Table 2.

Evidence classification for framework recommendations. Summary of bias mitigation practices included in the proposed framework, organized by strength of evidence. Established evidence base indicates practices supported by studies conducted in aesthetic facial evaluation contexts. Adapted evidence base indicates practices supported by evidence from adjacent domains (facial recognition, general medical AI, or computer vision) that have been translated to aesthetic evaluation by analogy and require validation in aesthetic-specific contexts before adoption as standard practice. Proposed evidence base indicates conceptual recommendations without empirical validation in any closely related domain, to be treated as research directions.

Bias source	Pipeline stage	Mitigation strategy	Evidence base	Residual gap
Targeting bias: narrow beauty standard definition	Data collection	Pillar 1: ≥7 ethnic categories Mixed-ancestry probabilistic labeling Continuous morphometric representation	Established	Discrete ethnic categories essentialize group-level patterns Intragroup variation (eg, nationality, socioeconomic background) is rarely captured
Annotation bias: culturally skewed rater judgments	Data collection	Pillar 1: diverse rater recruitment Structured training and calibration Tiered disagreement adjudication Ongoing score audits	Adapted	No validated rater calibration protocol exists for aesthetic evaluation Cultural feature weighting may persist despite diverse panels
Generative adversarial network amplification bias: synthetic augmentation	Data collection	Pillar 1: quality control gate for synthetic images—fairness audit, feature distribution check, and human review	Adapted	Mode collapse and feature exaggeration are documented in generative systems Quality control criteria not validated for aesthetic contexts
Modeling bias: fairness-unaware training	Model training	Pillar 2: adversarial debiasing Centroid fairness loss Skewness-aware reinforcement learning Debiasing variational autoencoder Multitask learning	Adapted	All techniques validated in facial recognition or general computer vision, not aesthetic evaluation Several rest on preprint evidence Combined validation absent
Domain shift bias: train/deploy distribution mismatch	Model training → deployment	Pillar 2: hybrid pretraining on standardized images Fine-tuning on clinical images Domain generalization evaluation prerelease	Proposed	No validated hybrid protocol for aesthetic AI Clinical image variation not systematically characterized
Evaluation bias: aggregate metrics obscure subgroup disparities	Evaluation	Pillar 3: layered fairness metrics with prioritization hierarchy Intersectional assessment Bayesian hierarchical modeling for rare subgroups	Adapted	Thresholds are proposed benchmarks without empirical derivation Metrics can conflict Intersectional sample sizes often insufficient
Explainability gap: black-box outputs in a cultural context	Evaluation → deployment	Pillar 4: Grad-CAMa, LIMEb, SHAPc with required human expert review Geometric/physics-based models as a longer-term goal	Adapted	Explainable AI tools cannot explain why features are culturally valued No method validated for cultural appropriateness verification in aesthetic AI
Human–AI decision bias: clinician interpretation and override	Deployment	Pillar 5 + 6: documentation and audit of AI recommendation override rates by patient demographic Clinician training on implicit bias	Proposed	No empirical data on differential override in aesthetic AI Audit infrastructure absent Accountability for remediation undefined
Deployment bias: commercial systems without governance	Deployment	Pillar 6: disclosure-based accountability for commercial developers FDA SaMDd framework alignment	Proposed	No enforcement mechanism for commercial tools Patient and clinician verification of compliance is currently impossible
Drift bias: postdeployment fairness degradation	Monitoring	Pillar 6: tiered monitoring—continuous process control, quarterly review, annual audit, drift-triggered escalation Designated AI clinical lead	Adapted	Drift thresholds not empirically derived for aesthetic AI Continuous monitoring may not be feasible for community practices

^aGrad-CAM: gradient-weighted class activation mapping.

^bLIME: local interpretable model-agnostic explanations.

^cSHAP: Shapley additive explanations.

^dFDA SaMD: Food and Drug Administration’s Software as a Medical Device.