Machine learning for the prediction of sunscreen sun protection factor and protection grade of UVA
Jiyong Shim Jun Man Lim Sun Gyoo Park
First published: 11 May 2019 https://doi.org/10.1111/exd.13958
Funding information:
The clinical studies presented in this article were funded by LG Household & Health Care, LTD.
SECTIONSePDFPDFTOOLS SHARE
Abstract
We report a prediction model for sunscreen sun protection factor (SPF) and protection grade of ultraviolet (UV) A (PA) based on machine learning. We illustrate with real clinical test results of UV protection ability of sunscreen for SPF and PA. With approximately 2200 individual clinical results for both SPF and PA level detection, individually, we were able to see that active ingredient information can provide accurate SPF and PA prediction rates through machine learning. Furthermore, we included four new factors—presence of pigment, concentration of pigment grade titanium dioxide, type of formulation and type of product—as additional information for the prediction model and were able to see increased prediction rates as results.
1 BACKGROUND
Overexposure to ultraviolet (UV) light can lead to harmful clinical consequences such as photo‐ageing and skin cancer, mostly due to UV rays, including UVA (320‐400 nm) and UVB (280‐320 nm).1, 2 Sunscreens play a crucial role in reducing the risk of melanoma and photo‐ageing by absorbing UV rays.3, 4 Sunscreens are labelled with sun protection factor (SPF) and protection grade of UVA (PA) levels according to their ability to protect against UVB and UVA, respectively. The commonly used SPF and PA simulators provide a final SPF or PA value based on complex calculations of each active ingredient's UV protection ability based on its concentration.5 Even though the concentration of UV filter substances is the principal influencer of the final SPF and PA values, various external factors are known to influence SPF and PA values.6-8 In order to increase the prediction rate of the SPF and PA prediction models based on machine learning, we included four important additional factors in the prediction model: presence of pigment, amount of pigment grade titanium dioxide (TiO2), type of formulation and type of product. In this study, we aimed to provide a rapid, inexpensive and efficient machine learning‐based SPF and PA prediction model, which will allow efficient development of new sun protection formulations.
1.1 Questions addressed
Can sunscreen prediction models based on machine learning provide cost‐effective, accurate and efficient results?
2 EXPERIMENTAL DESIGN
In this study, to increase the prediction rate of the SPF and PA prediction models based on machine learning, we included four important additional factors in the prediction model: presence of pigment, amount of pigment grade TiO2, type of formulation and type of product. Approximately 200 commercial and Ministry of Food and Drug Safety‐approved products, with values under SPF 70 (2377 individual clinical results) or PA 20 (2284 individual clinical results), were analysed by decision tree regression. Pearson's correlation coefficient was used for the correlation analyses employing IBM SPSS Statistics 25 (Armonk). SPF and PA testing was conducted in compliance with the regulatory requirements set forth by the International Sun Protection Factor Method (May 2006) or ISO 24444:2010 Test Method and ISO 24442 Test Method.9-11 Detailed information of the experimental procedures can be found within the Methods S1.
3 RESULTS
The combination of machine learning analysis and additional information of four factors increased both evaluation speed and prediction rate for the SPF and PA prediction models. In vivo SPF evaluations take at least 3 days for the UV radiation and MED evaluations, and in vivo PA evaluations take approximately 2‐24 hours for UVA radiation and minimum PPD evaluations.9-11 Even though in vitro SPF and PA are much faster than in vivo methods, they still require at least 15 minutes for the sunscreen sample to equilibrate at an ambient temperature to form a standard stabilized product film.12 The newly developed SPF and PA prediction model based on machine learning can provide prediction results for more than 1000 sunscreen formulations in less than a minute, which is much faster and more efficient than in vivo or in vitro SPF and PA evaluations.
The SPF prediction model showed good correlations with the SPF in vivo results, and the correlation improved when additional factors were included (Figure 1). The results for SPF as predicted by decision tree regression with active ingredient information showed very good correlation with SPF observed in vivo (r = 0.874; P < 0.01 for SPF in vivo and SPF predicted by machine learning) (Figure 1A). The coefficient of determination (R2), relative absolute error (RAE) and relative squared error (RSE) of active ingredient‐based correlation analysis were 0.76, 0.45 and 0.24, respectively (Table S2). The results of the SPF as predicted by decision tree regression with active ingredient information and additional factors showed a stronger correlation with SPF observed in vivo (r = 0.902; P < 0.01 for SPF in vivo and SPF predicted by machine learning) (Figure 1B). The R2, RAE and RSE of UV filter substances and additional factor‐based correlation analysis were 0.81, 0.41 and 0.20, respectively (Table S2). Thus, there was satisfactory correlation between the in vivo SPF values and decision tree regression predicted SPF values, and the correlation became stronger when SPF was predicted with the inclusion of additional factors.
image
Figure 1
Open in figure viewerPowerPoint
Correlations between sun protection factor (SPF) predicted by machine learning and SPF observed in vivo. A, Correlations between SPF predicted by machine learning based on concentration of ultraviolet (UV) filter substances with SPF observed in vivo. B, Correlations of SPF predicted by machine learning based on concentration of UV filter substances and four additional factors— presence of pigment, concentration of pigment grade titanium dioxide, type of formulation and type of product—with SPF observed in vivo.
