Skin Phototype Classification with Machine Learning Based on Broadband Optical Measurements

Abstract

The Fitzpatrick Skin Phototype Classification (FSPC) scale is widely used to categorize skin types but has limitations such as the underrepresentation of darker skin phototypes, low classification resolution, and subjectivity. These limitations may contribute to dermatological care disparities in patients with darker skin phototypes, including the misdiagnosis of wound healing progression and escalated dermatological disease severity. This study introduces (1) an optical sensor measuring reflected light across 410-940 nm, (2) an unsupervised K-means algorithm for skin phototype classification using broadband optical data, and (3) methods to optimize classification across the Near-ultraviolet-A, Visible, and Near-infrared spectra. The differentiation capability of the algorithm was compared to human assessment based on FSPC in a diverse participant population (n = 30) spanning an even distribution of the full FSPC scale. The FSPC assessment distinguished between light and dark skin phototypes (e.g., FSPC I vs. VI) at 560, 585, and 645 nm but struggled with more similar phototypes (e.g., I vs. II). The K-means algorithm demonstrated stronger differentiation across a broader range of wavelengths, resulting in better classification resolution and supporting its use as a quantifiable and reproducible method for skin type classification. We also demonstrate the optimization of this method for specific bandwidths of interest and their associated clinical implications.

Keywords: Fitzpatrick skin type; K-means clustering; dermatology; machine learning; skin optical properties; skin type classification.

Figure A1

Groupwise (leftmost column) and pairwise (columns 2–8) comparison analysis results with p-value listed of human evaluation grouping results. (K–W: Kruskal–Wallis test). Red font denotes statistical significance.

Figure A2

Groupwise (leftmost column) and pairwise (columns 2–8) comparison analysis results with p-value listed of K-means_410–940 grouping results. (K–W: Kruskal–Wallis test). Red font denotes statistical significance.

Figure A3

Groupwise (leftmost column) and pairwise (columns 2–8) comparison analysis results with p-value listed of K-means_410–535 grouping results. (K–W: Kruskal–Wallis test). Red font denotes statistical significance.

Figure A4

Groupwise (leftmost column) and pairwise (columns 2–8) comparison analysis results with p-value listed of K-means_560–705 grouping results. (K–W: Kruskal–Wallis test). Red font denotes statistical significance.

Figure A5

Groupwise (leftmost column) and pairwise (columns 2–8) comparison analysis results with p-value listed of K-means_730–940 grouping results. (K–W: Kruskal–Wallis test). Red font denotes statistical significance.

Figure 1

(a) Fitzpatrick Skin Type Scale (I–VI) and (b) Generalized penetration depths of various wavelengths of light through tissue structures of interest [11].

Figure 2

Block diagram of experimental procedures.

Figure 3

Sensor outside of packaging showing electronics, LEDs, and photodiodes.

Figure 4

K-means classification workflow diagram.

Figure 5

Normalized intensity of (a) human evaluation skin classification method vs. (b) K-means_410–940 across a broad spectrum bandwidth; Significant main effects (α = 0.05) of the group on irradiance intensity are reported. NS: no statistical difference, *: p < 0.05, **: p < 0.01, ***: p < 0.001, ****: p < 0.0001. All group-level statistical values across different wavelengths can be found in Appendix A, Figure A1 and Figure A2.

Figure 6

Normalized intensity of optimized (a) K-means_410–535, (b) K-means_560–705, and (c) K-means_730–940 across a 410–940 nm bandwidth. Green shading denotes optimized bandwidths in the K-means classification approach, whereas grey shading denotes neglected bandwidths. Significant main effects (α = 0.05) of the group on irradiance intensity are reported. NS: no statistical differences, *: p < 0.05, **: p < 0.01, ***: p < 0.001, ****: p < 0.0001. All group-level statistical values across different wavelengths can be found in Appendix A, Figure A3, Figure A4 and Figure A5.

Figure 7

Table of statistical differences for intra-grouping pairwise comparison results at various wavelengths under different classification methods. (a) human FSPC classification method, (b) K-means_410–940, (c) K-means_410–535, (d) K-means_560–705, (e) K-means_730–940. Colors in (a) represent Fitzpatrick skin type scales I–VI. NS: no statistical difference, *: p < 0.05, **: p < 0.01, ***: p < 0.001, ****: p < 0.0001. All statistical tests excluded single-participant groupings. All intra-group level statistical values across different wavelengths can be found in Appendix A, Figure A1, Figure A2, Figure A3, Figure A4 and Figure A5.