Agentic Radiology Incidental Findings Follow-up Orchestration

1. Problem / Context

Incidental findings in radiology—such as pulmonary nodules, thyroid nodules, and adrenal lesions—are common and clinically significant. Yet many mid-market health systems and radiology groups still rely on manual handoffs, spreadsheets, and inbox reminders to ensure guideline-based follow-up. The result is leakage and risk: patients don’t complete recommended imaging, providers lose track of next steps, and organizations face liability exposure and performance penalties. For $50M–$300M providers with lean teams, the operational load of tracking every incidental finding from report to closure is simply too high without automation.

At the same time, regulated environments demand guardrails. Any automation must respect PHI handling, produce auditable recommendations, and support human oversight. The practical question is not “Can AI help?” but “How do we orchestrate a governed, reliable workflow from radiology report to completed follow-up exam?”

2. Key Definitions & Concepts

• Incidental finding: A clinically relevant abnormality identified on an imaging study performed for another reason (e.g., a pulmonary nodule found on a CT abdomen).
• Guideline-based follow-up: Modality and interval recommended by professional guidelines (e.g., ACR) based on finding type, size, and risk factors.
• Agentic AI: An orchestration approach where AI systems perceive, reason, and act across tools and APIs—extracting details from reports, applying rules, opening EHR tasks, coordinating outreach, and escalating exceptions.
• Clinical NLP: Natural language processing tuned to radiology reports to extract finding attributes (type, size, location), map to ACR categories, and determine preliminary recommendations.
• Lakehouse foundations: Storing reports and events in Delta tables, governed by Unity Catalog for PHI controls; using Databricks Jobs for workflow execution and MLflow for model lifecycle.
• Human-in-the-loop (HITL): Navigators or radiologists validate recommendations, and ordering providers approve and sign orders. Exceptions route to specialists.
• Closure: The loop from initial finding through scheduled follow-up, completed imaging, documented results, and final resolution.

3. Why This Matters for Mid-Market Regulated Firms

Mid-market providers face the same clinical and regulatory stakes as large systems but with tighter budgets and staffing. Manual tracking doesn’t scale, and missed follow-ups can lead to patient harm, rework, malpractice exposure, and lost downstream revenue. Payers and accrediting bodies increasingly expect auditable, guideline-aligned processes. That means:

• Reducing time from finding to scheduled follow-up
• Raising closure rates while lowering outreach burden
• Enforcing PHI controls and audit trails across every recommendation, task, and order
• Minimizing IT lift with a pragmatic, interoperable architecture

A governed, agentic workflow built on the Lakehouse lets lean teams achieve this with confidence.

4. Practical Implementation Steps / Roadmap

1. Ingest radiology reports into Delta

   • Stream or batch reports from RIS/PACS or EHR into a Delta table. Include study metadata, report text, patient demographics, and ordering provider.
   • Use Unity Catalog to classify PHI and control access down to column and row.

2. Extract findings with clinical NLP

   • Run an NLP service to detect finding type (e.g., pulmonary nodule), size, features, and ACR category.
   • Normalize units and map to structured fields (e.g., “8 mm solid nodule, right lower lobe”). Track model confidences.

3. Apply guideline rules and risk stratification

   • A rules engine encodes ACR guideline logic to propose modality and interval (e.g., low-risk adult with 6–8 mm pulmonary nodule → CT at 6–12 months).
   • Personalize by age, smoking history, and prior imaging; fall back to conservative defaults when data is incomplete.

4. Human-in-the-loop validation

   • Present recommendation to a navigator or radiologist for quick validation. Capture overrides and rationale. Version both the rules and decisions.

5. Create EHR tasks and orders

   • Via EHR APIs, open tasks for the ordering provider, generate draft orders, and pre-populate scheduling workqueues. Ensure the provider reviews/signs orders.

6. Patient outreach and scheduling

   • Select outreach channel (portal message, SMS, phone, mail) based on consent and success rates.
   • Include clear instructions and available scheduling slots; log each outreach attempt with timestamps.

7. Track closure and escalate

   • Monitor task status, scheduled appointments, completed studies, and final reports. If deadlines approach without action, escalate to the provider or specialty clinic.
   • Close the loop by recording the completed follow-up and resolution.

8. Orchestration and tooling

   • Use Databricks Jobs to schedule pipelines, call the NLP service, trigger rules evaluation, and invoke EHR/tasking and messaging APIs.
   • Maintain state in Delta with event-driven tables (finding detected → recommendation validated → task created → outreach sent → follow-up completed).

[IMAGE SLOT: agentic AI workflow diagram connecting PACS/RIS, Delta Lake, clinical NLP, ACR rules engine, EHR task/order API, patient outreach channels, and a navigator console]

5. Governance, Compliance & Risk Controls Needed

• PHI governance with Unity Catalog: Tag sensitive columns, enforce attribute-based access, and separate sandboxes from production. Log all data access.
• Versioned guideline rules: Keep rules in a versioned repository with effective dates and approval signatures; show exactly which rule set drove each recommendation.
• MLflow model approvals: Register NLP models, require stage gates (Staging → Production), and track performance drift with shadow evaluations.
• Immutable audit: Record every recommendation, validation, order creation, outreach attempt, and outcome in append-only Delta tables. Preserve request/response payload hashes for key API calls.
• HITL and sign-off: Ensure a navigator or radiologist validates recommendations and the ordering provider signs orders. Capture accountability in the audit log.
• Vendor lock-in mitigation: Favor open formats (Delta, Parquet), standard APIs, and modular adapters for EHR and messaging to keep swap costs low.

