BRIEF REPORT ![](https://assets.reachrx.ai/images/10.1007%E2%98%8Ds13555-025-01570-4/_page_0_Picture_2.jpeg) # **A Combined Approach to Predict Tobacco‑Induced Facial Aging Using Dermatologist Knowledge Elicitation and Generative Models** Edouard Raynaud · Laudine Bertrand · Frederic Flament · Julien Despois · Sileye Ba · Emmanuelle Tancrède‑Bohin · Tao Li · Hussein Joun[i](http://orcid.org/0000-0003-4641-3161) Received: August 4, 2025 / Accepted: October 8, 2025 © The Author(s) 2025 ## ABSTRACT *Introduction***:** This research presents a novel, scientifcally validated facial aging simulator. It integrates two published methodologies for personalized aging predictions, particularly con‑ cerning the impact of tobacco. This contrasts with widespread simulations on social network and gaming applications, often lacking scientifc rigor and transparent methodology. *Methods***:** This simulator combines elicited knowledge from 28 expert dermatologists with an AI-powered image generation system. Based on this knowledge, the model predicts 15-year **Supplementary Information** The online version contains supplementary material available at [https://doi.org/10.1007/s13555-025-01570-4.](https://doi.org/10.1007/s13555-025-01570-4) E. Raynaud Bioclinical Research Center (CRB) of L'Oréal Advanced Research, Saint-Louis Hospital, Paris, France E. Raynaud · E. Tancrède‑Bohin L'Oréal Research and Innovation, Aulnay‑Sous‑Bois, France L. Bertrand · F. Flament · J. Despois · S. Ba · T. Li · H. Jouni (\*) L'Oréal Research and Innovation, Clichy, France e-mail: hussein.jouni@loreal.com E. Tancrède‑Bohin Dermatology Department, St Louis Hospital, APHP, Paris, France facial aging signs according to various intrinsic and extrinsic factors, providing personalized probabilities of reaching specifc aging stages (e.g., wrinkles, pigmentation). This work couples these 15-year predictions with a machine learn‑ ing model (AMGAN) that generates personalized facial aging simulation images. The AMGAN was initially trained on a dataset of 600 individuals with sign scores averaged from 15 expert graders. *Results***:** We present predicted independ‑ ent probabilities of reaching grades for differ‑ ent facial aging signs as a function of cumula‑ tive tobacco consumption. A simulation was performed using a 43-year-old subject's facial image to demonstrate the tool's capabilities in illustrating the impact of tobacco consumption. Confounding variables such as sun exposure, sunscreen use, and body mass index (BMI) were controlled, while cumulative smoking (less than 10 or greater than 20 pack-years) was varied. This tool effectively visualizes the known effects of smoking on aging and provides personalized quantifcation of its impact on specifc facial wrinkles. Recognizing that individuals prioritize different facial signs, the simulator allows users to focus on areas of personal concern. *Conclusion***:** By focusing on preserving skin longevity and preventing premature aging, the simulator, which can quantify the impact of smoking on individual aging trajectories, effectively motivates positive lifestyle changes. This personalized approach offers a promising preventative messaging strategy, particularly for audiences resistant to traditional methods, and strengthens the scientifc rationale regarding the impact of tobacco on specifc aging signs. # PLAIN LANGUAGE SUMMARY This work introduces a novel, scientifically robust facial aging simulator, addressing the lack of scientifc and dermatological validation in most existing aging simulators, especially those used in social networks and gaming appli‑ cations. Integrating two published research and leveraging dermatologist knowledge, the simu‑ lator generates personalized aging predictions (visualized as images and probabilities). This personalized, quantifed, and visual approach, focused on the impact of tobacco, empowers informed decisions through probability com‑ parisons of individual facial signs, motivating behavior change. Consequently, by focusing on preserving skin longevity and preventing pre‑ mature aging, it offers a novel preventative mes‑ saging tool, particularly effective for those unre‑ sponsive to traditional campaigns, by promoting healthier behaviors. The synergy between expert dermatological knowledge and AI-generated imagery reinforces the scientifc basis for the impact of smoking on facial aging. **Keywords:** Tobacco; Smoking; Skin aging; Predictive model; Elicitation; Artifcial intelligence; AI; Longevity ### **Key Summary Points** #### *Why carry out this study***?** Responding to a critical need for validated tools, this work introduces a novel, scientif‑ cally robust facial aging simulator, thereby addressing the lack of scientifc and derma‑ tological validation in most existing aging simulators. Integrating two published research and leveraging dermatologist knowledge, the simulator generates personalized aging predictions (visualized as images and probabilities). This personalized, quantifed, and visual approach, focused on the impact of tobacco, empowers informed decisions through prob‑ ability comparisons of individual facial signs, motivating behavior change. Consequently, it offers a novel preventa‑ tive messaging tool, particularly effective for those unresponsive to traditional campaigns, by promoting healthier behaviors. *What was learned fom the study***?