancestry
    Forehead wrinkles0.48638997
    Glabellar wrinkles0.38739599
    Crow’s feet wrinkles0.49618998
    Periorbital wrinkles0.61538293
    Nasolabial fold0.45659097
    Marionette lines0.53578796
    Pigmentary spots (density)0.55558595
    Ptosis of the lower part of the face0.55548797
    Diffuse redness0.41679499
    Cheek skin pores0.49609098
    Interocular wrinkles0.53578897
    Perioral lip wrinkles0.45649298
    Eye bags0.39739398

    aMAE: mean absolute error.

    Prediction Potential of the Automatic Scoring System

    Table 6 summarizes the performance of the AI-based automatic grading system. On average across all 13 signs, considering the condition of MAE<1, which is the accepted error during a live evaluation by a clinician based on a standardized photographic scale, reliable accuracy is achieved for East Asian (~98%), African ancestry (~96%), and European (~97%) vs experts’ clinical ground truth.

    Table 6. Performance of the AI-based automatic grading system.
    Facial signMAEaPercent |error|<0.5Percent |error|<1.0Percent |error|<1.5
    Panel A: East Asian ancestry
    Forehead wrinkles0.2488100100
    Glabellar wrinkles0.1894100100
    Crow’s feet wrinkles0.28789797
    Periorbital wrinkles0.2190100100
    Nasolabial fold0.228597100
    Marionette lines0.238795100
    Pigmentary spots (density)0.258795100
    Ptosis of the lower part of the face0.278894100
    Diffuse redness0.15100100100
    Cheek skin pores0.16100100100
    Interocular wrinkles0.3862100100
    Perioral lip wrinkles0.3381100100
    Eye bags0.218697100
    Panel B: African ancestry
    Forehead wrinkles0.38892100
    Glabellar wrinkles0.4779296
    Crow’s feet wrinkles0.219498100
    Periorbital wrinkles0.1794100100
    Nasolabial fold0.228894100
    Marionette lines0.189498100
    Pigmentary spots (density)0.248496100
    Ptosis of the lower part of the face0.29819296
    Diffuse redness0.2100100100
    Cheek skin pores0.29909696
    Interocular wrinkles0.44779296
    Perioral lip wrinkles0.248596100
    Eye bags0.169396100
    Panel C: European ancestry
    Forehead wrinkles0.2887100100
    Glabellar wrinkles0.1796100100
    Crow’s feet wrinkles0.35789098
    Periorbital wrinkles0.228699100
    Nasolabial fold0.21899799
    Marionette lines0.249199100
    Pigmentary spots (density)0.2192100100
    Ptosis of the lower part of the face0.318196100
    Diffuse redness0.1898100100
    Cheek skin pores0.259299100
    Interocular wrinkles0.387796100
    Perioral lip wrinkles0.39728996
    Eye bags0.17919898

    aMAE: mean absolute error.

    Considering the average of all ethnicities, the MAE values were below 0.3 for 11 signs, except for interocular and perioral lip wrinkles, which showed higher values. Interestingly, the comparison of results between Tables 5 and 6 clearly indicates that the automatic grading system induces less variability and better reliability than that of the panel of trained experts, regardless of ethnicity or facial clinical sign. Indeed, the MAE values obtained suggest 2 or 3 times lower errors and variability for the automatic grading system compared with the experts’ panel.

    Additional Agreement Analyses

    Additional analyses were conducted to further investigate how well the automatic grading system aligned with the expert assessments. Since pigmentary spots and vascular disorders were evaluated under both polarized and nonpolarized imaging conditions, these analyses included 15 facial aging sign/imaging condition pairings derived from the 13 facial signs assessed in the study.

    Among these 15 sign/imaging condition combinations, the MAE was 0.292 grading units, and the RMSE was 0.419 grading units. The higher RMSE compared to MAE suggests the presence of some larger outlier prediction errors even though the average error remained low. Furthermore, more than 90% of the automatic predictions were within 1 grading unit of the expert reference for 14 of the 15 sign/imaging condition combinations. The main outlier was CPS under polarized conditions, where only 66% of the predictions were within 1 grading unit of the expert score.

