Oncology Tumor Board Preparation and Summarization Automation

Oncology coordinators and fellows currently spend a full workday manually collating DICOM studies, PDF pathology reports, and genomic spreadsheets into a single presentation deck for weekly tumor boards. A custom agentic workflow automates this data fusion, narrative synthesis, and slide generation, reclaiming that time for higher-value patient coordination and clinical research.

Manual preparation leads to inconsistent formats, missing data points, and narrative gaps that force physicians to mentally reconstruct the case during the board. An automated workflow enforces a consistent, evidence-structured template—integrating segmented tumor volumes, key pathologic markers (e.g., PD-L1, KRAS), and relevant trial matches—so the multidisciplinary team can focus on decision-making, not data assembly.

The business value extends beyond meeting preparation. By structuring fragmented data into a computable patient profile, the workflow can automatically query internal trial databases and external registries (e.g., ClinicalTrials.gov) for potential matches based on genomic biomarkers and prior therapies. This turns the tumor board into an active recruitment engine, potentially reducing screening time for clinical trial operations by weeks.

Every automated briefing includes provenance tracking: which data sources were used, which AI models performed segmentation, and what logic generated the narrative summary. This creates a defensible audit trail for tumor board decisions, supports quality improvement initiatives, and generates structured, de-identifiable data lakes for retrospective research on treatment patterns and outcomes.

Implementation requires a LangGraph or similar orchestration layer coordinating specialized agents: a DICOM agent (fetches/parses segmented volumes from PACS), a pathology NLP agent (extracts key findings from unstructured reports in the EHR/LIS), a genomics agent (queries lab information systems), and a narrative synthesis agent (LLM-powered, with strict citation of source data). The workflow integrates via HL7/FHIR and direct APIs with Epic, Cerner, Varian ARIA, and digital pathology platforms.

Clinical safety is non-negotiable. The workflow is built with mandatory review gates. The generated briefing is presented to the assigned oncologist or coordinator in a secure web interface 24-48 hours before the board. They can edit, annotate, or flag discrepancies, with all overrides logged. The final, approved briefing is then distributed, ensuring the automated output is always under physician authority before guiding care decisions.

The core orchestration layer that ingests and aligns structured and unstructured data from PACS (segmented lesion volumes), the Laboratory Information System (LIS) for pathology reports, and the EHR/Genomic Data Commons for biomarker findings. Agents use retrieval-augmented generation (RAG) to query these systems, resolve patient identity, and temporally align findings to the current treatment timeline, creating a unified patient context for summarization.

A specialized LLM agent, governed by clinical guidelines and templated structures, that generates the draft tumor board summary. It synthesizes key imaging changes (e.g., -30% tumor volume per RECIST 1.1), critical pathology notes (e.g., PD-L1 80%), and actionable genomic variants into a coherent narrative. The agent cites source data, highlights discrepancies between modalities for human review, and formats output for presentation systems (PowerPoint, EHR-integrated viewer).

A mandatory human-in-the-loop control point where the oncology coordinator or designated physician reviews the auto-generated briefing. The interface displays source data, the AI's synthesis, and allows for edits, overrides, or requests for additional information. All changes are logged to an audit trail for governance and model improvement. Approved briefings are automatically distributed to the tumor board's schedule and participant portals.

The secure middleware that handles bidirectional data sync between the workflow and hospital IT. It listens for DICOM study-completion events from PACS, retrieves segmented results via REST APIs, fetches patient context via FHIR from the EHR (e.g., Epic, Cerner), and posts the final, approved summary back as a clinical note. This component manages authentication, error handling, and data persistence to ensure a seamless clinical workflow.

An observability layer that tracks operational and clinical metrics: case preparation time, auto-summary adoption rate, physician override frequency by data type (imaging vs. genomics), and user feedback scores. This data monitors workflow ROI, identifies bottlenecks (e.g., slow genomic data retrieval), and detects drift in source system data formats or summarization quality, triggering alerts for IT or clinical leadership.

Implementation logic that manages a staged pilot, typically starting with a single cancer type (e.g., breast) and one tumor board. This component controls feature flags, manages user access and training material delivery, and collects structured feedback. It ensures the workflow integrates with existing coordinator routines without disruption, allowing for iterative refinement of agents and interfaces before enterprise-scale deployment.