** This study further demonstrated that the synergy between expert dermatological knowledge and AI-generated imagery rein‑ forces the scientifc basis for the impact of smoking on facial aging. # INTRODUCTION This work introduces a novel, scientifically robust facial aging simulator designed to visualize personalized aging predictions, focusing on the impact of tobacco on premature facial aging [[1–](#page-6-0)[4](#page-7-0)]. Smoking-induced extrinsic aging exacerbates chronoaging through multiple interconnected and cumulative mechanisms. The most described among these are oxidative stress with the release of free radicals, infammation that exacerbates infammaging, and the stimulation of metalloproteinases, which collectively contribute to aberrant cross-links and the degradation of elastic and collagen fibers. These physiopathological mechanisms result in cumulative damage, notably manifesting as premature facial aging, characterized by the premature formation and accentuation of certain wrinkles, as observed in the literature and in this study. Simulations claiming to predict facial aging, frequently promoted as utilizing artifcial intelligence (AI) technologies, prominently featured within social network and gaming applications, often lack scientifc validation and methodological transparency, despite their broad accessibility. Leveraging the implicit knowledge of experienced dermatologists, who have individually assessed 75,000–150,000 patients over their careers, can substantially improve the precision of these tools. We present results generated by a simulator that predicts, modifes, and visualizes personalized facial appearances by varying extrinsic factors at different intensities. ## METHODS The generation of personalized predictive images is achieved by combining two new tools devel‑ oped in collaboration with dermatologists and data science experts: the Elicitation Based Aging Simulator (EBAS) and AgingMapGAN (AMGAN) [\[5–](#page-7-1)[7\]](#page-7-2). First, predictive models for skin aging were developed using the EBAS [[5](#page-7-1), [6\]](#page-7-3), leveraging expertise from 28 dermatologists worldwide (France, USA, China, Brazil, South Africa, and Senegal). This approach enables the quantifca‑ tion of expertise through knowledge elicitation, employing the Delphi method. Causal Bayesian Belief Networks model the probabilistic and causal relationships between lifestyle factors (e.g., tobacco use, sun exposure, photoprotec‑ tion), chronological age, and resulting facial aging. The models predict the 15-year cumu‑ lative impact on signs of aging (wrinkles, pig‑ mentary disorders, ptosis, etc.) in women of European, Asian, and African descent, starting at age 18 years. The output includes a probabil‑ ity distribution of potential clinical aging grades for each sign, accompanied by a mean predicted score. Grounded in expert knowledge, this approach ensures reliable prediction of aging impacts based on established dermatological expertise, particularly regarding photoprotection and the effects of tobacco use. The quality of the models was evaluated by a cohort of independ‑ ent dermatologists. These dermatologists did not take part in the initial model design process and relied on established skin aging references for their assessments. The simulator's results on the impact of smoking, aging, and sun exposure on the premature appearance of facial wrinkles were consistent with the literature. These results also aligned with the opinions of expert dermatolo‑ gists involved in the model's creation, as well as independent dermatologists who evaluated the model without prior participation [[1](#page-6-0)–[6](#page-7-3)]. Second, AgingMapGAN (AMGAN) [[7](#page-7-2)], a conditional Generative Adversarial Network (GAN) machine learning model, generates personalized facial aging simulation images. It was initially trained on a dataset of 600 individuals with sign scores averaged from 15 expert graders. The authors did not consider it necessary to submit this study to an ethics committee. All volunteers who participated in the generation of images transformed by the AMGAN simulator signed an image rights consent form authorizing both scientific publication and educational presentations. The images used to build the model are acquired using the Canfeld sensor. This device enables the acquisition of images under controlled illumination conditions. Thus, there are no illumination-related variations that would necessitate a very large dataset for model training. Furthermore, we employed standard data augmentation techniques commonly used in image-based prediction models to train the models, specifcally image mirroring (when possible), random cropping, and color/ illumination perturbations. This AI approach uses the