    The ICC values appeared to show good to excellent agreement for most of the evaluated signs. For CPS, especially under polarized lighting, the agreement was comparatively lower, with an ICC of approximately 0.66. Signs like upper lip wrinkles, ptosis, and cheek pores exhibited agreement within the good range, while most of the remaining sign/imaging condition combinations showed excellent agreement. The full ancestry-stratified statistics are provided in Multimedia Appendix 1.

    Bland-Altman analyses showed mean biases close to zero for most signs, indicating little systematic overestimation or underestimation at the group level. Some proportional bias was observed at higher severity grades, whereas the limits of agreement were generally within approximately ±1 grading unit (Multimedia Appendix 2). Calibration plots indicate good linearity between the AI-automated grading and expert scores for most signs, with regression slopes close to 1.0 and intercepts close to 0. Similar to observations with higher MAE and lower ICC, deviations from the identity line were noted for CPS (polarized), whereas nonpolarized images demonstrated better-calibrated predictions for pigmentation-related signs (Multimedia Appendix 3).

    PVA

    Figure 2 presents correlation coefficients between the 13 facial clinical grades predicted by the automatic grading system and the VA assigned by the naive panel. It clearly shows that most facial signs are interrelated in the perception of age, with the exception of CSP, PS (density), and diffuse redness (DR). They will, therefore, not be retained in the subsequent calculation of PVA.

    Figure 3 presents correlation coefficients between RA and the ages perceived by a panel of 253 US women. Although these appear strongly correlated (r=0.94; P<.001), the 468 photographed women were perceived to be older than they are by about 10 years in a large majority.

    ‎
    Figure 2. Correlation coefficients between all facial signs predicted by the AI-based automatic grading system and visible age (VA). The colored boxes show the levels of correlation with VA: green: ≥0.70; yellow: ≤0.5 and >0.7; orange: ≤0.2 and >0.5; red: ≥−0.1 and <0.2. CFW: crow’s feet wrinkles; CSP: cheek skin pores; DR: diffuse redness; EB: eye bags; FW: forehead wrinkles; GW: glabellar wrinkles; IOW: interocular wrinkles; ML: marionette lines; NF: nasolabial fold; PLF: ptosis lower part face; PLW: perioral lip wrinkles; PS: pigmentary spots; PW: periorbital wrinkles.
    ‎
    Figure 3. Correlation between the visible ages (assessed by the naïve panel) of the 468 US participants and their real ages (r=0.94).

    For each of the 10 retained signs, a second-order regression of the type Y = Ax2 + Bx + C (where x is the grade established by the automatic grading system) is adopted, where A, B, and C coefficients of each of the retained signs are indicated in Table 4. The sum of all values of all Aixi2 + Bixi + C, when averaged, allows obtaining the PVA of each of the 468 US participants.

    PVA was strongly correlated (r=0.93; P<.001) with the VA evaluated by the naïve panel (Figure 4). Nevertheless, this figure shows that despite the good correlation, gaps were found within the 40- and 60-year age ranges. However, about 70% of assessments were within 5 years of the VA assessed by the naïve panel.

    ‎
    Figure 4. Correlation between predicted visible ages of the 468 US participants and their visible ages, as assessed by the naïve panel (r=0.93).

    Principal Findings

    The present work aimed at analyzing how an AI-based process could predict the VA of a given participant from a standardized full-face photograph in a robust manner, while reducing the maximum of biases arising from human assessment. The study paired standardized image capture, clinical grading by experts, an independent naïve panel for VA assessment, and automatic grading of facial aging signs. The NEXA photographic system used in this study provided standardized image acquisition and good image quality together with its ease of use (rapidity, reproducibility, etc), that is, the primary conditions to allow the most precise analysis by a naïve panel, an expert group, or an automatic AI-based process. It is nevertheless clear that attributing a precise VA to a person from a single facial aspect is by no means an easy task. In real life, other criteria are used, such as hair presence or style, body positioning (straight or bent), walking pace, etc, as well as reflecting cultural, geographical, or environmental specificities. Facial traits remain, however, a suitable model of skin aging, abundantly studied as they integrate intrinsic biological changes and external impacts (sun exposure and lifestyles), as extensively reviewed [41].

