INTRODUCTION
Early diagnosis of skin disease, especially malignant neoplasms, can significantly reduce patient morbidity and mortality. Assessment of most skin conditions is often first performed by a nurse and/or general practitioner before a subsequent referral to a dermatologist is made. Previous publications by numerous groups report the ability of custom-built Artificial Intelligence (AI) systems to help assess skin conditions.1-6
In 2021, Jain et al examined the use of Google’s AI-based algorithm by primary care physicians and claimed it could improve the triage of skin conditions.7 The study however severely underrepresented patients with more highly pigmented skin. Specifically, the AI was validated on only 15 out of 152 (9.9% of total) histologically verified cases of skin conditions in Fitzpatrick skin types IV and V and no histologically verified cases of the darkest Fitzpatrick skin type (VI). Due to the class imbalance present, the test set was heavily skewed towards the 3 lightest skin types with 90.1% of the data representing 50% of the Fitzpatrick classes (I-III), while only 9.9% of the data represented the other 50% of the Fitzpatrick classes (IV-VI). These investigators did not generate sufficient evidence to support the study’s conclusion.
Recent studies have also highlighted the decreased diagnostic accuracy among clinicians and clinician-trainees on photos of patients with skin of color (SOC).8-10 Specifically, Diao et al reported that the diagnostic accuracy of identifying cutaneous/subcutaneous pathology in darker skin types was significantly lower at 44.3% compared with 50.5% in intermediate skin and 50.4% in light skin.8 This is likely due to poor representation of SOC in textbooks and instructional guides, comprising as low as 4% to 18% of the total number of photos in these teaching materials.11 Furthermore, previous studies have reported that people with SOC have worse prognoses and lower survival rates compared to individuals with light skin. These poorer patient outcomes are due to delayed or incorrect diagnoses.12-16
In 2021, Jain et al examined the use of Google’s AI-based algorithm by primary care physicians and claimed it could improve the triage of skin conditions.7 The study however severely underrepresented patients with more highly pigmented skin. Specifically, the AI was validated on only 15 out of 152 (9.9% of total) histologically verified cases of skin conditions in Fitzpatrick skin types IV and V and no histologically verified cases of the darkest Fitzpatrick skin type (VI). Due to the class imbalance present, the test set was heavily skewed towards the 3 lightest skin types with 90.1% of the data representing 50% of the Fitzpatrick classes (I-III), while only 9.9% of the data represented the other 50% of the Fitzpatrick classes (IV-VI). These investigators did not generate sufficient evidence to support the study’s conclusion.
Recent studies have also highlighted the decreased diagnostic accuracy among clinicians and clinician-trainees on photos of patients with skin of color (SOC).8-10 Specifically, Diao et al reported that the diagnostic accuracy of identifying cutaneous/subcutaneous pathology in darker skin types was significantly lower at 44.3% compared with 50.5% in intermediate skin and 50.4% in light skin.8 This is likely due to poor representation of SOC in textbooks and instructional guides, comprising as low as 4% to 18% of the total number of photos in these teaching materials.11 Furthermore, previous studies have reported that people with SOC have worse prognoses and lower survival rates compared to individuals with light skin. These poorer patient outcomes are due to delayed or incorrect diagnoses.12-16
