Nikesh Kotecha on Blueprint for Trust: How ChatEHR Establishes a Framework for Responsible AI in Clinical Care
Affiliated with Stanford University, Stanford Health Care was awarded a 2025 InterSystems Impact Award for its AXIOM initiative for fast access to electronic health records. In turn, AXIOM underpins the ChatEHR medical chatbot that Nikesh Kotecha, head of the Data Science team at Stanford Health Care, presents below. The Impact Award recognizes the work of the team to deliver low-latency, comprehensive access to patient data from complex electronic health record (EHR) systems. This foundation has been essential to enabling the ChatEHR platform.
Contribution by Nikesh Kotecha
The integration of Large Language Models (LLMs) into health systems holds the promise of reducing administrative burden and cognitive overload for clinical teams. However, deploying such tools in clinical workflows requires a commitment to safety, governance, and continuous monitoring. Stanford Medicine’s ChatEHR platform, an initiative developed as part of an enterprise investment in establishing a Data Science pillar, serves as an essential case study for any health system aiming for responsible AI adoption.
By embedding a real-time AI platform directly within the electronic health record (EHR), ChatEHR allows clinicians and staff to query patient data in plain language, improving efficiency and freeing time for direct care.
Setting up a responsible clinical AI system requires focus on three key areas: a robust platform architecture, transparent governance, and continuous evaluation.
1. The Foundational Platform: Real-Time Data and Scalability
A significant barrier to integrating AI tools into clinical workflows is often fragmented data sources and latency in data access. Traditional reporting systems were unsuitable for the real-time applications needed at the point of care.
The ChatEHR Platform tackles this with a sophisticated, four-pillar architecture:
- LLM Router: Provides secure access to a variety of models, standardizing calls and handling centralized logging.
- Real-Time Data Access: Fetches and organizes clinical information using Fast Healthcare Interoperability Resources (FHIR). This enables near real-time responsiveness by combining FHIR and HL7v2 messaging with optimized query performance.
- Function Server: Transforms generic AI capabilities into healthcare-specific functions and task-specific endpoints, powering workflow automations.
- EHR Integration: Manages secure connections and integrates the custom UI directly into Epic Hyperspace, maintaining authentication and patient context.
This scalable architecture, underpinned by Stanford’s AXIOM framework, proved critical for performance. For instance, the system cut retrieval times by over 95%—from nearly two minutes to four seconds—for some AI workflows. This focus on low-latency data access is fundamental to clinical decision-making and the AXIOM initiative, which provides this breakthrough data capability, was recognized as an InterSystems Impact Award winner in 2025.
2. Governance and Oversight
Responsible AI deployment starts with organizational frameworks and strong governance models. ChatEHR is explicitly guided by Stanford Medicine’s Responsible AI Lifecycle. Oversight is managed through the Data Science Executive Committee (DSEC), ensuring transparency and alignment throughout the project life cycle. Furthermore, all system operations adhere to HIPAA and HITECH standards, utilizing role-based access controls inherited directly from the EHR to limit data exposure. This structured approach ensures that AI deployment is treated not as a one-off technical build, but as a governed, enterprise-wide strategy.
3. Evaluation and Monitoring with MedHELM
Crucial to maintaining trust and safety is developing approaches for evaluating LLM performance, a task made challenging by the flexibility and variability of generative AI outputs.ChatEHR implemented MedHELM (Medical Holistic Evaluation of Language Models), a generative AI evaluation framework developed in collaboration with the Stanford Center for Biomedical Informatics Research (BMIR) and the Institute for Human-Centered AI (HAI )
By combining governance, an embedded platform for low-latency data access, and monitoring via the MedHELM framework, the ChatEHR initiative provides a replicable blueprint for health systems seeking to translate AI innovation into measurable clinical and operational impact while upholding standards of safety and responsibility.
………………………………………
Nikesh Kotecha, Head of Data Science at Stanford Health Care, tasked with building a group to develop and deploy AI-guided models into clinical and operational workflows.
Prior to that, Nikesh started and led the informatics efforts at the Parker Institute for Cancer Immunotherapy, a non-profit organization bringing together 7 top cancer centers
____________________________
HEALTHCARE PROVIDERS
Organization: Stanford Health Care
Innovation Name: AXIOM
World Winner – Revenue Generation
Regional Winner – North America
Stanford Health Care’s AXIOM initiative (Advance Extraction for Intelligent Orchestration and Medical Insight) is a transformative solution that enables fast access to complex electronic health record (EHR) data. Using a unique recursive data retrieval method, AXIOM minimizes duplication and reduces query times from minutes or hours to seconds. It leverages a FHIR (Fast Health Interoperability Resources) repository on InterSystems IRIS, to power advanced AI and large language model (LLM) applications. This architecture enables real-time clinical insights, enhances care team collaboration, and supports use cases such as augmented triage and optimized patient flow across a variety of care settings.
Sponsored by InterSystems and selected by an independent panel of judges from Massachusetts Institute of Technology (MIT).