    A relatively low prediction errors for most facial signs were obtained when the results provided by the automatic system were compared with those from the expert panel. These findings were further confirmed by additional analyses, with good to excellent agreement obtained for most evaluated sign/imaging condition combinations. In the Bland-Altman analyses, the mean biases approached zero for most signs, and calibration overall was satisfactory. Still, performance was not homogeneous across all facial signs. Pigmentary spots exhibited lower agreement when assessed under polarized lighting. These findings suggest that the performance of the automatic grading system partially depends on the facial sign and imaging condition evaluated.

    VA and Human Perception

    With regard to the assessments by the naïve panel, comparing VA and RA, although highly correlated (r=0.94; P<.001), leads to blurred results. Why this panel “viewed” almost all participants as much older (by 10 y or more at younger ages) than they really were is an open question [9,12]. In contrast to the blind/neutral feature of the automatic system, the perception of the panel is processed at a glance that rapidly integrates all facial aspects, which might attribute more “weight” to some facial regions or signs (upper or lower face, ptosis, etc) than others; this is a possibility that could be checked through an eye-tracking methodology (not explored here). To what extent such focus on specific facial regions obeys cultural or socioeconomic influences is another open question. In contrast, the automatic system provides, in the end, a realistic predictive estimate of the VAs of the 468 women studied, although Figure 3 shows larger gaps in the 40 and 60 years range, noting that these follow a normal Gaussian distribution. It is likely that such a result first illustrates the period when the process of facial aging leads to a larger variability among women with regard to hormonal changes (perimenopausal and postmenopausal periods). The latter indeed induce profound and widely varying effects among the women concerned in both physiological (skin included) and psychological status [42-45].

    Second, the spontaneous attribution of a VA to a person likely differs in complexity according to age; that is, it is probably more precisely established at age ranges where markers of aging are either not very pronounced (18–30 y) or already settled (>60 y). It is noteworthy that 3 signs related to homogeneity (PS, CSP, and DR) showed low correlations with VA and were not included in the PVA calculation. When different ancestries are considered together, facial signs such as wrinkles and sagging may contribute more consistently to the overall perception of VA.

    Performance Across Ancestry Groups

    The choice of performing the study in the United States enabled the inclusion of women from different ancestry groups and age ranges. The automatic grading system showed generally high agreement with expert assessments across the ancestry groups represented in the study. The additional analyses, however, showed variation depending on both facial sign and ancestry group. Hence, the overall performance estimates should not be read as exactly the same for every subgroup. In a diverse US female population, similar work had also demonstrated the feasibility of automated grading of facial signs using selfies, which depended on facial features, especially those associated with pigmentation [46].

    These findings are important in the context of potential algorithmic bias. Although the inclusion of women of East Asian ancestry, African ancestry, and European ancestry increased diversity within the development dataset, it does not establish fairness or generalizability to all populations. In particular, performance across age groups was not separately characterized to the same extent as ancestry-related performance. Further evaluation on larger and more independent datasets will, therefore, be needed to assess whether performance stays comparable across the different demographic groups.

    Potential Relevance to Digital Dermatology and Health Care Setting

    These results demonstrate that the application of standardized image analysis provides a reproducible digital outcome to assess visible facial aging. This approach might be useful for complementing traditional clinical grading by providing standardized quantitative assessment of facial aging signs and VA in the fields of dermatology and cosmetic dermatology.

    Further potential applications could be found in the longitudinal monitoring of visible facial changes or in standardized assessments of appearance-related outcome measures in research and clinical practice. From this perspective, emerging concepts such as intelligent mirrors can offer a framework for the repeated, standardized, and noninvasive acquisition of skin phenotypes over time, supporting a transition from single time point assessments toward longitudinal skin monitoring [47]. These tools might additionally aid practitioners’, patients’, and consumers’ communication about visible change.

