Exploring sensitive skin to design reliable measurements

Dear Editor, Sensitive skin (SS) is a sensory syndrome in which stimuli usually not inducing a reaction lead to unpleasant sensations (burning, pain, pruritus, tingling, etc.).1 Generally not associated with any clinical sign, SS can be triggered by chemical, environmental or physiological factors.2 Objective assessment of SS severity remains challenging because of its subjective nature and the lack of visible symptom. To fill that gap, this letter proposes two methods to quantitatively assess SS severity. A clinical evaluation was conducted during winter 2019 to assess SS in 90 Caucasian females (1865 years, mean ± SD = 42 ± 14, phototype I–IV). The subjects answered 60 yes/no questions to self-assess SS-inducing factors and reactions and 4 questions to estimate their sensation intensities (none, slight, moderate, marked). Based on these four last replies, a trained assessor classified subjects as non-sensitive or sensitive. A lactic acid sting test3 also provided a stinger/non-stinger classification. SpectraFace hyperspectral front face images4 of all subjects were collected. In the questionnaires, subjects tended to select the cheeks or the cheekbones as their most sensitive facial region (34% and 28% of replies, respectively) against the mouth contour, the nose, the forehead, and the eye contour. A mapping of frequency scores is presented in Figure 1. Comparison of the distribution of basic skin characteristics (typology, shine, redness, or dryness) between the sensitive panel (70 subjects classified by the assessor) and the non-sensitive one (20 subjects) revealed no difference. Even if these characteristics are combined in a linear regression, the classification prediction remains not reliable: the Matthew's correlation coefficient5 (MCC) equals to 0.49, with a good precision on sensitive panel (0.95) and a poor one on non-sensitive panel (0.45). Most subjects classified as sensitive were also declared stinger-negative (Figure 2), confirming that the sting test is an imperfect predictive tool.6 To better apprehend the subtle SS perception over the panel and using the existing data, a linear regression analysis was conducted to generate a continuous scale instead of a binary classification. The regression inputs were the questionnaire answers, turned into numerical scores (No = 0, Yes = 1, None = 0, Slight = 1, Moderate = 2, Marked = 3) and the targets were the trained assessor binary classification (Non-sensitive = 0, Sensitive = 1). For readability reasons, the output score was rescaled between 0 (not sensitive) and 10 (very sensitive), leading to a continuous SS score. By thresholding the SS score with a sensitive limit set to 2.3, defined by tests and trials, the expert categorisation could then be predicted with MCC of 0.97, a precision on sensitive panel of 0.98 and a precision on non-sensitive panel of 1.00 (Table 1). Therefore, the SS score based on questionnaires provided an improved description of self-assessed sensitivity. However, it requires a rather long questionnaire. To ensure unbiased answers, questions must be asked by a trained assessor also guaranteeing good comprehension. To strengthen this approach of SS measurement, it was decided to rely on objective data from hyperspectral images with the endpoint of more robust and precise SS assessment tool, using a machine-learning [ML] model. ML model would generalise the information from a group of subjects, making the small individual imprecisions in self-assessment not so impacting. From SpectraFace images, colour parameters (CIE L*a*b*, CIE L*C*h, ITA, IWA7), homogeneity (H76), oxygen saturation, melanin and haemoglobin concentrations (Figure 3) were computed, on several facial zones (cheeks, cheek bones, nasolabial folds). A regression multi-layer perceptron [MLP] was trained on the data from a subset of 80% randomly selected subjects from the panel, using the above-described continuous SS score as ground truth. A coefficient of determination of R2 = 0.93 was reached on the training set and of R2 = 0.81 (Figure 4) on the remaining 20% of the data. Therefore, this instrumental SS index, called Sym'Index, can reliably estimate SS based on the sole hyperspectral acquisition of the subject's face. To experiment with the index ability to follow SS evolution, another double-blind clinical evaluation was conducted during winter 2020 on 25 sensitive subjects selected from the previous panel. Two products were compared: a soothing formula (VIVO 1902.A.02) and a placebo (VIVO 1902.B.02). The soothing formula contains SymRelief 100 (INCI name: Bisabolol, Zingiber Officinale (Ginger) Root Extract). It is a synergistic blend of synthetic bisabolol with high purity and natural ginger extract. This ingredient offers a comprehensive solution targeting inflammatory mediators (IL-1α & PGE-2)8. SpectraFace images, acquired just before product application [D0], after 5 days [D5] and 10 days [D10] of repeated applications (twice a day [morning/evening] on a clean and dry skin, as much as needed until complete absorption), were analysed. SS indices were automatically computed for each subject's half-face and at each time point. Results (Table 2) indicate that, overtime, the soothing formula tends to induce a decrease of the instrumental SS index while the placebo would increase the index. Even if differences between products are not significant, the changes induced by each remain coherent and promising. The index is impacted enough by slight skin aspect changes to highlight product effects. To conclude, this initial analysis opens new perspectives of SS assessment, using for the first time hyperspectral face skin imaging and artificial intelligence model. This approach is combining classical appraisal of SS based on questionnaires and innovative, instrumental, and objective multiparametric analysis, highly correlating with expert's assessments. The proposed methodology could help to follow up the sensitivity evolution and to evaluate effects of soothing products. In the next steps, the instrumental SS index will be improved to underline finer evolutions, by increasing the panel size and distribution over sensitivity grades, by making the ground truth more robust and repeatable (asking several times the questionnaire), by adding other face areas and by removing redundant analysis parameters. The data that support the findings of this study are available from the corresponding author upon reasonable request.