The PA prediction model showed good correlations with the PA in vivo results, and the correlation became stronger when additional factors were included in the evaluation (Figure 2). The results of the PA as predicted by decision tree regression with active ingredient information showed a very good correlation with PA observed in vivo (r = 0.879; P < 0.01 for PA in vivo and PA predicted by machine learning) (Figure 2A). The R2, RAE and RSE of active ingredient‐based correlation analysis were 0.77, 0.45 and 0.23, respectively (Table S2). The results of the PA as predicted by decision tree regression with active ingredient information and additional factors showed a stronger correlation with PA observed in vivo (r = 0.894; P < 0.01 for PA in vivo and PA predicted by machine learning) (Figure 2B). The R2, RAE and RSE of UV filter substances and additional factor‐based correlation analysis were 0.80, 0.43 and 0.20, respectively (Table S2). Thus, there was satisfactory correlation between the in vivo PA values and decision tree regression predicted PA values, and the prediction rate became stronger with the inclusion of additional factors.
image
Figure 2
Open in figure viewerPowerPoint
Correlations between protection grade of ultraviolet A (PA) predicted by machine learning and PA observed in vivo. A, Correlations between PA predicted by machine learning based on concentration of ultraviolet (UV) filter substances with PA observed in vivo. B, Correlations of PA predicted by machine learning based on concentration of UV filter substances and four additional factors—presence of pigment, concentration of pigment grade titanium dioxide, type of formulation and type of product—with PA observed in vivo.
4 CONCLUSION
Machine learning analysis has not been used for the prediction of SPF and PA levels. With the available technology, we were able to develop complex SPF and PA prediction models based on machine learning. Our results suggest that the SPF and PA prediction models developed with decision tree regression analysis based on additional information—including concentration of UV filter substances, presence of pigment, amount of pigment grade TiO2, type of formulation and type of product—are valuable resources for the future development of sunscreens.
ACKNOWLEDGEMENTS
We thank Seonghwan Kong, Yeon Hee Kang and Jee Hee Jung for their assistance with the research.
CONFLICT OF INTEREST
The authors have declared no conflicting interests.
AUTHOR CONTRIBUTION
JS designed and performed the SPF and PA prediction model research and wrote the manuscript. JML designed and organized the research. SKP designed and analysed the research study.
Supporting Information
REFERENCES
Jiyong Shim Jun Man Lim Sun Gyoo Park
First published: 11 May 2019 https://doi.org/10.1111/exd.13958
Funding information:
The clinical studies presented in this article were funded by LG Household & Health Care, LTD.
SECTIONSePDFPDFTOOLS SHARE
Abstract
We report a prediction model for sunscreen sun protection factor (SPF) and protection grade of ultraviolet (UV) A (PA) based on machine learning. We illustrate with real clinical test results of UV protection ability of sunscreen for SPF and PA. With approximately 2200 individual clinical results for both SPF and PA level detection, individually, we were able to see that active ingredient information can provide accurate SPF and PA prediction rates through machine learning. Furthermore, we included four new factors—presence of pigment, concentration of pigment grade titanium dioxide, type of formulation and type of product—as additional information for the prediction model and were able to see increased prediction rates as results.
1 BACKGROUND
Overexposure to ultraviolet (UV) light can lead to harmful clinical consequences such as photo‐ageing and skin cancer, mostly due to UV rays, including UVA (320‐400 nm) and UVB (280‐320 nm).1, 2 Sunscreens play a crucial role in reducing the risk of melanoma and photo‐ageing by absorbing UV rays.3, 4 Sunscreens are labelled with sun protection factor (SPF) and protection grade of UVA (PA) levels according to their ability to protect against UVB and UVA, respectively. The commonly used SPF and PA simulators provide a final SPF or PA value based on complex calculations of each active ingredient's UV protection ability based on its concentration.5 Even though the concentration of UV filter substances is the principal influencer of the final SPF and PA values, various external factors are known to influence SPF and PA values.6-8 In order to increase the prediction rate of the SPF and PA prediction models based on machine learning, we included four important additional factors in the prediction model: presence of pigment, amount of pigment grade titanium dioxide (TiO2), type of formulation and type of product. In this study, we aimed to provide a rapid, inexpensive and efficient machine learning‐based SPF and PA prediction model, which will allow efficient development of new sun protection formulations.
1.1 Questions addressed
Can sunscreen prediction models based on machine learning provide cost‐effective, accurate and efficient results?
2 EXPERIMENTAL DESIGN
In this study, to increase the prediction rate of the SPF and PA prediction models based on machine learning, we included four important additional factors in the prediction model: presence of pigment, amount of pigment grade TiO2, type of formulation and type of product. Approximately 200 commercial and Ministry of Food and Drug Safety‐approved products, with values under SPF 70 (2377 individual clinical results) or PA 20 (2284 individual clinical results), were analysed by decision tree regression. Pearson's correlation coefficient was used for the correlation analyses employing IBM SPSS Statistics 25 (Armonk). SPF and PA testing was conducted in compliance with the regulatory requirements set forth by the International Sun Protection Factor Method (May 2006) or ISO 24444:2010 Test Method and ISO 24442 Test Method.9-11 Detailed information of the experimental procedures can be found within the Methods S1.