    The present study has not evaluated diagnostic accuracy for diseases, nor has it explored the influence on clinical decisions, treatment outcomes, patient management, or health system outcomes. Thus, at this stage, the system cannot be regarded as being a diagnostic or prognostic tool. Furthermore, since the evaluation used standardized photographs captured with one imaging device, applicability for use at POS, in clinics, on smartphones, or in consumer-facing situations under less standardized conditions requires further demonstration.

    The present approach may potentially be integrated with other measurement modalities (eg, cutaneous biomarkers) to explore relationships between visible facial aging and biological processes. However, such work represents a future research perspective.

    Ethical Considerations

    While the application of AI-based facial grading and PVA in dermatology offers great promise, several ethical considerations must be addressed. First, more longitudinal studies are needed to validate the long-term imbrication between biological age and VA. While some treatments have been shown to influence biomarkers in the short term, their long-term impact on skin health and longevity remains unclear. Additionally, the use of biological age or VA as metrics for treatment decisions raises ethical questions, such as whether patients should be informed of their biological age if it differs significantly from their chronological age, and if so, how this information should be communicated alongside VA and how it might influence a patient’s perception. Another challenge lies in the accessibility of these advanced, expensive, and specialized tools in POS or clinics, limiting their availability to a broader patient population.

    Addressing these barriers, for example, with nomad solutions on smartphones, is essential to ensure equitable access to these innovations, preventing disparities between patients who can afford high-end aesthetic or dermo-cosmetic treatments and those who cannot. Additionally, differences in life expectancy between men and women, influenced by hormonal and environmental factors, add complexity to longevity research. These variations underscore the importance of developing sex-specific approaches for personalized and effective treatments [48].

    Limitations and Future Perspectives

    Several limitations of the study should be noted. The data are restricted to women in North America who were photographed with a standardized imaging device. Although women across various ages and ethnicities were included, this diversity does not, by itself, establish generalizability to men, populations outside North America, other image-capture devices, or less-standardized real-life conditions. Additionally, there was no validation of results on an independently collected dataset.

    Further investigations will be necessary, including the use of independent datasets containing more varied human populations, as well as alternative image acquisition devices under variable imaging conditions, before broader clinical or consumer-facing applications can be considered.

    Overall performance was satisfactory across all ancestry groups included in the study, with variations depending on the facial signs and subgroup. To further assess the robustness of, and possible biases associated with, each group, additional investigations would be necessary, as the current analyses are based on limited datasets that lack balance. The naïve panel consisted solely of North American women aged 20 and 60 years, and thus the VAs might partially represent their specificities and perceptions.

    It is worth noting that the algorithm trained to predict VA is based on subjective human evaluations. As such, PVA reflects an assessment of visible changes related to chronological age rather than objectively measurable signs of biological aging. No biological, proteomic, or physiological markers were assessed in the present study, which further implies that while the idea of VA being a potential “skin clock” is appealing, its relationship with biological age, health span, or overall health should currently be considered exploratory. This AI-based process could, in future studies, be associated concomitantly with biological measurements extracted from the skin surface. This new tool will support engagement, compliance, and long-term monitoring. It is nevertheless clear that this system calls for some extensions, such as its application to men—for whom technical challenges regarding beards, for example, could be addressed in a new AI-based model—or a deeper study of countries’ specificities to tackle cultural shifts through larger cohorts of the major ethnic backgrounds.

    In addition, building a global ecosystem to accompany patients and consumers through integrative longevity strategies is at hand, as we compare the MAE obtained here for the automatic grading system across all ethnicities (~97%) with that obtained with a similar system developed with Skin Aging Atlas references for smartphone selfies (~98%) [38,49]. This aspect will allow us to go further into North American women and controlled imaging. Another instrumental introduction will be leveraging technologies such as GANs for photorealistic and accurate personalized simulations [17,22,50] of future visible manifestations of longevity treatments of aging root causes—or the absence of them.

