Quantifying expert knowledge to leverage new data: example of multi‐ethnic skin aging prediction with elicitation

Several areas of dermatology—such as dermatology for skin of color (SOC)—lack clinical data, although implicit expert knowledge exists. We evaluated the effectiveness of causal predictive models utilizing expert knowledge elicitation to address this data scarcity. To illustrate the benefits of quantifying the implicit expert knowledge, we used examples from a prior model constructed with 28 expert dermatologists from six countries spanning five continents to predict facial aging. This model enables personalized quantification of expert intuition according to different intrinsic and extrinsic aging factors. Based on Causal Bayesian Belief Networks, the model predicts how a baseline grade of facial aging signs is likely cumulatively impacted by environmental (extrinsic) and intrinsic modulators, using evolution probability elicited using the Delphi approach and constituting prior state-of-the-art knowledge.1 The output is a distribution of possible aging sign grades, each associated with a probability of occurrence. This model generated mean wrinkle scores and probabilities of reaching a defined wrinkle score at a given age, depending on photoprotection and smoking from age 18, for women of European, Asian, and African descent, as well as women with three different tones of SOC. It provided results consistent with existing literature and generated new data simulating long-term changes in a multiethnic population. Here, we presented four outcomes generated using our predictive model and forehead wrinkle score: sun exposure impact on European skin aging, interethnic comparison of aging and intraethnic comparison between various skin tones, and aging trajectory simulations in African descent skin. In European-descent skin, 83% of the aging signs were markedly impacted by sun exposure by ±0.1 severity grade, consistent with existing research.2 Moreover, in phototypes 1–4, 50% of aging signs were impacted by ≥2 grades, including glabella, under-eye, perioral wrinkles, cheek folds, pigmentary heterogeneity, and vascular disorders. We evidenced considerable variation in the probability of developing forehead wrinkles of comparable severity (Figure 1a) by age 50 in European, Asian, and African descent skins, depending on skin history (i.e., sun exposure, photoprotection, and smoking since the age of 18) (Figure 2a). Moreover, we generated predictions for grade-4 forehead wrinkles at age 50 in African-descent skin, considering skin history and precise skin tone, as measured by individual typology angle (ITA) (Figure 2a).3 This demonstrated the graded protective effects of melanin against forehead wrinkle development. In African descent skin of −30° < ITA < 10°, we generated aging trajectory simulations depicting the probability of reaching forehead wrinkles of grade ≥3 (Figure 1b) from age 18 to 65 (Figure 2b). The models could incorporate the impact of past harmful habits, such as insufficient photoprotection, to predict the effects of lifetime photoprotection practices on this wrinkle type. Our results align with published SOC data and underscore the importance of photoprotection for all skin types. The elicitation sessions enabled the creation of a model that can predict the distribution of forehead wrinkle grades, illustrating interindividual variability. These methods could apply to other dermatology subspecialties lacking longitudinal data, leveraging experts' knowledge of prognosis factors,4 for instance, to address the current underrepresentation of SOC populations in clinical studies. Moreover, these models could be used in various populations for photodamage-associated cancer risk prevention by adapting our model to correlate photoaging and cancer risk or combining the model with facial recognition technology to develop personalized prevention tools. In conclusion, quantifying expertise through elicitation can generate new data in clinical situations requiring trials with long-term follow-up of numerous subgroups. Furthermore, visualizing lifestyle impact through predictive evolution offers potential preventive applications. We thank the international working group of independent dermatologists who also contributed to the models presented in this study: Pierre André, Soraya Benamor, Pierre-Patrice Cabotin, Jin Chen, Zekai Chen, Katleen Conceiçao, Ncoza Dlova, Catherine Figoni-Laugel, Xianwei Han, Jenna Lester, Dongni Li, Isabelle Pansé, Mira Pavlovic-Ganascia, Sylvie Niverd-Rondelé, Nokubonga Khoza, Antoine Petit, Marie-Estelle Roux, Yu Shi, Isabelle Tardy-Bastide, Shanqing Wang, Youli Wang, Nan Xu, Yuehua Yan. We also thank Stéphanie Nouveau, Virginie Hourblin, Anissa Azouaoui, Janet Wangari Olivero, Stéphane Diridollou, and Céleste Grossgold for their support on the project and Santé Active Edition for medical writing assistance.