predictions from elicitation models to modify an image accordingly, thus visually demonstrating the evolution of clinical signs (skin aging sign crops, such as periocular, glabellar, nasolabial, interocular, undereye, peribuccal, and forehead wrinkles). As quantitative evaluation, we calculated the mean average error (MAE). The Spearman correlations between clinical sign ground truth and predicted values were determined. Spearman correlations are greater than 89% for the clinical signs used in this article (see Table S1 in the electronic supplementary material for details). The maximum of the skin atlases scores vary between 6 to 8. In this regard, clinical score prediction MAE values are low. Hence, it allows the user to visualize aging by taking into account individual and lifestyle factors (e.g., age, sun exposure, photoprotection, smoking). Both source image crops and target scores are defned according to skin atlas grading, which employs standardized and linear photographic scales that have been validated [\[8](#page-7-4)[–10\]](#page-7-5). ## RESULTS This work demonstrates the capabilities of a novel tool through a simulation showcasing the impact of tabacco consumption on facial aging. A high-resolution facial image (Fig. [1a](#page-3-0)) of a 43-year-old subject, acquired using the Nexa® device (Canfeld Beauty, Parsippany, USA), was transformed via AMGAN technology to simulate the addition of two grades to each clinical sign of aging. The selection of two grades resulted from discussions with dermatologists and clinicians familiar with these atlases, taking into account the typical worsening of wrinkles observed over a 15-year period. This represents a signifcant delta in the worsening of wrinkle scores, which is clearly visible across all wrinkles considered, as demonstrated in the photographs (Fig. [1b](#page-3-0)). This approach facilitates a clear visualization of these wrinkle deltas for readers unfamiliar with such atlases. For scientifc rigor and ease of protocol interpretation, it was crucial that the same grade delta be applied consistently across all wrinkles. Furthermore, during the elicitation sessions, der‑ matologists evaluated the worsening by score unit, consequently making a two-grade incre‑ ment the most coherent choice. A single-grade ![](https://assets.reachrx.ai/images/10.1007%E2%98%8Ds13555-025-01570-4/_page_3_Picture_5.jpeg) Fig. 1 Illustration of a two-grade wrinkle increase simulated by AMGAN. a Baseline facial wrinkles of a 43-yearold volunteer. b AMGAN-simulated image from a depicting a two-grade augmentation in all wrinkles. Tese fgures visually exemplify the specifc two-grade increase in wrinkles whose probabilities are calculated by the EBAS model and detailed in Fig. [2](#page-4-0) ![](https://assets.reachrx.ai/images/10.1007%E2%98%8Ds13555-025-01570-4/_page_4_Figure_1.jpeg) Fig. 2 Probabilistic comparison for a 43-year-old volunteer (Fig. [1](#page-3-0)a), illustrating the probability of achieving at least two additional wrinkle severity grades 15 years later (Fig. [1](#page-3-0)b) across multiple wrinkles, based on tobacco consumption. For instance, the probability to reach at least two additional grades for nasolabial fold wrinkles by age 58 is 28.54% for a non-smoker, whereas it almost doubles to 57.92% for a person with tobacco consumption exceeding 20 pack-years (>20 PY). Dermatologists involved in the delta would have been insuffcient to differen‑ tiate extrinsic aging factors, especially when considering the effects of chronological aging alone over a 15-year period. Figure [2](#page-4-0) displays the predicted probabilities of reaching these grades independently as a function of cumu‑ lative tobacco consumption, as determined by the EBAS models. Confounding variables were minimized by maintaining constant levels of sun exposure (low daily and recreational), sun‑ screen use (SPF 50+ during recreational expo‑ sure, none daily), and body mass index (BMI) (normal) over a 15-year period. Only cumulative smoking (less than 10 or greater than 20 packyears over 15 years) was varied. Starting with the baseline wrinkle scores visualized in Fig. [1](#page-3-0)a (age 43), Fig. [2](#page-4-0) shows the probabilities of reaching the increased grades depicted in Fig. [1](#page-3-0)b by age 58. For instance, the probability of reaching at least two additional grades for nasolabial fold wrinkles by age 58 is 28.54% for a non-smoker, whereas it almost doubles to 57.92% for a person with tobacco consumption exceeding 20 pack-years. This development of the EBAS simulator utilized wrinkle atlases to simulate the 15-year progression of individual wrinkles. Tis simulation was guided by their expert intuition for each specifc wrinkle type and informed by various intrinsic and extrinsic aging factors. Following extensive days of simulations and data acquisition, the simulator was able to calculate the probabilities of reaching a specifc wrinkle grade from a defned baseline grade tool not only visualizes the known effect of smoking on aging but also provides personalized quantification