Closing Incidental Finding Loops: What NLP Follow-Up Systems Achieve in Practice

Diagnostic imaging reports contain thousands of critical follow-up recommendations that disappear into electronic records. Natural language processing models extracting unstructured report text have lifted follow-up completion from 42% to 68%, but direct staff outreach remains mandatory to overcome patient barriers.

The clinical scope of dropped incidental findings

When a patient undergoes a computed tomography (CT) scan for suspected kidney stones, the radiologist frequently detects unexpected abnormalities: a 6-millimeter ground-glass pulmonary nodule, an indeterminate thyroid nodule, or an adrenal mass. Professional guidelines (such as the Fleischner Society criteria) specify exact follow-up imaging intervals to rule out early malignancy.

Yet in real-world outpatient practice, between 35% and 50% of these actionable recommendations fail to result in completed follow-up exams. Primary care providers, overwhelmed by hundreds of daily in-basket notifications, fail to register recommendations buried in narrative radiology impressions, leaving patients vulnerable to delayed cancer diagnoses.

How natural language processing extracts unstructured recommendations

Recent multi-institution trials published in the Journal of the American College of Radiology (JACR) demonstrate that transformer-based natural language processing (NLP) models can extract follow-up recommendations from narrative radiology reports with sensitivity exceeding 94% and specificity above 96%.

Instead of relying on radiologists to manually enter structured codes, these systems continuously parse free-text impressions, identify recommended modalities (e.g., non-contrast chest CT in 6 months), and automatically populate an ambulatory tracking registry. In an October 2024 prospective deployment across a five-hospital regional network, automated registry tracking increased documented 12-month follow-up completion from 42.1% to 68.4%.

Despite high algorithmic accuracy in registry creation, the intervention cohort highlighted a critical behavioral barrier: patient engagement. When the health system initially sent automated portal messages and form letters alerting patients to their uncompleted nodule scans, scheduling rates rose by only 8 percentage points.

Patients reported that technical diagnostic language ('indeterminate pulmonary nodule concerning for neoplasm') caused acute anxiety, leading many to avoid booking appointments. Follow-up completion surged to 68.4% only when the registry was coupled with dedicated nurse navigators who conducted outbound telephone calls to explain the routine monitoring rationale and coordinate scheduling.

Operational essentials for closing diagnostic loops

Medical specialty groups and health systems implementing loop-closure tools should establish clear workflow governance:

Responsibility assignment: Determine whether the ordering physician, radiologist, or central navigation team owns overdue outreach.
Direct booking links: Enable patients to initiate scheduling requests via telephone or two-way messaging without navigating complex web portals.
Structured handoffs: Route incoming patient scheduling requests directly into front-desk scheduling worklists.

How Fluent handles follow-up scheduling requests

Fluent does not parse inpatient diagnostic imaging feeds or replace enterprise radiology registry engines. Instead, Fluent provides the essential inbound communication bridge when patients call to schedule their follow-up examinations. When a patient calls a clinic following an outreach letter, Fluent's AI phone receptionist captures their request, verifies patient identity, and queues an actionable WorkItem in the staff Inbox.

This ensures patient calls are never lost to busy signals or unreturned voicemails, closing the critical loop between registry identification and completed clinical care.

Sources & reporting notes

Published November 7, 2026. Sources are linked above; operational recommendations are Fluent’s editorial analysis.