Fitzpatrick Skin Type (FST) Overview

• Fitzpatrick Skin Type (FST) is a six-category system classifying skin based on UV response, guiding clinical phototherapy dosing and risk assessment.
• Recent work integrates colorimetric measures and neural network models to enhance annotation accuracy and address inter-rater variability in FST determination.
• Studies reveal significant dataset imbalances and biases across FST bins, prompting the development of continuous, objective metrics for fairer AI applications.

The Fitzpatrick Skin Type (FST) scale is a six-point categorical system originally devised to stratify human skin according to its response to ultraviolet (UV) radiation, not as a direct measure of constitutive skin color. Though developed for clinical purposes such as phototherapy guidance and risk assessment for sun-induced dermatoses, FST has become the de facto standard for stratifying skin tone subgroups in medical AI, biometrics, and computer vision fairness research. The scale's growing role in model auditing, dataset stratification, and bias mitigation has prompted rigorous examination of its measurement properties, annotation protocols, biases, and limitations.

1. Conceptual Foundations and Dermatological Context

FST was introduced by Thomas B. Fitzpatrick in 1988 as an ordinal classification (Types I–VI) based on a patient’s burning and tanning history after sun exposure. The six classes are:

Type	Descriptor	UV Reactivity
I	Very fair, "always burns, never tans"	Highly sensitive, maximal burn risk
II	Fair, "usually burns, tans minimally"	Very sun sensitive
III	Medium, "sometimes mild burn, tans uniformly"	Moderately sun sensitive
IV	Olive, "rarely burns, tans easily"	Minimally sun sensitive
V	Brown, "very rarely burns, tans profusely"	Sun insensitive, easily tans
VI	Dark brown/black, "never burns, deeply pigmented"	Sun insensitive, deep pigmentation

Clinically, FST remains integral to phototherapy dosing, skin cancer risk stratification, and interpretation of morphologic dermatoses (where lesion appearance can vary markedly by FST). However, its application as a proxy for skin tone in computer vision and AI fairness greatly exceeds its original intent (Groh et al., 2021, Krishnapriya et al., 2021, Thong et al., 2023).

2. FST Measurement Protocols: Annotation, Inter-Rater Reliability, and Automation

Image-based FST assignment in modern datasets relies on several protocols:

• Expert and Crowdsourcing Protocols: Board-certified dermatologists or qualified annotators assign FST labels to images, referencing standardized charts and example photographs. Dynamic consensus mechanisms are used to optimize inter-rater consistency, with protocols specifying thresholds for consensus, qualification, and expert adjudication (Groh et al., 2021, Groh et al., 2022). Human annotation achieves high reliability: expert-expert and expert-crowd Pearson correlations are ρ ≈ 0.85–0.88, with diminishing returns beyond ~12 independent ratings per image (Groh et al., 2022).
• Manual Rating Variability: Even with color-calibrated imagery and exemplars, substantial inter-rater variation persists. Exact agreement for typical face datasets stands at 31–36%, but allowing for ±1 FST bin increases agreement to ~96% between any two raters (Krishnapriya et al., 2021). Color-correction (e.g., using an 18% gray calibration card) marginally improves agreement (Krishnapriya et al., 2021).
• Algorithmic Assignment (ITA-based Methods): The Individual Typology Angle (ITA), computed from the CIE L*b* color space, provides a continuous lightness–chromaticity measure:

$$\mathrm{ITA} = \arctan\Bigl(\frac{L^* - 50}{b^*}\Bigr)\times \frac{180}{\pi}$$

ITA thresholds are mapped to FST bins; e.g., ITA > 41.1 (I), 28.4 < ITA ≤ 41.1 (II), down to ITA ≤ –30.0 (VI) (Groh et al., 2022, Krishnapriya et al., 2021, Benčević et al., 6 Apr 2025). However, agreement of ITA-derived FST with expert/crowd-human annotation is weaker (ρ ≈ 0.52–0.57), with segmentation-based and clustering-based ITA approaches being more robust than patch-based methods (Groh et al., 2022, Groh et al., 2021, Benčević et al., 6 Apr 2025).

• Neural Network Estimators: Ordinal regression models (e.g., VGG-11+CORAL, EfficientNet-B0) trained on curated and/or synthetic data achieve balanced accuracy 0.72 and κ ≈ 0.53–0.95 in simulation, with performance approaching but not surpassing skilled human raters (Benčević et al., 6 Apr 2025, Benčević et al., 10 Feb 2026).

3. Distributional Skew, Dataset Bias, and Implications for Model Fairness

Large-scale medical datasets exhibit pronounced underrepresentation of darker skin types. In the Fitzpatrick 17k dataset, only 13% of images fall into FST 5–6; in ISIC 2020 and MILK10k dermatoscopic datasets, <1% of subjects are predicted as FST V–VI by colorimeter-supervised neural estimators (Groh et al., 2021, Benčević et al., 10 Feb 2026). Across face and dermatology datasets, the FST distribution typically skews toward Types II–IV (light/intermediate), with implications for downstream classifier performance and algorithmic fairness (Barros et al., 2023).

This class imbalance directly propagates to model calibration and accuracy: models trained primarily on light FST data exhibit marked decreases in accuracy (sometimes by >20 percentage points) on test sets from underrepresented FST bins, particularly when training/testing across non-overlapping FST groups (Groh et al., 2021, Sagers et al., 2022, Barros et al., 2023).

Fairness assessments now typically stratify error, sensitivity, calibration, and ROC-AUC by FST group, employing metrics such as:

• Equal Opportunity Difference (|TPR_light – TPR_dark|) (Pakzad et al., 2022)
• Normalized Accuracy Range (NAR) across FST groups (Pakzad et al., 2022)
• Fairness Gap (FG, Δ): max difference in a metric (AUC/BACC/ECE) across FST bins (Shah et al., 14 Sep 2025).

