AI Training Data Transparency

De-identified clinical event data from partner long-term care facilities; voice biomarker datasets through strategic partner; publicly available cognitive assessment benchmarks. All data sources are properly licensed or owned by Scienza Health or its partners.

Training machine learning models for cognitive screening (0.89 AUC Lancet-validated methodology); voice biomarker pattern recognition for early cognitive decline detection; healthcare workflow automation for intake, follow-up, and clinical documentation.

Our AI models are trained on 11M+ clinical events and 2M patient records, representing diverse populations across memory care, skilled nursing, assisted living, and primary care settings.

All training data is properly licensed from healthcare data partners or owned by Scienza Health. Proprietary algorithms and methodologies are patented where applicable. digitalhumanOS™ and Gia™ are registered trademarks.

Licensed from healthcare data partners under data use agreements; collected through IRB-approved research studies with informed consent; aggregated from consented clinical deployments with appropriate data governance protocols.

All data is de-identified in compliance with HIPAA Safe Harbor standards. No directly identifiable patient information is included in training datasets. Personal health information (PHI) is never used in model training without proper de-identification.

Synthetic data augmentation is used for edge cases and underrepresented scenarios following healthcare machine learning best practices. Synthetic data generation methodology follows established protocols to ensure clinical validity.

Data collection period: 2010-2025. First use in AI model training: 2024. Training datasets are updated quarterly to incorporate new validated clinical data while maintaining data quality standards.

HIPAA de-identification protocols applied to all data; outlier detection and removal using statistical methods; quality validation against established clinical standards; demographic representation balancing to reduce bias.