    Conclusions

    The present study demonstrated that an AI-based process can automate the evaluation of some features of visible facial aging with relatively low prediction errors and provide a predicted VA that has a strong correlation with human-assessed VAs under standardized image acquisition conditions. The automatic grading system demonstrated good agreement with expert assessments across diverse ancestral backgrounds represented in the study, although performance varied by facial sign and subgroup.

    These findings support the potential use of AI-based facial assessment as a reproducible digital outcome for facial aging research and longitudinal monitoring. Nevertheless, external validation in independent, more diverse, and less standardized conditions is required before any wider clinical or consumer-facing rollout. As predicted, VA is derived from quantified facial clinical signs; it reflects measurable features of visible facial aging, while its relationship with biological age, overall health, or its potential as a validated “skin clock” remains to be established through future studies integrating biological biomarkers.

    Acknowledgments

    The authors want to thank Mr Didier Saint-Leger, Mr Soheil Seyfaie, Mr Aurelien Utzeri, Mr Paul Tartrat, Mr Cedric Larcher, Mr Chengda Ye, Mrs Celeste Grossgold, Mrs Olivia De Drezy, Mr Rawad Abdayem, Mrs Madiiha Mandary, Mr Thierry Lageat, Mr Frederic Woodland, Mr Adrien Chretien, Mrs Komal Patel, Mrs Margarita Yatskayer, Mrs Vania Lacascade, and Mr Guive Balooch not only for their great help in completion of this paper and global study but also for their strong support and enthusiasm.

    AI was used as part of the study methodology for facial assessment and analysis, as described in the manuscript. No generative AI tools were used to generate scientific data, results, or conclusions.

    Funding

    All costs of the present work were entirely covered by L’Oréal Research & Innovation.

    Data Availability

    The data that support the findings are available from the corresponding author upon reasonable request.

    Authors' Contributions

    FF, PAB, and TL were involved in the design, conduct, interpretation, preparation, and decision to submit the manuscript for publication. They had full access to all the data in the study and took responsibility for the integrity of the data and the accuracy of the data analysis. AA was involved in the conduct, collection, management, and analysis of the quiz study. TL, PM, JD, and SB designed algorithms and improved them through the training and validation phases. HJ and LB conducted data analysis for assessments of automatic grading system performance.

    TD was involved in the design and interpretation. DH and HC made substantial contributions to revising the manuscript.

    All authors reviewed and approved the manuscript before submission and have agreed to be accountable for their contributions.

    Conflicts of Interest

    FF, PAB, TL, SB, JD, HJ, LB, PM, and TD are employees of the L’Oréal group. The other authors declare no conflicts of interest.

    Multimedia Appendix 1

    Full agreement statistics stratified by ancestry group.

    DOCX File, 34 KB

    Multimedia Appendix 2

    Bland-Altman plots.

    DOCX File, 437 KB

    Multimedia Appendix 3

    Calibration plots.