Imbalanced FST representation yields systematic disparities in algorithmic outcomes—models generalize poorly across the full skin tone spectrum unless explicitly designed to do so (Barros et al., 2023, Sagers et al., 2022, Pakzad et al., 2022).

4. Critical Evaluation of FST: Measurement Properties, Biases, and Limitations

Multiple studies expose structural and measurement limitations of the FST scale:

• Non-colorimetric Basis and Subjectivity: FST remains a text-only UV-reactivity questionnaire, with no well-accepted, direct mapping to colorimeter or CIE L* values (Cook et al., 2024, Howard et al., 2021). In image datasets, its use as a skin color label is an appropriation, not a direct proxy for melanin content or reflectivity.
• Low Sensitivity/Resolution: Each FST step spans ≈14.7 L* units, compared to finer-grained colorimetric scales such as CST (ΔL* ≈ 4.9) and the Monk Skin Tone Scale (ΔL* ≈ 7.4). Only ~66% of the 1–6 ordinal range is utilized across real samples (Cook et al., 2024).
• Demographic Bias: After controlling for measured skin lightness, racial/ethnic group identity remains a significant predictor of self-reported FST (e.g., White raters select ≈0.32 bins lower than Black raters with the same L*) (Cook et al., 2024). FST–L* correlation within race is weak (Kendall’s τ = 0.23); most variance is explained by race rather than phototype (Howard et al., 2021).
• Original Design and Categorization Bias: The original Fitzpatrick system (1988) defined only I–IV for lighter skin; V–VI extensions are coarser, introducing granularity bias and a categorical skew toward light/intermediate bins, further reinforced in AI-driven grouping practices (Shah et al., 14 Sep 2025, Thong et al., 2023).
• Annotation Variability: Inter-rater and expert–algorithm agreement are moderate (expert–expert ρ ≈ 0.85, expert–ITA ρ ≈ 0.55), and image-based FST assignment is confounded by lighting, camera processing, and anatomical variation, particularly without calibration objects (Krishnapriya et al., 2021, Groh et al., 2022).

5. Alternatives and Enhanced Measurement Strategies

The acknowledged limitations have catalyzed proposals for continuous and multidimensional measurement:

• Colorimetric Skin Tone (CST) Scale: A ten-swatch palette derived from colorimeter measurements (L* = 70–20, ΔL* = 5) achieves double the sensitivity of MST and triple that of FST, and removes subjectivity by anchoring to calibrated color values. CST reduces racial bias in assignments and achieves consistency (ICC ≈ 0.90) (Cook et al., 2024).
• Monk Skin Tone Scale: A palette of 10–12 calibrated shades designed to represent the diversity of human skin reflectance, with finer granularity and improved sensitivity for both light and dark tones (Shah et al., 14 Sep 2025).
• Continuous ITA / CIE L*a*b* Measures: Many groups now advocate extracting continuous L* (tone) and h* (hue angle) directly from segmented skin pixels, reporting model accuracy across the joint (L*, h*) plane to surface intersectional biases. Thresholds may be set empirically for binning but are traceable to objective reflectometric values (Thong et al., 2023, Benčević et al., 6 Apr 2025, Benčević et al., 10 Feb 2026).
• Neural Networks Supervised by Colorimeter/Clinical Gold Standards: Recent neural estimators regress FST and ITA using in-person gold-standard Fitzpatrick/CIELAB labels from colorimeter devices. Properly regularized models match or exceed traditional human agreement rates and enable reproducible, scalable annotation of new datasets for fairness auditing (Benčević et al., 10 Feb 2026, Benčević et al., 6 Apr 2025).

6. Impact on Model Development, Fairness Auditing, and Future Directions

The continued use and critique of FST profoundly shapes practices in dataset curation, dermatology AI development, and clinical translation:

• Equitable Model Training: Data resampling, loss reweighting, synthetic image generation (via diffusion models such as DALL·E 2), and subgroup-specific regularization loss are now deployed to mitigate the underrepresentation and bias documented by FST analyses (Sagers et al., 2022, Pakzad et al., 2022, Groh et al., 2021).
• Performance Auditing: Rigorous stratification of sensitivity, specificity, confidence, AUC, and calibration error by FST or finer-grained alternatives is now standard. Subgroup-specific audits reveal systematic degradation outside of well-represented FST bins, with sometimes dramatic drops in F1/Balanced Accuracy for minority types (Barros et al., 2023, Sagers et al., 2022, Shah et al., 14 Sep 2025).
• Best Practices and Recommendations:
  • Collect and annotate datasets to deliberately balance FST representation, or preferably use continuous colorimetric annotation (Groh et al., 2021, Barros et al., 2023, Benčević et al., 10 Feb 2026).
  • Replace or supplement FST with CST, MST, or direct ITA/L* metrics; report subgroup metrics using continuous, multidimensional skin color embeddings (Cook et al., 2024, Thong et al., 2023, Shah et al., 14 Sep 2025).
  • Implement qualified-crowdsourcing protocols with explicit dynamic consensus and expert adjudication for large-scale labeling; reserve ITA-based automation for pre-screening or for datasets with validated colorimeter groundtruth (Groh et al., 2022).

Recent research converges on the need for objective, high-sensitivity, and demographically robust skin tone measurement protocols to enable reproducible fairness audits and to support the clinical reliability of AI technologies across the true spectrum of human skin phenotypes. The FST, while historically foundational, is increasingly supplemented—or supplanted—by rigorous colorimetric and continuous measurement scales to better align technology with population diversity and clinical consequences (Cook et al., 2024, Benčević et al., 10 Feb 2026, Shah et al., 14 Sep 2025, Thong et al., 2023).