Fitzpatrick Skin Tone (FST) Scale Overview

• The Fitzpatrick Skin Tone Scale is a six-level classification originally designed to assess UV sensitivity, now used as a proxy for visible skin tone.
• It is operationalized via self-report, expert assessment, and digital annotation, serving multiple roles in dermatology, biometrics, and AI fairness studies.
• Recent research critiques its limited granularity and calls for supplementation with continuous measurements like ITA or alternative scales such as Monk Skin Tone.

Fitzpatrick Skin Tone, more precisely Fitzpatrick Skin Type (FST), is a six-point dermatologic classification that was originally created to classify skin response to ultraviolet exposure and later became a common proxy for visible skin tone in dermatology datasets, computer vision, biometrics, wearable sensing, and fairness auditing. Across these literatures, the scale is used both in its original medical framing and as an operational grouping variable for appearance-based analyses, yet multiple recent studies emphasize that it was not designed as a direct colorimetric taxonomy of apparent skin color and does not fully capture the diversity of human skin appearance (Howard et al., 2021, Thong et al., 2023, Matias et al., 2 Mar 2026).

1. Dermatologic origin and category structure

The Fitzpatrick system is conventionally organized as six ordered types, I through VI. In the medical framing summarized in recent biometric and annotation studies, it is a classification of UV sensitivity and tanning or burning tendency rather than a direct measurement of pigmentation. One paper states that the original system had four types and was later extended to six, while another notes that the scale is “originally used to assess the suntanning pathways of skin” and is now routinely reused as a skin-color classification tool in cosmetics, dermatology, and computer vision (Howard et al., 2021, Thong et al., 2023).

Type	Description
I	Highly sensitive, always burns, never tans
II	Very sun sensitive, burns easily, tans minimally
III	Sun sensitive to skin, sometime burns, slowly tans to light brown
IV	Minimally sun sensitive, burns minimally, always tans to moderate brown
V	Sun insensitive skin, rarely burns, tans well
VI	Sun insensitive, never burns, deeply pigmented

These six categories are repeatedly treated as an ordered progression from lighter to darker skin. Some image-generation work restates the range visually as “Type I (Pale White)” through “Type VI (Dark Brown/Black),” but the broader literature represented here consistently notes that the original construct is photobiologic rather than purely visual (Balestri, 12 Feb 2026).

2. Operationalization in datasets, annotation, and modeling

In contemporary research, FST is used in several distinct ways. It appears as self-report via the standard dermatologic survey, as in-person expert assessment, as dermatologist or crowd annotation from images, as grouped metadata in public datasets, and as a model output derived from ordinal regression or palette matching. This diversity of use is central to the present literature: the same label name can denote sun-response self-report, image-based appearance judgment, or an automatically inferred subgroup variable (Howard et al., 2021, Groh et al., 2021, Benčević et al., 10 Feb 2026, Balestri, 12 Feb 2026).

Large public datasets made FST especially important for subgroup analysis. Fitzpatrick17k contains 16,577 clinical images annotated with Fitzpatrick labels and 114 skin conditions, and its labels are used both at the full six-level resolution and in grouped forms such as lighter skin types $1\text{--}2$, medium skin types $3\text{--}4$, and darker skin types $5\text{--}6$ (Groh et al., 2021). DDI, by contrast, already provides grouped categories, and work on dermatology classification and fairness commonly harmonizes labels into FST 1/2, 3/4, and 5/6 or the equivalent T12, T34, and T56 notation (Shah et al., 14 Sep 2025, Aayushman et al., 2024). Other studies retain all six categories and treat FST as the central subgroup variable for threshold selection, domain generalization, or fairness evaluation, rather than collapsing them into a binary light-versus-dark split (Stothers et al., 8 Oct 2025, Pakzad et al., 2022).

The modeling treatment also varies by task. Some analyses enter FST as a categorical covariate, as in mixed-effects models for smartwatch physical activity energy expenditure error, where device, body fat percentage, Fitzpatrick skin type, and all pairwise interactions were included as fixed effects (Kostrna et al., 29 Jan 2026). Other work treats FST explicitly as an ordinal variable and therefore analyzes it with medians, distributions, contingency-based statistics, or ordinal-regression heads such as CORAL rather than ordinary nominal classification or continuous averages (Balestri, 12 Feb 2026, Benčević et al., 10 Feb 2026).

