"Waking Up" Data in Clinical Workflows with AI - with Mathew Paruthickal of Sanofi

Paruthickal outlines the crawl-walk-run methodology for integrating diverse data sources in regulated environments. The approach emphasizes starting with small user bases, demonstrating quick wins, and building modular systems that layer capabilities over time while maintaining business alignment through co-authored problem statements.

• •Start with crawl phase: show what's possible in weeks, not months, to demonstrate value of connecting data and documents
• •Build modular systems with iterative versioning rather than attempting full-scale deployment on day one
• •Embed product thinking into engineering pipelines through co-authored problem statements with business stakeholders
• •Involve business early and often to ensure engineering evolution matches business priorities
• •Balance technical capability with realistic timelines - 'Rome wasn't built in a day' approach to platform development

Discussion of how life sciences organizations must architect AI systems with traceability and compliance as day-one priorities. Paruthickal explains the technical requirements for creating explainable AI frameworks that can track data provenance, document versions, and provide audit trails for regulatory purposes.

• •Set boundaries, not bottlenecks - encourage experimentation while maintaining governance standards
• •Build explainability as a day-one priority, not a post-deployment addition to AI systems
• •Implement auditability frameworks that track data sources, document versions, and lineage for every AI output
• •Ensure AI proves both accuracy AND auditability to meet regulatory requirements
• •Ground AI answers in verifiable truth with traceable references to original source documents

Exploration of the two critical human roles in AI systems: experts who verify outputs and domain specialists who design explainable systems. Paruthickal emphasizes that in life sciences, humans must certify, sign off, and be held accountable, with AI serving to amplify rather than replace human judgment.

• •AI can accelerate, summarize, and recommend, but only humans can certify and be held accountable in regulated industries
• •Design editable components that allow human experts to validate and modify AI-generated content
• •Build systems around trust, traceability, and ethical responsibility rather than just compliance
• •Generative AI amplifies, protects, and documents human judgment rather than replacing it
• •Require human sign-off for critical decisions like medical accuracy validation and risk interpretation

Concrete use cases showing how AI transforms manual-heavy workflows in adverse event reporting, ICSR document generation, and clinical narrative writing. Paruthickal details specific techniques including auto-extraction from free-form text, LLM-based pre-filling, and reinforcement learning for model fine-tuning.

• •Auto-extract data from adverse event reports to eliminate manual transcription of patient calls
• •Pre-fill Individual Case Safety Reports (ICSR) using LLM training on structured adverse event data
• •Apply reinforcement learning and fine-tuning based on historical case data to improve accuracy and speed
• •Generate high-quality clinical narratives from structured study data with human validation
• •Prioritize regulatory precision, patient safety, and compliance as non-negotiable requirements for all AI applications

Framework for creating proactive AI systems that surface risks and optimize workflows through continuous learning. Paruthickal explains how capturing human overrides, corrections, and approvals creates training signals that improve AI accuracy over time, while monitoring ensures ongoing trust and regulatory compliance.

• •Capture every human feedback action (overrides, corrections, approvals) as training signals for continuous improvement
• •Design pipelines that listen to triggers and connect business problems with technical solutions
• •Implement few-shot learning that starts simple and improves through accumulated human feedback
• •Build monitoring as a first-class citizen - 'if you cannot monitor it, you can't trust it'
• •Pair technical pipelines with cultural habits of feedback, monitoring, and continuous refinement for proactive workflows
• •Ensure continuous improvement and compliance are baked into system architecture, not added later