Here’s a high-level view of how a Cortellis-style database update typically works, balancing AI automation with human curation. Note that exact internal workflows are proprietary, but these elements reflect common industry practices.
- Data sources and intake
- Ingest from diverse sources: regulatory agencies (FDA, EMA, PMDA), journals, clinical trial registries, patents, company reports, press releases, and other public/partner feeds.
- Raw data arrives in structured and unstructured formats (text, PDFs, PDFs-to-text, etc.).
- AI-driven extraction and normalization
- Natural language processing and machine learning extract entities (e.g., compounds, targets, pathways, indications, regulatory actions, trial results) and relations (drug–target interactions, mechanism of action).
- Normalization and mapping to standard ontologies/elements (e.g., CAS/InChI keys for chemicals, MeSH/UMLS terms, regulatory statuses).
- Deduplication and conflict resolution at the data-object level (e.g., same drug from multiple sources).
- Rule-based validation and quality checks
- Consistency checks (units, dosages, dates, jurisdiction-specific rules).
- Cross-source reconciliation to resolve discrepancies (e.g., conflicting trial results or regulatory statuses).
- Human curation and validation
- Subject-matter experts (SMEs) review AI-flagged data, disambiguate entities, resolve ambiguities, and add nuanced annotations (mechanisms, manufacturing forms, dose information, regulatory pathways, patent details).
- Complex interpretations, rare data types, or high-stakes entries go through expert review.
- Quality assurance (QA) and governance
- Multi-layer QA: automated tests plus manual spot checks; traceability to source documents.
- Provenance capture: each data item links back to source(s) and the update rationale.
- Integration, impact assessment, and publishing
- Updates are integrated into the existing data model, with deduplication and relationship linking to other objects (e.g., compound–target networks, trial cohorts).
- Change logs or audit trails are generated to document updates, additions, deletions, and corrections.
- Update cadence and delivery
- Continuous ingestion with automated refreshes for new/updated sources.
- Periodic “full refresh” cycles (e.g., monthly or quarterly) to refresh mappings and recalculate derived fields.
- Release notes accompany updates to highlight new entries, updated statuses, and any data corrections.
- AI model maintenance and learning
- AI models are continuously retrained on corrected and newly curated data to improve extraction accuracy.
- Human feedback loops help fine-tune models for higher precision in niche areas.
- Data governance and security
- Access controls, provenance tracking, and compliance with licensing of source data.
- Regular audits of data quality and adherence to internal standards.
If you want, I can tailor this to a specific Cortellis module (e.g., Regulatory Intelligence, Drug Discovery Intelligence, or Patent/IP data) and outline a more concrete, module-specific flow. Also, tell me what you’re most interested in (update frequency, data provenance, AI vs human roles, or how to interpret change logs) and I’ll expand that part.