LH2-data-labs/multimodal-oncology-atlas

LH2 Data — Multimodal Oncology Dataset

A large-scale, multimodal oncology dataset built around a principle rare in the field: placing non-Caucasian patient populations at the centre, not the periphery.

---

Dataset Summary

The vast majority of oncology datasets used to train diagnostic, prognostic, and treatment AI models are drawn overwhelmingly from Caucasian, Western populations — a well-documented limitation that undermines model generalisability and equity in real-world deployment. This dataset addresses this directly, offering a deeply multimodal oncology resource — genomic, clinical, imaging, and biospecimen — anchored across the Global South.

This sample contains structured clinical records for 20 de-identified cancer patients demonstrating the full schema. The complete dataset includes ~25,000 enrolled patients (target: 50,000 within 12 months) across 15 countries spanning the Indian Subcontinent, Southeast Asia, Latin America, and the Middle East.

Cohort at a Glance

Dimension	Scale
Patients Enrolled	~25,000 (sample: 20)
12-Month Target	50,000 patients sequenced
Geographic Coverage	Indian Subcontinent, Southeast Asia, Latin America, Middle East
Cancer Types	Multiple solid tumour types (prostate, colon, lung, and others)
Clinical Data Categories	13 per patient
Data Modalities	Genomic (WES) + Digital Pathology (WSI) + Clinical + Biobank

Data Modalities

Modality	Description	Format
Genomic Sequencing	Whole exome sequencing (WES) and targeted panel data. Variant calls available.	FASTQ, BAM, VCF
Digital Pathology	Whole slide images (WSIs) linked to genomic and clinical data across multiple cancer types.	SVS, TIFF, DICOM
Clinical Records	Structured longitudinal data across 13 categories: staging (AJCC pTNM), surgery, chemotherapy, targeted therapy, immunotherapy, radiation, PET-CT, molecular pathology (NGS, IHC), serum biomarkers, haematology, biochemistry, and performance status.	Parquet, CSV, JSON
Biobank	Residual FFPE tissue and extracted DNA/RNA for re-analysis and future modality generation.	Physical specimens
Multimodal Linked	Every modality joined at the patient level — DNA, tissue imaging, and clinical trajectories as linked records.	Patient-level joins

Clinical Record Schema (28 fields)

Field	Type	Description
patient_id	string	De-identified patient identifier
region	string	Geographic region (Indian Subcontinent / Southeast Asia / Latin America / Middle East)
cancer_type	string	Primary cancer diagnosis (named types: Prostate, Colon, Lung; others anonymised)
histology_grade	string	Tumour differentiation grade (G1–G4)
ajcc_stage_ptnm	string	AJCC pTNM staging (I through IVB)
age_at_diagnosis	int	Patient age at diagnosis
sex	string	Male / Female
surgery	string	Surgical intervention performed
chemotherapy	string	Chemotherapy regimen
chemotherapy_cycles	int	Number of chemotherapy cycles administered
targeted_therapy	string	Targeted therapy agent
immunotherapy	string	Immunotherapy agent
radiation	string	Radiation modality and intent
msi_status	string	Microsatellite instability status (MSI-High / MSI-Low / MSS)
molecular_pathology_ngs	string	NGS mutation panel result
molecular_pathology_ihc	string	IHC expression result
pet_ct	string	PET-CT metabolic activity findings
serum_biomarkers	string	Serum biomarker status
haematology_status	string	Haematology assessment
biochemistry_status	string	Biochemistry assessment
performance_status_ecog	int	ECOG performance status (0–4)
treatment_response	string	Best response (Complete / Partial / Stable / Progressive)
recurrence	bool	Whether recurrence was detected
survival_months	int	Survival duration in months
vital_status	string	Alive / Deceased
has_wes_data	bool	Whether WES data is available for this patient
has_wsi_data	bool	Whether WSI data is available for this patient
has_biobank_specimen	bool	Whether biobank specimen is available

Key Differentiators

• Non-Caucasian patient populations placed at the centre of dataset design — a principle rare in oncology AI
• Multimodal from day one: genomic + imaging + clinical + biospecimen, all linked per patient
• 13 clinical data categories per patient including staging, surgery, chemotherapy, targeted therapy, molecular pathology (NGS + IHC), PET-CT, serum biomarkers, haematology, and biochemistry
• 15 countries across the Indian Subcontinent, Southeast Asia, Latin America, and the Middle East — populations systematically absent from TCGA, ICGC, and most major oncology training datasets
• Physical biobank enables generation of future modalities (spatial transcriptomics, single-cell RNA-seq)
• Scalable cohort: 25,000 today → 50,000 target in 12 months

IP & Compliance

• Full commercial and AI/ML training rights secured under exclusive licensing agreement
• Ethics and informed consent frameworks aligned with source jurisdictions
• De-identified patient data — no personally identifiable information (PII)
• Compliant with applicable data protection and cross-border transfer regulations (DPDP Act 2023, HIPAA, GDPR as applicable)

Citation

@dataset{lh2_oncology_2025,
  title={LH2 Data — Multimodal Oncology Dataset},
  author={LH2 Data Labs},
  year={2025},
  publisher={Hugging Face},
  note={Multimodal oncology dataset: 25K+ patients across the Global South}
}