[IMAGE SLOT: governance and compliance control map showing Unity Catalog PHI zones, MLflow model registry gates, versioned ACR rules, and immutable audit tables with HITL checkpoints]

6. ROI & Metrics

Leaders should track a concise set of operational and clinical metrics:

• Closure rate: Percentage of incidental findings with a completed, guideline-aligned follow-up within the recommended interval.
• Cycle time: Days from initial report to scheduled appointment; days from scheduled to completed study.
• Error/exception rate: Percentage requiring manual correction due to NLP misclassification, incomplete data, or rule ambiguity.
• Provider burden: Average touches per case (navigator minutes, provider reviews) before and after automation.
• Leakage reduction: Recovered downstream imaging and specialty visits attributable to the program.
• Payback: Compare avoided adverse events/rework plus retained revenue with program cost.

Example: A 12-site radiology group processes ~4,000 incidental findings per quarter. Baseline closure is 55%, with an average 24-day time-to-schedule and heavy navigator workload. After deploying governed agentic orchestration, closure rises to 78–85%, time-to-schedule drops to 10–12 days, and navigator touches per case decrease by 30–40%. Even with conservative valuation of retained follow-up imaging, the payback period typically lands within two to three quarters.

[IMAGE SLOT: ROI dashboard with closure rate trend, time-to-schedule distribution, navigator workload bar chart, and payback projection]

7. Common Pitfalls & How to Avoid Them

• Brittle automation (RPA-only): Templated scripts fail on varied report language. Use robust NLP with confidence thresholds and fallbacks.
• Stale guidelines: ACR updates periodically. Version rules, track effective dates, and alert when a rule set changes.
• Incomplete context: Missing prior imaging or smoking history skews risk. Integrate with longitudinal data and escalate low-confidence cases for HITL review.
• EHR integration gaps: Tasking, ordering, and messaging APIs differ by vendor. Build adapters, test in non-prod, and confirm provider sign-off workflows.
• Outreach drop-off: One-size-fits-all outreach underperforms. Personalize channel and cadence; retry with intelligent sequencing.
• Weak auditability: Without immutable logs, defensibility suffers. Persist each decision and action with timestamps and provenance.

30/60/90-Day Start Plan

First 30 Days

• Discovery: Inventory report types, volumes, and top incidental categories (pulmonary, thyroid, adrenal).
• Data checks: Validate feeds from RIS/PACS/EHR; land sample reports in Delta with Unity Catalog tags for PHI.
• Governance boundaries: Define HITL checkpoints, provider sign-off requirements, and access policies.
• Rules baseline: Load initial ACR rule sets with versioning; identify local practice variations.

Days 31–60

• Pilot workflows: Stand up NLP for one category (e.g., pulmonary nodules) and connect rules engine.
• Agentic orchestration: Use Databricks Jobs to chain extraction → rules → HITL → EHR task creation.
• Security controls: Enforce least-privilege access, enable audit tables, and register models in MLflow with approval gates.
• Evaluation: Measure precision/recall of NLP, recommendation acceptance rate, and time-to-schedule.

Days 61–90

• Scaling: Add thyroid/adrenal pathways; expand outreach channels; harden adapters for EHR/tasking and messaging.
• Monitoring: Implement drift monitors, rule-change alerts, and SLA dashboards for closure metrics.
• Stakeholder alignment: Review ROI, provider feedback, and compliance sign-off; finalize operating playbook and handoffs.

9. Industry-Specific Considerations

• EHR variability: Epic, Cerner, and Meditech differ in tasking/order APIs. Plan for site-by-site adapter tuning and non-production validation.
• ACR nuance: Categories (e.g., TI-RADS for thyroid) and thresholds require careful encoding and testing, with local practice overlays where appropriate.
• Patient access: Consent and language preferences affect outreach effectiveness; support multi-channel and interpreter workflows.
• Multi-entity governance: Radiology groups serving multiple hospitals need tenant-aware data zones and audit segregation.

10. Conclusion / Next Steps

A governed, agentic approach transforms incidental finding follow-up from a brittle, manual chase into a reliable, auditable workflow that protects patients and clinicians. By combining clinical NLP, versioned guideline rules, EHR tasking and ordering, and intelligent outreach with HITL oversight, mid-market organizations can raise closure rates, shorten cycle times, and reduce burden on lean teams.

Kriv AI serves mid-market providers as a governed AI and agentic automation partner—helping teams get data-ready on the Lakehouse, operationalize MLOps, and implement the governance controls auditors expect. With practical adapters for EHR/tasking and messaging and a navigator console and closure dashboard, Kriv AI helps turn pilots into production-ready systems.

If you’re exploring governed Agentic AI for your mid-market organization, Kriv AI can serve as your operational and governance backbone.