3 RESULTS
The combination of machine learning analysis and additional information of four factors increased both evaluation speed and prediction rate for the SPF and PA prediction models. In vivo SPF evaluations take at least 3 days for the UV radiation and MED evaluations, and in vivo PA evaluations take approximately 2‐24 hours for UVA radiation and minimum PPD evaluations.9-11 Even though in vitro SPF and PA are much faster than in vivo methods, they still require at least 15 minutes for the sunscreen sample to equilibrate at an ambient temperature to form a standard stabilized product film.12 The newly developed SPF and PA prediction model based on machine learning can provide prediction results for more than 1000 sunscreen formulations in less than a minute, which is much faster and more efficient than in vivo or in vitro SPF and PA evaluations.
The SPF prediction model showed good correlations with the SPF in vivo results, and the correlation improved when additional factors were included (Figure 1). The results for SPF as predicted by decision tree regression with active ingredient information showed very good correlation with SPF observed in vivo (r = 0.874; P < 0.01 for SPF in vivo and SPF predicted by machine learning) (Figure 1A). The coefficient of determination (R2), relative absolute error (RAE) and relative squared error (RSE) of active ingredient‐based correlation analysis were 0.76, 0.45 and 0.24, respectively (Table S2). The results of the SPF as predicted by decision tree regression with active ingredient information and additional factors showed a stronger correlation with SPF observed in vivo (r = 0.902; P < 0.01 for SPF in vivo and SPF predicted by machine learning) (Figure 1B). The R2, RAE and RSE of UV filter substances and additional factor‐based correlation analysis were 0.81, 0.41 and 0.20, respectively (Table S2). Thus, there was satisfactory correlation between the in vivo SPF values and decision tree regression predicted SPF values, and the correlation became stronger when SPF was predicted with the inclusion of additional factors.
image
Figure 1
Open in figure viewerPowerPoint
Correlations between sun protection factor (SPF) predicted by machine learning and SPF observed in vivo. A, Correlations between SPF predicted by machine learning based on concentration of ultraviolet (UV) filter substances with SPF observed in vivo. B, Correlations of SPF predicted by machine learning based on concentration of UV filter substances and four additional factors— presence of pigment, concentration of pigment grade titanium dioxide, type of formulation and type of product—with SPF observed in vivo.
The PA prediction model showed good correlations with the PA in vivo results, and the correlation became stronger when additional factors were included in the evaluation (Figure 2). The results of the PA as predicted by decision tree regression with active ingredient information showed a very good correlation with PA observed in vivo (r = 0.879; P < 0.01 for PA in vivo and PA predicted by machine learning) (Figure 2A). The R2, RAE and RSE of active ingredient‐based correlation analysis were 0.77, 0.45 and 0.23, respectively (Table S2). The results of the PA as predicted by decision tree regression with active ingredient information and additional factors showed a stronger correlation with PA observed in vivo (r = 0.894; P < 0.01 for PA in vivo and PA predicted by machine learning) (Figure 2B). The R2, RAE and RSE of UV filter substances and additional factor‐based correlation analysis were 0.80, 0.43 and 0.20, respectively (Table S2). Thus, there was satisfactory correlation between the in vivo PA values and decision tree regression predicted PA values, and the prediction rate became stronger with the inclusion of additional factors.
image
Figure 2
Open in figure viewerPowerPoint
Correlations between protection grade of ultraviolet A (PA) predicted by machine learning and PA observed in vivo. A, Correlations between PA predicted by machine learning based on concentration of ultraviolet (UV) filter substances with PA observed in vivo. B, Correlations of PA predicted by machine learning based on concentration of UV filter substances and four additional factors—presence of pigment, concentration of pigment grade titanium dioxide, type of formulation and type of product—with PA observed in vivo.
4 CONCLUSION
Machine learning analysis has not been used for the prediction of SPF and PA levels. With the available technology, we were able to develop complex SPF and PA prediction models based on machine learning. Our results suggest that the SPF and PA prediction models developed with decision tree regression analysis based on additional information—including concentration of UV filter substances, presence of pigment, amount of pigment grade TiO2, type of formulation and type of product—are valuable resources for the future development of sunscreens.
ACKNOWLEDGEMENTS
We thank Seonghwan Kong, Yeon Hee Kang and Jee Hee Jung for their assistance with the research.
CONFLICT OF INTEREST
The authors have declared no conflicting interests.
AUTHOR CONTRIBUTION
JS designed and performed the SPF and PA prediction model research and wrote the manuscript. JML designed and organized the research. SKP designed and analysed the research study.
Supporting Information
REFERENCES
Δεν υπάρχουν σχόλια:
Δημοσίευση σχολίου