Supplementary Material for "Artificial Intelligence for Diagnosis and Treatment Recommendation in Infantile Hemangioma: A Retrospective Multi-Center Diagnostic Accuracy Study"

Description:
This dataset contains supplementary materials supporting the findings of the multicenter diagnostic accuracy study titled "Artificial Intelligence for Diagnosis and Treatment Recommendation in Infantile Hemangioma: A Retrospective Multi-Center Diagnostic Accuracy Study".

Research Hypothesis:
The study hypothesized that the DeepIH system—an artificial intelligence model designed for end-to-end clinical decision support—would demonstrate diagnostic accuracy for infantile hemangiomas (IHs) comparable to that of specialized clinicians and provide treatment recommendations aligning with expert consensus in a multi-center real-world validation setting.

Data Content:
The supplementary material consists of two files:
1. Supplementary Table S1: Detailed demographic and clinical characteristics of the 400 included cases across the four participating institutions. Variables include patient age, lesion location, and source hospital.
2. Supplementary Figure S1: Concordance analysis between DeepIH's top-3 treatment recommendations and the treatment choices made by the expert panel for the 271 consensus-diagnosed IH cases. Data are presented as percentage agreement across the six treatment categories (topical timolol, oral propranolol, injection, follow-up, laser, and surgery).

Key Findings:
The primary findings from this validation study include:
1. DeepIH achieved an overall diagnostic accuracy of 85.9% (95% CI: 81.9-89.2%) against expert consensus, with high sensitivity (89.4%) and moderate specificity (72.8%).
2. For treatment recommendations, the AI's top-3 suggestions agreed with expert choices in 79.0% of IH cases.
3. Inter-expert agreement on treatment selection was only moderate (Fleiss' Kappa = 0.342), highlighting the inherent variability in clinical practice that the AI must accommodate.
4. Subgroup analyses (presented in Figure 1 of the main manuscript) demonstrated robust diagnostic performance during the early proliferative phase (0-6 months) and for cosmetically sensitive cephalofacial lesions, with treatment concordance rates of 59.8-92.8% across subgroups.
5. Multivariate analysis identified torso lesion location (OR=4.79) and unanimous expert consensus on IH diagnosis (OR=4.41) as independent predictors of correct AI diagnosis.

Data Interpretation and Usage:
These data support the conclusion that the DeepIH system can serve as a viable clinical decision-support tool, particularly in settings where specialist access is limited. The supplementary tables and figures provide granular detail on the case mix and the AI's performance across different clinical scenarios.