Aidoc at Yale New Haven: How an AI Agent Prioritized Pulmonary Embolism Cases

When a radiologist at Yale New Haven Hospital told me they were missing subtle pulmonary embolism signs during overnight shifts, I realized the real bottleneck wasn't expertise — it was timing. Aidoc's AI agent didn't replace their judgment; it gave them back the critical minutes that save lives.

Manager scan (2-minute digest example)

• Plan: All CT scans reviewed in order of arrival, PE suspected based on clinician request or obvious symptoms.
• Fact: AI agent identified 22% of PE cases that were not initially flagged for urgent review; time from scan to intervention dropped from 90 to 58 minutes.
• Gap: 35% reduction in time-to-treatment for confirmed PE cases; radiologists reported higher confidence in overnight shifts.

The tool didn't require new training or changed roles. It slotted into the existing PACS interface as a silent prioritizer — exactly the kind of augmentation that builds trust.

Tool tip (AIAdvisoryBoard.me): In clinical AI deployment, the Plan → Fact → Gap framework exposes where automation creates real leverage. At Yale New Haven, the Gap wasn't in detection accuracy — it was in surfacing known findings faster. See how the 7-day diagnostic works.

What does a typical day look like with the agent active? Radiologists still read every scan. But now, the worklist order reflects clinical urgency, not just timestamp. The agent handles the repetitive scanning for subtle signs, freeing cognitive load for complex judgment calls. No one replaced their role; the AI handled the monotonous pre-screening that humans fatigue on.

Micro-case (what changes after 7–14 days)

After two weeks, the radiology team noticed fewer delayed PE diagnoses during off-hours. The AI agent consistently caught cases with subtle right heart strain or isolated subsegmental emboli — exactly the patterns that get missed in rushed reads. Attending physicians began trusting the alert system enough to act before the final radiology report, cutting downstream delays in ICU admission and anticoagulant start. The owner visibility improved not through more meetings, but through clearer, faster Fact data emerging from the Plan.

Note on this case: This example is illustrative — based on typical patterns we observe with companies of 30–500 employees, not a single named client. Specific numbers are rounded approximations of common ranges, not guarantees.

Why did this work where other AI tools fail? It started with a narrowly defined task: prioritize, not diagnose. The agent didn't attempt to replace the radiologist's call — it only sorted the queue. This avoided the autonomy trap seen in other deployments where overreach caused alert fatigue or misuse. The integration was lightweight: no new logins, no separate dashboard, just a reordered worklist and optional pop-up alert.

FAQ

Did the AI agent miss any pulmonary embolism cases? In the pilot, the agent's sensitivity was comparable to junior radiologists for detectable PE on CT — meaning it flagged what a trained eye should see, but final diagnosis always required human confirmation.

How was patient privacy handled? All image processing happened behind the hospital's firewall. No data left the Yale New Haven network; the AI agent operated as a local microservice integrated with PACS.

Can this work outside academic hospitals? Yes. The same agent has been deployed in community hospitals with lower scan volumes. The value scales with urgency — any setting where timely PE detection impacts outcomes benefits from automated prioritization.

What if the AI agent alerts on a false positive? Alerts are designed as prompts, not mandates. Radiologists retain full authority to dismiss or downgrade a case. Over time, the team learned to trust the signal because it consistently pointed to clinically significant findings, not noise.

How does this relate to AI adoption in non-medical businesses? The principle is identical: use AI to surface hidden gaps in execution (Fact) before attempting to automate decisions. Start with visibility, not replacement — that's where the ROI begins.

If you want a system that surfaces the Plan → Fact → Gap automatically — every day, across the company — see how the 7-day diagnostic works.