    DOCX File, 524 KB
    1. World Health Statistics 2024: monitoring health for the SDGs, Sustainable Development Goals. World Health Organization; 2024. URL: https://iris.who.int/server/api/core/bitstreams/74b12494-f213-4b5b-9533-18442147e1fb/content [Accessed 2026-10-03]
    2. Masfiah S, Kurnialandi A, Meij JJ, Maier AB. Definitions of healthspan: a systematic review. Ageing Res Rev. Sep 2025;111:102806. [CrossRef] [Medline]
    3. Garmany A, Yamada S, Terzic A. Longevity leap: mind the healthspan gap. NPJ Regen Med. Sep 23, 2021;6(1):57. [CrossRef] [Medline]
    4. Haykal D, Flament F, Othman K, Siow R. Lifespan, healthspan, and the expanding role of cosmetic dermatology in longevity science. J Cosmet Dermatol. Dec 2025;24(12):e70610. [CrossRef] [Medline]
    5. Haykal D, Flament F, Mora P, Balooch G, Cartier H. Unlocking longevity in aesthetic dermatology: epigenetics, aging, and personalized care. Int J Dermatol. Dec 2025;64(12):2204-2214. [CrossRef] [Medline]
    6. López-Otín C, Blasco MA, Partridge L, Serrano M, Kroemer G. Hallmarks of aging: an expanding universe. Cell. Jan 19, 2023;186(2):243-278. [CrossRef] [Medline]
    7. Jacques E, Herzog C, Ying K, et al. Invigorating discovery and clinical translation of aging biomarkers. Nat Aging. Apr 2025;5(4):539-543. [CrossRef] [Medline]
    8. Biomarkers of Aging Consortium, Herzog CMS, Goeminne LJE, Poganik JR, et al. Challenges and recommendations for the translation of biomarkers of aging. Nat Aging. Oct 2024;4(10):1372-1383. [CrossRef] [Medline]
    9. Matts PJ, Fink B, Grammer K, Burquest M. Color homogeneity and visual perception of age, health, and attractiveness of female facial skin. J Am Acad Dermatol. Dec 2007;57(6):977-984. [CrossRef] [Medline]
    10. Porcheron A, Mauger E, Soppelsa F, et al. Facial contrast is a cross-cultural cue for perceiving age. Front Psychol. 2017;8:1208. [CrossRef] [Medline]
    11. Russell R, Porcheron A, Sweda JR, Jones AL, Mauger E, Morizot F. Facial contrast is a cue for perceiving health from the face. J Exp Psychol Hum Percept Perform. Sep 2016;42(9):1354-1362. [CrossRef] [Medline]
    12. Flament F, Abric A, Adam AS. Evaluating the respective weights of some facial signs on perceived ages in differently aged women of five ethnic origins. J Cosmet Dermatol. Mar 2021;20(3):842-853. [CrossRef] [Medline]
    13. Christensen K, Thinggaard M, McGue M, et al. Perceived age as clinically useful biomarker of ageing: cohort study. BMJ. Dec 10, 2009;339:b5262. [CrossRef] [Medline]
    14. Mekić S, Pardo LM, Gunn DA, et al. Younger facial looks are associated with a lower likelihood of several age-related morbidities in the middle-aged to elderly. Br J Dermatol. Feb 22, 2023;188(3):390-395. [CrossRef] [Medline]
    15. Turner C, Pardo LM, Gunn DA, et al. Deep learning predicted perceived age is a reliable approach for analysis of facial ageing: a proof of principle study. J Eur Acad Dermatol Venereol. Dec 2024;38(12):2295-2302. [CrossRef] [Medline]
    16. Dobos G, Lichterfeld A, Blume-Peytavi U, Kottner J. Evaluation of skin ageing: a systematic review of clinical scales. Br J Dermatol. 2015;172(5):1249-1261. [CrossRef] [Medline]
    17. Flament F, Bokaris PA, Despois J, et al. Ethnic differences in women’s perception of simulated facial aging over a 15-year horizon: a GAN-based model approach. Cosmetics. 2025;12(4):154. [CrossRef]
    18. Raynaud E, Bertrand L, Flament F, et al. A combined approach to predict tobacco-induced facial aging using dermatologist knowledge elicitation and generative models. Dermatol Ther (Heidelb). Jan 2026;16(1):629-636. [CrossRef] [Medline]
    19. Young AT, Xiong M, Pfau J, Keiser MJ, Wei ML. Artificial intelligence in dermatology: a primer. J Invest Dermatol. Aug 2020;140(8):1504-1512. [CrossRef] [Medline]
