AI Integration for GE Emergency Department PACS

AI integration for GE Emergency Department PACS targets three primary surfaces: the reading worklist, the reporting module, and the notification/alerting system. The goal is to inject AI analysis—typically from a containerized inference service—between the modality and the radiologist's final read. When a CT or X-ray study is completed in the ED, the DICOM study is sent to the PACS and simultaneously routed via a DICOM Send or HL7 ORM/ORU message to a secure AI inference queue. Algorithms for trauma (e.g., ICH, pneumothorax, spine fracture) and non-trauma (e.g., appendicitis, bowel obstruction) are run in parallel, returning structured results as a DICOM SR (Structured Report) or via a REST API callback.

The returned AI findings are then used to dynamically reorder the radiologist's Centricity PACS worklist, pushing studies with high-probability critical findings to the top. For confirmed critical results, the integration can trigger automated critical result notifications via the existing hospital communication system (e.g., secure chat, pager), embedding a direct link to the flagged study in the zero-footprint viewer. Within the reporting interface, AI can pre-populate draft findings sentences and suggest relevant report macros, reducing dictation time for high-volume ED shifts. This creates a closed-loop workflow where AI acts as a force multiplier, ensuring the sickest patients are seen first while automating administrative overhead.

A production rollout requires careful governance. AI results should be presented as non-interruptive overlays or side-panels within the PACS viewer, requiring radiologist verification before final sign-off to maintain legal responsibility. An audit trail must log every AI inference, its result, and the radiologist's action (accepted, modified, rejected) for performance monitoring and regulatory compliance. Starting with a pilot—such as AI triage for non-contrast head CTs in suspected stroke—allows for workflow validation and change management before scaling to the full ED imaging portfolio. The integration should be designed to fail gracefully; if the AI service is unavailable, the PACS workflow continues uninterrupted, falling back to standard first-in-first-out reading.

Integration begins at the DICOM Study Router within the GE PACS ecosystem. As CT and X-ray studies are completed in the ED, the PACS automatically pushes anonymized DICOM images and metadata via a DICOM C-STORE SCU to a secure, on-premises or cloud-based AI Inference Gateway. This gateway, built on containerized services, manages the queuing, load balancing, and secure execution of AI models—such as those for detecting intracranial hemorrhage, pneumothorax, or fractures—against the incoming imaging data. Results, including bounding boxes, confidence scores, and structured findings, are packaged as DICOM Structured Reports (SR) or HL7 messages and sent back to the PACS via DICOM C-STORE SCP or a HL7 ADT/A04 interface.

Within the GE ED PACS, the AI results trigger two primary workflow actions. First, the Radiologist Worklist is dynamically re-prioritized. Studies flagged with high-confidence critical findings are elevated to the top of the list, tagged with visual alerts. Second, for the most time-sensitive detections (e.g., large vessel occlusion), the system can initiate an automated Critical Results Notification. This uses the PACS's existing notification framework or integrated secure messaging (like TigerText or Spok) to send an immediate alert with a de-identified case preview to the designated ED physician and stroke team, bypassing the wait for a radiologist's initial read. This creates a parallel, AI-assisted triage lane operating alongside the standard radiology workflow.

Rollout follows a phased, governance-first model. Initial deployment targets a single high-volume AI algorithm (e.g., ICH detection on non-contrast head CTs) in a silent mode, where results are logged but do not alter the worklist or trigger notifications. This allows for validation against ground truth and establishes baseline performance. After a governance committee approves, the system shifts to assistive mode, where worklist prioritization is enabled. Full notification workflows are the final phase, requiring clear protocols for alert acceptance, escalation, and documentation. All AI interactions are logged to a dedicated audit trail within the PACS or a separate system, capturing the original images, AI inference payload, result action, and any overrides by clinical staff for compliance and continuous model monitoring.

In the Emergency Department, AI integration must prioritize patient safety and clinician trust. For GE's ED PACS, this means implementing a multi-layered governance model. Key controls include:

• Role-Based Access Control (RBAC): Restricting which users (e.g., attending radiologists, ED physicians) can see AI findings and override them.
• Audit Logging: Every AI inference, user interaction, and result modification is logged to the study's DICOM metadata or an external audit database, creating a traceable chain of custody.
• Human-in-the-Loop Mandate: AI outputs are configured as non-interpretive findings suggestions or priority flags within the PACS viewer (e.g., a "Critical Finding Alert" overlay). The radiologist's final report is the sole source of truth.
• Model Performance Monitoring: Integrating with GE's Edison AI Platform or a separate LLMOps dashboard to track model drift, inference latency, and accuracy metrics (e.g., false positive rates for pneumothorax detection) against a validation set of ED studies.

A phased rollout mitigates risk and allows for workflow optimization. A typical four-phase approach for GE ED PACS is:

1. Silent Pilot: AI processes incoming DICOM studies (CTs, X-rays) in the background via a service listening to the PACS' DICOM SCP. Results are stored in a separate database but not displayed. This validates technical integration and establishes baseline performance without impacting care.
2. Non-Interruptive Alerts: AI findings are displayed as a non-obtrusive sidebar or color-coded flag in the GE viewer worklist and within the study hanging protocol. This introduces the AI as an assistive tool without disrupting the primary reading workflow.
3. Context-Aware Integration: AI insights become more embedded. For trauma pan-scans, AI-powered organ segmentation and injury scoring templates are suggested. For non-trauma chest X-rays, AI draft findings for common pathologies are populated into the speech recognition macro palette. All suggestions require explicit acceptance.
4. Workflow Orchestration: AI-driven prioritization automatically reshuffles the ED radiologist's worklist based on detected criticality (e.g., large vessel occlusion, tension pneumothorax). This phase integrates with the hospital's critical result notification system, using AI as a trigger for automated, protocol-driven alerts to the ED team via secure messaging, while maintaining the radiologist's final approval.

Post-deployment, governance is operational. A Clinical AI Committee—comprising radiologists, ED physicians, IT, and compliance—reviews audit logs, performance reports, and incident feedback quarterly. Change control procedures are established for any prompt updates, model retraining, or integration modifications. This structured, safety-first approach ensures the AI integration enhances the speed and accuracy of emergency care within GE's trusted clinical environment, building the foundation for scalable adoption across other high-acuity imaging workflows.