Large Language Model Reads Radiologists' Notes to Flag Patients for Follow-Up Imaging

Why It Matters

The solution tackles a major source of diagnostic error—missed follow‑ups—by automating detection in high‑volume settings, thereby enhancing patient safety and care continuity.

Key Takeaways

• LLM extracts follow‑up needs from unstructured radiology notes
• Flags 6.18× more cases than prior macro system
• Achieves 97% recommendation detection accuracy
• Determines timing, exam type, diagnosis with 94% accuracy
• Boosts patient follow‑up scheduling in high‑volume settings

Pulse Analysis

Diagnostic errors often stem from missed follow‑up imaging, a problem amplified in health systems that process hundreds of thousands of radiology studies annually. Traditional electronic health record (EHR) workflows rely on structured templates or simple macros, which struggle to capture nuanced recommendations embedded in narrative reports. By deploying a large language model (LLM) that can interpret free‑text clinical impressions, hospitals gain a more reliable safety net that flags patients before gaps in care emerge, directly addressing a critical vulnerability in diagnostic pathways.

The Parkland Health study demonstrates the practical impact of this technology. After training the LLM on a random sample of 10,000 radiology notes, researchers expanded the evaluation to 120,000 imaging studies over three months. The model correctly identified 97% of follow‑up recommendations and outperformed the existing macro system by a factor of 6.18, raising flagged cases from 83 to 513. Moreover, it achieved 94% accuracy in pinpointing the appropriate timing, exam type, and underlying diagnosis, enabling care teams to prioritize and schedule scans with unprecedented precision.

Beyond immediate workflow gains, the integration of LLM‑driven decision support signals a broader shift toward AI‑augmented clinical operations. Health systems can scale this approach across specialties, reducing reliance on manual chart reviews and freeing staff to focus on patient interaction. As regulatory frameworks evolve and data‑privacy safeguards mature, such tools are poised to become standard components of diagnostic safety strategies, ultimately driving higher adherence rates, lower repeat imaging costs, and better health outcomes for patients nationwide.