AI‐Based Objective Severity Assessment of Atopic Dermatitis Using Patient Photos in a Real‐World Setting: A Digital Biomarker Approach

Abbreviations

AD

atopic dermatitis

AI

artificial intelligence

EASI

Eczema Area and Severity Index

Itch‐NRS‐5

itch intensity on a 0–5 numeric rating scale

SCORAD

SCORing Atopic Dermatitis

TIS

Three Item Severity

To the Editor,

Atopic dermatitis (AD) is a chronic, relapsing skin condition requiring long‐term management [1]. Social media and smartphone‐based platforms facilitate symptom tracking and patient education while enabling continuous documentation of symptoms [2]. However, discrepancies between patient‐reported outcomes and objective disease severity measures highlight the need for standardized evaluation methods and the potential role of digital biomarkers [3, 4]. “Atopiyo,” Japan's largest online AD platform, allows 28,000+ users to share 57,000+ photos and comments about their symptoms, providing a rich dataset for real‐world disease monitoring (Figure 1A and Appendix S1) [5]. This study aimed to develop an artificial intelligence (AI)‐based model to assess eczema severity from user‐uploaded photos, offering a potential tool for objective assessment.

FIGURE 1.

(A) Atopiyo. Users can compare photos with previous ones, record a 0–5 Itch Numeric Rating Scale (Itch‐NRS‐5), and post comments. (B) The AI model incorporates three algorithms. (C) Algorithm for body part and lesion detection. A Single‐Shot Multibox Detector for object detection, employing a convolutional neural network. (D) Algorithm for determining the Three Item Severity (TIS) score using stacking ensemble learning. (E) Evaluation scales. TIS is a simple objective assessment with three components. SCORAD evaluates six intensity items, the rash area, and subjective symptoms; objective‐SCORAD focuses on the objective components. A higher score means worse symptoms in all instruments.

To develop the AI model, we trained and integrated three algorithms: body part detection, skin lesion detection, and severity assessment (Figure 1B). Single Shot MultiBox Detector and Convolutional Neural Network were utilized for body part and lesion detection (Figure 1C and Appendix S1). Since the model analyzes representative rather than full‐body images, we adopted the Three Item Severity (TIS) score, a localized severity assessment ranging from 0 to 9, evaluating erythema, edema/papulation, and excoriation [6]. TIS was calculated using a stacking ensemble technique (AI‐TIS: Figure 1D,E and Appendix S1). For validation, we adopted SCORing Atopic Dermatitis (SCORAD), highly correlated with EASI (Eczema Area and Severity Index), another key indicator of AD severity. SCORAD includes both objective measures like intensity parameters (oozing/crusts, lichenification, and dryness) and eczema extent, as well as subjective measures like pruritus and sleep loss (Figure 1E) [6]. Users recorded itch intensity on a 0–5 numeric rating scale (Itch‐NRS‐5) alongside their uploaded photos (Figure 1E). For those who visited physicians, TIS, SCORAD, and objective‐SCORAD were also clinically assessed (Figure 2A and Appendix S1). Pearson correlation analysis was performed to examine the relationship between Itch‐NRS‐5, TIS, SCORAD, and AI‐TIS using SPSS (version 26.0). The study was approved by Kyoto Prefectural University of Medicine (ERB‐C‐2611, ‐3047), and informed consent was obtained upon app download. Further methodological details are available in the Appendix S1.

FIGURE 2.

(A) Study workflow for developing the AI model and conducting correlation analysis. The AI model was trained with 880 images, following the test with 220 images. The established AI‐TIS examined the correlation with Itch‐NRS‐5, TIS, SCORAD, and objective‐SCORAD. (B) Participants' characteristics. (C) Validation of the developed AI model. The AI‐TIS accurately identified body parts at 98% and eczema lesions at 100%, strongly correlating with TIS assessments made by dermatologists or allergy specialists. (D–G) Correlation between AI‐TIS and other scores: (D) Itch‐NRS‐5, (E) TIS, (F) SCORAD, and (G) objective‐SCORAD. Objective scales (TIS and objective‐SCORAD) correlated well with AI‐TIS.

From the 57,429 images uploaded from August 2018 to January 2024, 9656 images (n = 900) with itch scores, excluding unclear images, were included to establish and validate AI‐TIS (Figure 2A). Most users were relatively young (median age: 33; range: 2–71) and predominantly female (68%, Figure 2B and Appendix S1). The trained AI model correctly detected 98% of body parts and 100% of eczema areas, and correlated well with TIS scores determined by board‐certified dermatologists and/or allergists (R = 0.73, p < 0.001) across 220 test images (Figure 2C). AI‐TIS showed a lower correlation with Itch‐NRS‐5 (R = 0.11, p < 0.001) across 8556 images (n = 602) (Figure 2D). Correlations of AI‐TIS with TIS, SCORAD, and objective‐SCORAD in 15 patients were R = 0.61 (p = 0.01), R = 0.4 (p = 0.12), and R = 0.53 (p = 0.04), respectively (Figure 2E–G and Appendix S1).

This study demonstrates that AI‐TIS can successfully identify body parts, eczema‐affected areas, and TIS scores from smartphone‐uploaded images in a non‐clinical setting. The strong correlation between AI‐TIS and objective measures, including TIS and objective‐SCORAD, supports its clinical utility for assessing objective outcomes and patient perception. In contrast, the weaker correlation between AI‐TIS and Itch‐NRS‐5 (R = 0.11) suggests that pruritus in AD does not always correspond to eczema severity [3]. Given the variability in subjective perception of pruritus, digital biomarkers such as smartphone applications may provide a more precise assessment method [4].

To enhance the model's versatility and standardization, future iterations should include patients of all ages and diverse skin types and incorporate elements from SCORAD or EASI. The AI model developed in this study has the potential to help patients with AD objectively assess their skin condition, facilitating timely and appropriate treatment. This study lays the groundwork for future advancements in AI‐driven dermatological assessments, enhancing both patient care and clinical research.

Author Contributions

U.O.K., M.H., and T.A. planned the strategy and were major contributors to writing the manuscript, supported by C.T., A.M., R.A., and S.N., supervised by S.Y., Y.O., M.F., T.In., T.Is., Y.I., K.M., S.S., N.K., and H.M. All authors reviewed the manuscript and approved the final manuscript.

Conflicts of Interest

R.A. is the representative partner of Atopiyo LLC and has received grants from Iwasa Educational and Cultural Foundation and Himawari Venture Development Foundation. M.H. has received grants from TAKEDA Science Foundation, Bayer Pharma Co. Ltd., Topcon Corp., and Tokai Optical Holdings. S.N. has served as a consultant and/or speaker (received honoraria) for Sanofi, Maruho, and Pfizer. S.N. is a member of an industry‐academia collaboration course with Maruho. T.In. reports non‐financial support from Lion Corporation and Sony Network Communications Inc.; grants from Yuimedi Inc., ROHTO Pharmaceutical Co. Ltd., Kobayashi Pharmaceutical Co. Ltd., and Kandenko Co. Ltd.; and personal fees from Santen Pharmaceutical Co. Ltd. and InnoJin Inc., outside the submitted work. N.K. has received honoraria as a speaker/consultant for Sanofi, Maruho, Abbvie, Ely‐Lilly Japan, Torii Pharmaceutical, and Otsuka Pharmaceutical, and has received grants as an investigator from Maruho, Sun Pharma, and Leo Pharma. T.A. has received grants from SECOM Science and Technology Foundation and Nakayama Foundation for Human Science. Other authors declare no conflicts of interest.

Supporting information

Data S1.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.