3. Relationship to objective colorimetry and continuous skin measurements

A major line of recent work argues that FST should be distinguished sharply from objective skin-color measurement. The most direct formulation appears in the biometrics study on “Reliability and Validity of Image-Based and Self-Reported Skin Phenotype Metrics,” which states that FST is a dermatology classification system originally created to classify UV sensitivity, not to quantify visible skin tone, and concludes that FST is poorly predictive of skin-tone when compared with calibrated device measurements (Howard et al., 2021). In that study, the correlation between FST and ground-truth Face Area Lightness Measures was $\tau = 0.51$, whereas the correlation between race and ground-truth FALMs was $\tau = 0.68$; within each race category, the correlation between FST and ground-truth FALMs dropped to $\tau = 0.23$. Linear models likewise gave $R^2 = 0.48$ for $L_f \sim FST$, $R^2 = 0.72$ for $L_f \sim race$, and $3\text{--}4$0 for $3\text{--}4$1 (Howard et al., 2021).

Several papers compare FST with continuous colorimetric representations, especially the Individual Typology Angle (ITA), computed in CIELAB space as

$3\text{--}4$2

In Fitzpatrick17k, agreement between ITA-derived labels and human Fitzpatrick labels was only moderate even under a $3\text{--}4$3 criterion: overall concordance reached 45.87% or 60.34% for full-image approaches and 53.30% or 70.38% with YCbCr masking, depending on the thresholding scheme (Groh et al., 2021). In dermatoscopy, a colorimeter-supervised estimator treated Fitzpatrick as an ordinal target and ITA as a continuous regression target derived from predicted $3\text{--}4$4, achieving Fitzpatrick weighted kappa 52.98% on MSKCC dermatoscopic images and ITA ICC3 approximately 94% against average triplicate colorimeter ITA, while also showing that fewer than 1% of subjects in ISIC 2020 and MILK10k were predicted to belong to Fitzpatrick types V and VI (Benčević et al., 10 Feb 2026).

Manual-rating studies reach similar conclusions about the scale’s limited colorimetric precision. In a large self-rating experiment with calibrated colormeter measurements, FST produced adjusted $3\text{--}4$5 against measured skin tone, compared with 0.5709 for MST and 0.6093 for CST, and one FST rating step corresponded to 14.7 $3\text{--}4$6 units, making it the least sensitive of the evaluated scales to actual lightness variation (Cook et al., 2024). This pattern supports a recurring conclusion in the literature: FST is clinically familiar and operationally useful, but it is not equivalent to a direct measurement of apparent skin color.

4. FST as a subgrouping variable in machine learning and fairness research

The scale became central to AI fairness work because it allows performance to be stratified by skin appearance when race or ethnicity labels are unavailable, heterogeneous, or legally sensitive. In dermatology classification, the Fitzpatrick 17k study showed that models trained on restricted FST ranges perform best on test images with Fitzpatrick types closest to those seen in training, making subgroup generalization failure directly visible (Groh et al., 2021). CIRCLe formalized this use further by treating FST as the protected attribute $3\text{--}4$7, using a StarGAN-based image-to-image translation model $3\text{--}4$8 to generate counterfactual versions of the same lesion under different skin types and regularizing latent representations with

$3\text{--}4$9

thereby improving both overall accuracy and multi-group fairness on Fitzpatrick17K (Pakzad et al., 2022).

FST is also used pragmatically for post-processing and subgroup-aware model design. “Fitzpatrick Thresholding for Skin Image Segmentation” trains psoriasis segmentation models without tone input and then chooses different probability thresholds for different Fitzpatrick groups at inference time; for Fitz VI, this lifted held-out test performance by up to +31.46% bIoU and +24.13% Dice on U-Net, with smaller but similarly directed gains on ResU-Net and SETR-small (Stothers et al., 8 Oct 2025). In lesion classification, training subgroup-specific models on FST 1/2, FST 3/4, and FST 5/6 reduced fairness gaps relative to a balanced general model, while further coarsening from separate FST 1/2 and FST 3/4 labels into a single FST 1/2/3/4 group degraded performance, especially for FST 1/2 (Shah et al., 14 Sep 2025). PatchAlign likewise uses Fitzpatrick labels as the sensitive attribute on Fitzpatrick17k and DDI and reports improved average accuracy together with improved fairness metrics such as DPM and EOM, including a large gain for T6 on Fitzpatrick17k (Aayushman et al., 2024).