    20. du Crest D, Madhumita M, Enbiale W, et al. Skin and digital—the 2024 narrative. Mayo Clin Proc Digit Health. Sep 2024;2(3):322-330. [CrossRef] [Medline]
    21. Haykal D, Garibyan L, Flament F, Cartier H. Hybrid cosmetic dermatology: AI generated horizon. Skin Res Technol. May 2024;30(5):e13721. [CrossRef] [Medline]
    22. Despois J, Flament F, Perrot M. AgingMapGAN (AMGAN): high-resolution controllable face aging with spatially aware conditional GANS. Presented at: European Conference on Computer Vision; Aug 23-28, 2020. [CrossRef]
    23. Chan S, Reddy V, Myers B, Thibodeaux Q, Brownstone N, Liao W. Machine learning in dermatology: current applications, opportunities, and limitations. Dermatol Ther (Heidelb). Jun 2020;10(3):365-386. [CrossRef] [Medline]
    24. Jain A, Way D, Gupta V, et al. Development and assessment of an artificial intelligence-based tool for skin condition diagnosis by primary care physicians and nurse practitioners in teledermatology practices. JAMA Netw Open. Apr 1, 2021;4(4):e217249. [CrossRef] [Medline]
    25. Elder A, Ring C, Heitmiller K, Gabriel Z, Saedi N. The role of artificial intelligence in cosmetic dermatology—current, upcoming, and future trends. J Cosmet Dermatol. Jan 2021;20(1):48-52. [CrossRef] [Medline]
    26. Goh CL. Will digitization, big data and artificial intelligence- and deep learning-based algorithm govern the practice of medicine? J Eur Acad Dermatol Venereol. Jul 2022;36(7):947. [CrossRef] [Medline]
    27. Liu Y, Jain A, Eng C, et al. A deep learning system for differential diagnosis of skin diseases. Nat Med. Jun 2020;26(6):900-908. [CrossRef] [Medline]
    28. Kim S, Cho D, Kim J, et al. Smartphone-based multispectral imaging: system development and potential for mobile skin diagnosis. Biomed Opt Express. Dec 1, 2016;7(12):5294-5307. [CrossRef] [Medline]
    29. Seité S, Khammari A, Benzaquen M, Moyal D, Dréno B. Development and accuracy of an artificial intelligence algorithm for acne grading from smartphone photographs. Exp Dermatol. Nov 2019;28(11):1252-1257. [CrossRef] [Medline]
    30. Sangers TE, Kittler H, Blum A, et al. Position statement of the EADV artificial intelligence (AI) task force on AI‐assisted smartphone apps and web‐based services for skin disease. J Eur Acad Dermatol Venereol. Jan 2024;38(1):22-30. [CrossRef] [Medline]
    31. Flament F, Jacquet L, Ye C, et al. Artificial intelligence analysis of over half a million European and Chinese women reveals striking differences in the facial skin ageing process. J Eur Acad Dermatol Venereol. Jul 2022;36(7):1136-1142. [CrossRef] [Medline]
    32. Foucher A, Nouveau S, Piffaut V, Marque S, Aguilar L, Cavusoglu N. Clinical vs. chronological skin age: exploring determinants and stratum corneum protein markers of differential skin ageing in 351 healthy women. Sci Rep. Oct 9, 2024;14(1):23643. [CrossRef] [Medline]
    33. Klare BF, Burge MJ, Klontz JC, Vorder Bruegge RW, Jain AK. Face recognition performance: role of demographic information. IEEE Trans Inf Forensics Secur. 2012;7(6):1789-1801. [CrossRef]
    34. Drozdowski P, Rathgeb C, Dantcheva A, Damer N, Busch C. Demographic bias in biometrics: a survey on an emerging challenge. IEEE Trans Technol Soc. 2020;1(2):89-103. [CrossRef]
    35. Buolamwini J, Gebru T. Gender shades: intersectional accuracy disparities in commercial gender classification. Presented at: 1st Conference on Fairness, Accountability and Transparency (FAccT); Feb 23-24, 2018. URL: https://proceedings.mlr.press/v81/buolamwini18a/buolamwini18a.pdf [Accessed 2026-10-03]
    36. Bazin R, Flament F. Skin Aging Atlas Volume 2: Asian Type. Editions Med’Com; 2010. ISBN: 9782354030339
    37. Flament F, Bazin R, Qiu H. Skin Aging Atlas Volume 5: Photo-Aging Face & Body. Éditions Med’Com; 2017. ISBN: 9782354032463