of the impact of smoking on individual facial wrinkles. ### DISCUSSION This innovative method, combining predictive image generation with probability assessments of personalized aging simulation, enhances the educational and preventive potential of the model by emphasizing the potential impact of lifestyle changes. This personalized and quan‑ tifed approach focuses on the probability of reaching specifc severity grades for various facial aging signs. Such a precise and tailored assess‑ ment is important, as each individual focuses on and prioritizes distinct facial aging manifesta‑ tions. Furthermore, by presenting simulations of everyday situations rather than extreme, life‑ long exposure, the simulator more accurately refects common scenarios, thereby enhancing its applicability for behavioral changes. Stud‑ ies have consistently shown the effectiveness of image-based education in promoting preventa‑ tive health behaviors, particularly concerning tobacco use [[11](#page-7-6)]. Indeed, images offer a stronger and more lasting impact than text in educating about risks and facilitating behavioral changes. This work contributes to this field by intro‑ ducing a new tool for leveraging the power of images in prevention, using data based on the clinical expertise of dermatologists. The present study minimized extrinsic aging confounding factors and focused on the impact of tobacco consumption. Although the effects of sun exposure were not analyzed in this study, the simulator is capable of predicting facial aging signs related to sun exposure. The simulator can further predict the likelihood and intensity of developing pigmentary lesions by varying levels of sun exposure or photoprotection. This version does not account for complexion or radiance, both of which are known to be affected by chronic smoking. Although not presented here, the simulator is also capable of generating multi-ethnic predictive visualizations. To ensure the widespread use of its tools in prevention campaigns, future development of the simulator must be capable of transforming smartphone and selfe images while maintaining scientifc robustness. Our model was developed for the Nexa Canfeld acquisition system, which allows one to capture facial images under controlled lighting conditions. To extend our models to uncontrolled lighting acquisition settings, training data in such conditions will be required. Similarly, to address clinical signs such as skin radiance and certain pigmentation disorders, annotated data is also required. We will organize data collections to build models that can reliably address adverse illumination conditions and new clinical signs [\[12\]](#page-7-7). Furthermore, while the models we propose are based on the GAN framework, newer generative models, such as denoising diffusion models, have shown higher performances for image generation. We have already started investigating the use of such models. The ability to process images captured under non-standardized conditions and the necessity of universal applicability across all skin tones are key development priorities for such tools. Work is currently underway to integrate these enhancements. A critical priority during this process is to ensure that the scientific robustness of these tools, along with the invaluable expertise of the dermatologists involved in their development, remains uncompromised. These anticipated improvements are expected to enable broader public accessibility, thereby significantly contributing to the dissemination of public health prevention messages. # CONCLUSION The combination of our knowledge elicitation data and AI-generated images clearly demon‑ strates the impact of smoking on facial aging and the potential to mitigate premature facial aging through behavioral change. This person‑ alized approach, tailored to individual lifestyles and expectations regarding signs of aging, underscores the value of our model. Realistic image simulation models based on expert der‑ matological knowledge prove valuable in pre‑ ventative strategies. Renewed approaches to preventative messaging, particularly for audi‑ ences previously unreceptive to traditional methods, are crucial for addressing public health concerns such as smoking, photopro‑ tection, premature skin aging, and promoting longevity. # ACKNOWLEDGEMENTS Authors would like to acknowledge the international working group of dermatologists [[5,](#page-7-1) [6](#page-7-3)]: Pierre André, Soraya Benamor, Pierre-Patrice Cabotin, Jin Chen, Zekai Chen, Katleen Conceiçao, Ncoza Dlova, Catherine Figoni-Laugel, Xianwei Han, Jenna Lester, Dongni Li, Fatimata Ly, Isabelle Pansé, Mira Pavlovic-Ganascia, Valérie Harvey, Sylvie Niverd-Rondelé, Nokubonga Khoza, Antoine Petit, Marie-Estelle Roux, Yu Shi, Isabelle Tardy-Bastide, Neelam Vashi, Shanqing Wang, Youli Wang, Nan Xu, Yuehua Yan. The authors also thank Matthieu Cassier, Anne Colonna, Stéphane Diridollou, Charles Gomes, Lionel Jouffe (Bayesia), Axel Sala-Martin, for their contributions to the models presented in this study or for their strong support during this project. The 28 dermatologists involved in earlier stages of this research were not included as co-authors in this new manuscript. Their contribution was limited to the development and validation of the EBAS predictive simulator, for which they were appropriately recognized as co-authors in the original publication [[5\]](#page-7-1). They were not involved in the current publication because modulators such as tobacco had already been validated and published in submission collaboration with them. Their input was not required for the operation of either the EBAS simulator (which generates probabilities for each facial sign) or the AMGAN simulator (which generates a full face with predicted grades) in this specifc study. *Author Contributions.