This body of work establishes FST as a practical infrastructure for auditing subgroup disparities, selecting operating points, and defining sensitive groups. At the same time, the same papers repeatedly caution that the scale is a proxy rather than a complete description of skin appearance, and that dataset imbalance across its categories can itself distort fairness conclusions.

5. Empirical limits of FST as an explanatory variable

A notable feature of recent work is that FST often remains biologically or operationally plausible while failing to explain task performance as strongly as other factors. In smartwatch energy-expenditure estimation, the hypothesized mechanism was optical: “Melanin in darker skin absorbs more of the green light used by common sensors, attenuating the pulsatile signal and contributing to higher error rates in HR detection.” Yet in the actual mixed-effects models, “Fitzpatrick skin type showed no robust main effect (p = .89),” and the discussion concluded that there was “no measurable effect of Fitzpatrick skin tone (III-V) on PAEE error,” whereas device brand, body fat percentage, and their interaction were significant (Kostrna et al., 29 Jan 2026). Because that study only included Hispanic adults with Fitzpatrick III–V, with 51.7% type III, 41.4% type IV, and 6.9% type V, the authors explicitly warned that the null result should be interpreted with caution (Kostrna et al., 29 Jan 2026).

A related pattern appears in dermoscopic lesion segmentation. “Exploring the Impact of Skin Color on Skin Lesion Segmentation” used Fitzpatrick grouping as a baseline stratification variable but found that global tone measures, including Fitzpatrick grouping and mean ITA, had weak association with segmentation performance, with $5\text{--}6$0 for most metrics. By contrast, lesion–skin pigment contrast derived from pixel-wise ITA distributions and Wasserstein distances was much more predictive of segmentation success or failure; low lesion-skin contrast was consistently associated with lower segmentation accuracy and higher error rates (Paxton et al., 31 Mar 2026). The paper therefore argues that fairness auditing for dermoscopic segmentation should move beyond discrete Fitzpatrick bins toward contrast-aware, distribution-based analysis.

These findings do not show that FST is irrelevant. Rather, they indicate that in some sensor-mediated or pixel-level tasks, the scale may be too coarse, too underpowered, or too conceptually distant from the actual mechanism of failure. A plausible implication is that FST often functions best as a first-pass stratification variable rather than as a sufficient physical description of skin appearance.

6. Critiques, alternatives, and current directions

The main critiques of FST in recent literature are conceptually consistent across domains. First, the scale was originally developed for sun response in predominantly white populations and is therefore a repurposed rather than native skin-color taxonomy (Balestri, 12 Feb 2026). Second, it has restricted granularity for darker tones and is often described as less inclusive than newer scales such as Monk Skin Tone (MST) (Matias et al., 2 Mar 2026). Third, it collapses apparent skin color into a light-to-dark axis and inadequately captures hue variation, even though image-based fairness failures can depend on red–yellow differences as well as lightness (Thong et al., 2023). Fourth, image-based FST annotation is often subjective, inconsistent across raters, and sensitive to illumination, capture conditions, or color normalization choices (Krishnapriya et al., 2021, Cook et al., 2024).

Recent alternatives follow three main directions. One direction replaces FST with visually grounded scales such as the 10-tone Monk Skin Tone scale; the STW benchmark argues that MST provides finer granularity, especially among darker tones, and is more suitable for in-the-wild computer vision auditing than the traditional 6-tone categorical scales (Matias et al., 2 Mar 2026). A second direction keeps FST for comparability but supplements it with objective or continuous measures such as ITA, colorimeter-derived $5\text{--}6$1, or neural estimators trained against physical measurements (Benčević et al., 10 Feb 2026). A third direction moves beyond one-dimensional tone altogether: “Beyond Skin Tone” recommends multidimensional skin color scales relying on both skin tone and hue, while the CST work proposes a new Colorimetric Skin Tone scale and reports that it is more sensitive, consistent, and colorimetrically accurate than FST or MST, though still vulnerable to perceptual and race-related bias in manual annotation (Thong et al., 2023, Cook et al., 2024).

Current research therefore treats FST in a dual manner. It remains a common language for subgroup reporting, threshold selection, and historical comparison across dermatology AI, biometrics, and fairness studies. Simultaneously, the most technically detailed recent work increasingly treats it as a coarse, subjective, and medically repurposed proxy that should be paired with broader and more balanced subgroup representation, better objective optical measurements, continuous colorimetric variables, or multidimensional representations of apparent skin color (Kostrna et al., 29 Jan 2026, Paxton et al., 31 Mar 2026, Benčević et al., 10 Feb 2026).