    38. Jiang R, Kezele I, Levinshtein A, et al. A new procedure, free from human assessment that automatically grades some facial skin structural signs. Comparison with assessments by experts, using referential atlases of skin ageing. Int J Cosmet Sci. Feb 2019;41(1):67-78. [CrossRef] [Medline]
    39. He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. Presented at: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR); Jun 27-30, 2016. [CrossRef]
    40. Kingma DP, Ba J. Adam: a method for stochastic optimization. Presented at: 3rd International Conference on Learning Representations (ICLR 2015); May 7-9, 2015. URL: https://www.scirp.org/reference/ReferencesPapers?ReferenceID=1611214 [Accessed 2026-10-03]
    41. Swift A, Liew S, Weinkle S, Garcia JK, Silberberg MB. The facial aging process from the “inside out”. Aesthet Surg J. Sep 14, 2021;41(10):1107-1119. [CrossRef] [Medline]
    42. McNeil MA, Merriam SB. Menopause. Ann Intern Med. Jul 2021;174(7):ITC97-ITC112. [CrossRef] [Medline]
    43. Talaulikar V. Menopause transition: physiology and symptoms. Best Pract Res Clin Obstet Gynaecol. May 2022;81:3-7. [CrossRef] [Medline]
    44. US Preventive Services Task Force, Mangione CM, Barry MJ, Nicholson WK, et al. Hormone therapy for the primary prevention of chronic conditions in postmenopausal persons: US Preventive Services Task Force recommendation statement. JAMA. 2022;328(17):1740-1746. [CrossRef]
    45. Soares CN. Depression and menopause: an update on current knowledge and clinical management for this critical window. Med Clin North Am. Jul 2019;103(4):651-667. [CrossRef] [Medline]
    46. Flament F, Jiang R, Houghton J, et al. Accuracy and clinical relevance of an automated, algorithm-based analysis of facial signs from selfie images of women in the United States of various ages, ancestries and phototypes: a cross-sectional observational study. J Eur Acad Dermatol Venereol. Jan 2023;37(1):176-183. [CrossRef] [Medline]
    47. Haykal D, Flament F. Intelligent mirrors for longitudinal skin phenotyping. J Invest Dermatol. Jul 2026;146(7):1733-1735. [CrossRef] [Medline]
    48. Engelbrecht HR, Merrill SM, Gladish N, et al. Sex differences in epigenetic age in Mediterranean high longevity regions. Front Aging. 2022;3:1007098. [CrossRef] [Medline]
    49. Flament F, Lee YW, Lee DH, et al. The continuous development of a complete and objective automatic grading system of facial signs from selfie pictures: Asian validation study and application to women of three ethnic origins, differently aged. Skin Res Technol. Mar 2021;27(2):183-190. [CrossRef] [Medline]
    50. Kips R, Jiang R, Ba S, et al. Real‐time virtual‐try‐on from a single example image through deep inverse graphics and learned differentiable renderers. Comput Graph Forum. May 2022;41(2):29-40. [CrossRef]

    ‎
    CSP: cheek skin pores
    DR: diffuse redness
    ICC: intraclass correlation coefficient
    MAE: mean absolute error
    POS: points of sale
    PVA: predicted visible age
    RA: real age
    ReLU: rectified linear unit
    RMSE: root mean squared error
    VA: visible age

    Edited by Robert Dellavalle; submitted 12.Feb.2026; peer-reviewed by Liz González, Tirumala Ashish Kumar Manne, Valentina Palama; final revised version received 21.Sep.2026; accepted 23.Sep.2026; published 28.Oct.2026.

    Copyright

    © Frederic Flament, Panagiotis-Alexandros Bokaris, Tao Li, Aurelie Abric, Sileye Ba, Julien Despois, Hussein Jouni, Laudine Bertrand, Paul Michel, Thomas Dauxerre, Diala Haykal, Hughes Cartier. Originally published in JMIR Dermatology (http://derma.jmir.org), 28.Oct.2026.

    This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Dermatology, is properly cited. The complete bibliographic information, a link to the original publication on http://derma.jmir.org, as well as this copyright and license information must be included.