LH2-data-labs/longitudinal-dermatology-atlas

LH2 Data — Longitudinal Dermatology Atlas

A longitudinal multimodal dermatology dataset with real-world per-patient follow-up across the full medical dermatology spectrum — not isolated snapshots, not cosmetic procedures. South Asian skin types at scale.

---

Dataset Summary

Most dermatology AI datasets are cross-sectional snapshots predominantly representing Fitzpatrick I–III skin types. This dataset inverts both limitations: it provides longitudinal clinical records with multi-year per-patient follow-up (5+ years for mature cohorts) and is anchored entirely in South Asian populations (Fitzpatrick IV–VI), addressing the well-documented skin-type bias in dermatology AI.

This sample contains structured visit-level metadata for 10 de-identified patients across 197 clinical visits, demonstrating the longitudinal schema. The complete dataset includes 60,000+ longitudinal records with active enrolment ongoing.

This sample is metadata-only. Clinical photography (300+ high-resolution images per patient across anatomical regions per visit) is available under gated access with a signed Data Use Agreement.

Cohort at a Glance

Dimension	Scale
Longitudinal Records	60,000+ (sample: 197 visits across 10 patients)
Photos per Patient	300+ across anatomical regions per visit
Longitudinal Depth	5+ years follow-up for mature cohorts
Fitzpatrick Coverage	IV, V, VI (South Asian skin types)
Disease Focus	Chronic inflammatory, autoimmune, pigmentary, acne spectrum

Disease Coverage

Category	Conditions
Chronic Inflammatory	Psoriasis, Eczema / Atopic Dermatitis, Lichen Planus
Acne Spectrum	Hidradenitis Suppurativa, Severe Acne Vulgaris
Autoimmune / Immunobullous	Pemphigus Vulgaris, Bullous Pemphigoid, Cutaneous Lupus
Pigmentary Disorders	Vitiligo
Hair Disorders	Alopecia Areata

Data Modalities

Modality	Description	Availability
Clinical Photography	300+ DSLR high-resolution images per patient across anatomical regions per visit. Longitudinal series showing disease progression and treatment response.	Gated (DUA required)
Prescriptions & Treatment Regimens	Full pharmacotherapy stack — topicals, immunosuppressants, biologics, isotretinoin. Dosing, duration, and switches documented per visit.	In this sample
Serial Lab Monitoring	CBC, LFT, renal, lipid, autoimmune, and biologic-screening panels per visit.	In this sample
Visit-Level Metadata	Date, severity assessment, PASI score (psoriasis), anatomical regions, photo counts, treatment response.	In this sample
Multimodal Linked	Photography, prescriptions, and labs — joined per patient and visit.	Full dataset

Visit Record Schema (16 fields)

Field	Type	Description
patient_id	string	De-identified patient identifier
visit_number	int	Sequential visit number (1, 2, 3...)
visit_date	string	Date of clinical visit (YYYY-MM-DD)
diagnosis	string	Primary dermatological diagnosis
disease_category	string	Condition category (chronic_inflammatory / autoimmune / pigmentary / acne_spectrum / hair_disorder)
fitzpatrick_type	string	Fitzpatrick skin type (IV / V / VI)
age_at_first_visit	int	Patient age at first recorded visit
sex	string	Male / Female
severity	string	Clinical severity assessment (Mild / Moderate / Severe / Very Severe)
pasi_score	float	Psoriasis Area and Severity Index (psoriasis patients only; null otherwise)
photos_taken	int	Number of clinical photographs taken at this visit
anatomical_regions	string	Pipe-separated list of anatomical regions photographed
current_medication	string	Current pharmacotherapy regimen
lab_panels	string	Pipe-separated list of lab panels ordered at this visit
follow_up_months	int	Months since first visit
treatment_response	string	Clinical response assessment (Improving / Stable / Worsening / Flare / Remission)

Why This Dataset Matters

Existing dermatology datasets on HuggingFace and elsewhere share three structural limitations that this dataset addresses directly:

Cross-sectional vs. Longitudinal. HAM10000, Fitzpatrick17k, DDI, and Derm1M are all single-visit snapshot datasets. They cannot capture disease trajectory, treatment response over time, medication switches, or the visual progression of chronic conditions. This dataset provides 10–30 visits per patient spanning 2–5+ years.

Fitzpatrick I–III bias. The majority of training data for skin AI models comes from populations with lighter skin types. Per the 2025 SkinMap audit, Fitzpatrick V–VI representation in the largest public datasets is approximately 5.8%. This dataset is 100% Fitzpatrick IV–VI.

Cosmetic vs. Medical. Many skin datasets mix cosmetic dermatology (mole classification, benign lesion detection) with medical dermatology. This dataset covers exclusively medically-managed conditions requiring systemic therapy — psoriasis, pemphigus, HS, cutaneous lupus — where longitudinal treatment data is critical.

Key Differentiators

• Longitudinal: 5+ years follow-up with 300+ photos per patient — not single-visit snapshots
• South Asian skin: 100% Fitzpatrick IV–VI — directly addresses the most documented bias in dermatology AI
• Medical focus: Chronic inflammatory, autoimmune, and immunobullous conditions requiring systemic therapy
• Multimodal linked: Photography + prescriptions + serial lab panels joined per patient and visit
• Full pharmacotherapy stack: Topicals, immunosuppressants (methotrexate, cyclosporine, MMF), biologics (secukinumab, adalimumab), isotretinoin — real treatment trajectories, not just diagnostic labels

IP & Compliance

• Full commercial and AI/ML training rights under exclusive licensing agreement
• De-identified patient data — no personally identifiable information
• Clinical photography available under gated access with signed Data Use Agreement
• Re-identification of patients from clinical photographs is expressly prohibited
• Compliant with India's DPDP Act 2023 for data licensing and cross-border transfer

Citation

@dataset{lh2_derma_2025,
  title={LH2 Data — Longitudinal Dermatology Atlas},
  author={LH2 Data Labs},
  year={2025},
  publisher={Hugging Face},
  note={60K+ longitudinal dermatology records with South Asian Fitzpatrick IV-VI coverage}
}