* Conceptualization: Edouard Raynaud, Hussein Jouni. All authors (Edouard Raynaud, Laudine Bertrand, Frederic Flament, Julien Despois, Sileye Ba, Emmanuelle Tancrède-Bohin, Tao Li and Hussein Jouni) con‑ tributed to the study conception and design. Material preparation, data collection and anal‑ ysis were performed by all authors. The frst draft of the manuscript was written by Edouard Raynaud and all authors commented on previ‑ ous versions of the manuscript. All authors read and approved the fnal manuscript. *Funding.* Sponsorship for this study and Rapid Service Fee were funded by L'Oréal Research and Innovation, Clichy, France. *Data Availability.* The data that support the fndings of this study are not openly avail‑ able due to reasons of sensitivity and confden‑ tiality. This contains proprietary data which belongs to L'Oreal Research and Innovation. Data can be available upon reasonable request; please contact the corresponding author (Hus‑ sein Jouni). #### *Declarations* *Confict of Interest.* Hussein Jouni, Laudine Bertrand, Frederic Flament, Sileye Ba, Emma‑ nuelle Tancrede-Bohin, Tao Li are employees of L'Oréal. Julien Despois was employee of L'Oréal. Edouard Raynaud works as a consultant derma‑ tologist for L'Oréal. *Ethical Approval.* The authors did not con‑ sider it necessary to submit this study to an eth‑ ics committee. All volunteers who participated in the generation of images transformed by the AMGAN simulator signed an image rights con‑ sent form authorizing both scientifc publication and educational presentations. *Open Access.* This article is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License, which permits any non-commercial use, sharing, adaptation, distri‑ bution and reproduction in any medium or for‑ mat, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indi‑ cated otherwise in a credit line to the material. If material is not included in the article's Crea‑ tive Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain per‑ mission directly from the copyright holder. To view a copy of this licence, visit [http://creativeco](http://creativecommons.org/licenses/by-nc/4.0/) [mmons.org/licenses/by-nc/4.0/.](http://creativecommons.org/licenses/by-nc/4.0/) ## REFERENCES - 1. Krutmann J, Schikowski T, Morita A, Berneburg M. Environmentally-induced (extrinsic) skin aging: exposomal factors and underlying mechanisms. J Invest Dermatol. 2021;141:1096–103. - 2. Krutmann J, Bouloc A, Sore G, et al. The skin aging exposome. J Dermatol Sci. 2017;85:152–61. - 3. Kadunce DP, Burr R, Gress R, Kanner R, Lyon JL, Zone JJ. Cigarette smoking: risk factor for premature facial wrinkling. Ann Intern Med. 1991;114(10):840–4. - 4. Uitto J. Understanding premature skin aging. N Engl J Med. 1997;337(20):1463–5. - 5. Jouni H, Jouffe L, Tancrede-Bohin E, et al. Predict‑ ing the evolution of clinical skin aging in a multiethnic population: developing causal Bayesian networks using dermatological expertise. Skin Res Technol. 2024. [https://doi.org/10.1111/srt.13602.](https://doi.org/10.1111/srt.13602) - 6. Raynaud E, Vashi N, Harvey V, et al. Quantifying expert knowledge to leverage new data: example of multi-ethnic skin aging prediction with elicita‑ tion. Int J Dermatol. 2025;64(3):575–7. - 7. Despois J, Flament F, Perrot M, AgingMapGAN (AMGAN). High-resolution controllable face aging with spatially-aware conditional gans. computer vision and pattern recognition (cs.CV). [arXiv:](http://arxiv.org/abs/2008.10960) [2008.10960](http://arxiv.org/abs/2008.10960) - 8. Bazin R, Doublet E. Skin Aging Atlas, vol. 1. Cau‑ casian type. Paris: Editions Med'Com; 2007. - 9. Bazin R, Flament F. Skin Aging Atlas, vol. 2. Asian type. Paris: Editions Med'Com; 2010. - 10. Bazin R, Flament F, Giron F. Skin Aging Atlas, vol. 3. Afro-American type. Paris: Editions Med'Com; 2012. - 11. Fong GT, Hammond D, Hitchman SC. The impact of pictures on the effectiveness of tobacco warn‑ ings. Bull World Health Organ. 2009;87(8):640–3. - 12. Jiang R, Kezele I, Levinshtein A, et al. A new pro‑ cedure, free from human assessment that auto‑ matically grades some facial skin structural signs. Comparison with assessments by experts, using referential atlases of skin ageing. Int J Cosmet Sci. 2019;41(1):67–78. **Publisher's Note** Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affliations.