Radiology has become very good at generating follow-up recommendations and increasingly good at detecting them automatically. What it has not solved is the harder downstream problem: ensuring the recommended next step actually happens for the right patient, at the right level of urgency. Treating every recommendation as if it were the same—routing all of them through a generic, uniform notification or scheduling process—results in two failures at once. Low-risk findings get overworked, and genuinely dangerous findings slip through because the system has no way to treat them differently. Effective follow-up requires stratification: matching the intensity of communication, intervention, and tracking to the risk and urgency of the finding and to the barriers a given patient faces, and then measuring completion rather than mere notification.
One Workflow Does Not Fit Every Finding
Radiology findings differ systematically in risk, urgency, and the firmness of the recommendation attached to them, and that heterogeneity is precisely what a single workflow cannot accommodate. Actionable incidental findings (AIFs) are common on cross-sectional and emergency imaging, and the rising use of imaging in the emergency department has increased the volume of patients carrying findings unrelated to the reason they were scanned (Fu et al., 2023). These range from trivial to life-threatening, yet they often end up in the same undifferentiated queue.
Compounding this, radiologists themselves vary widely in their recommendations for follow-up. In a retrospective analysis of 318,366 reports, Cochon et al. (2019) found roughly a sevenfold variation among radiologists in whether a report contained a follow-up recommendation at all, after accounting for patient, modality, and setting factors. If the input to the follow-up process is that inconsistent—firm versus soft, urgent versus routine—then handling every recommendation generically guarantees a mismatch between clinical need and organizational response.
Professional communication standards already assume stratification rather than uniformity. The Royal College of Radiologists (2012) distinguishes explicitly among critical, urgent, and unexpected-but-less-urgent findings, with correspondingly different communication expectations. And the American College of Radiology has proposed a nationally standardized three‑tier urgency model—Immediate, Soon, and Routine—for critical imaging results (Leet et al., 2025). The implication is that follow-up and communication pathways should be tiered by design.
A simple contrast makes the point concrete. A scheduled 12-month surveillance CT for a stable, benign lesion can reasonably proceed via automated scheduling and a patient-portal notification; the baseline risk of deferral is low, and the process can tolerate light-touch handling. A new, potentially malignant incidental lung nodule identified in the ED cannot. It calls for immediate communication with the referral provider and active tracking until the follow-up is verified as complete. Sending both through the same channel either wastes scarce navigation resources in the first case or, far worse, lets the second one disappear into an overwhelmed inbox.
Match the Intervention to the Case
Detection is necessary but not sufficient. Automated, natural language processing–based systems can identify follow-up recommendations in free-text reports with high accuracy—Mabotuwana et al. (2019) reported 97.9% detection accuracy across large datasets—but identifying the recommendation is not the same as completing it. Adherence depends on the intervention that follows detection, and different interventions yield different results.
The evidence supports a graded spectrum of interventions, matched to the tier of the finding:
- Automated notification alone (EHR alerts, portal messages)—appropriate for low-risk, routine surveillance where baseline completion is relatively high.
- Staff outreach indicated for high-risk findings and for overdue routine cases. When Mannix et al. (2021) notified providers of overdue recommendations, they converted a meaningful share of stalled cases into completed or appropriately closed cases and, in the process, surfaced clinically important diagnoses.
- Patient navigation additional support is necessary when social or financial barriers are likely to impede adherence. Closed-loop programs increasingly build in a dedicated navigator and a safety-net that intervenes when imaging is not completed by the expected date (Fu et al., 2023; Kapoor et al., 2023).
- Leadership escalation reserved for repeated failures in high-risk cases, where the breakdown is organizational rather than individual and accountability needs to move up the chain.
The unifying principle is proportionality: the intervention should cost what the risk justifies. Portal messages are cheap and fine for benign surveillance; a suspicious mass in a hard-to-reach patient warrants a navigator, not another unread message.
Account for Patient Barriers
Even a correctly categorized, well-communicated recommendation can fail at the patient level, and those failures are not randomly distributed. Placing an order does not guarantee completion when transportation, cost, insurance complexity, or the absence of a usual source of care impedes it.
In a study of patient-level predictors of adherence, Ángel-González Calvillo et al. (2022) found that patients with Medicaid had roughly 4 times lower odds of completing recommended follow-up imaging than those with commercial insurance (odds ratio, 0.24; 95% CI, 0.06–0.88; p = .032). Notably, in that analysis, several factors often assumed to drive disparities—age, sex, race/ethnicity, primary language, body mass index, and neighborhood socioeconomic status—were not independently associated with completion once insurance was accounted for, underscoring that insurance-related and structural barriers deserve direct attention rather than assumption. Work using automated closed-loop tracking similarly links adverse social determinants of health to a higher risk of not completing clinically necessary follow-up imaging, which is why such programs pair tracking with safety-net outreach (Kapoor et al., 2023).
These findings map directly onto workflow design:
- Transportation, cost, and insurance barriers mean that high-risk findings may require navigation and financial counseling pathways, not just an order in the chart.
- Health literacy and portal access shape whether a patient grasps the urgency of a finding; higher-risk cases should trigger a phone call or in-person explanation rather than relying on a portal message that may never be opened.
- Fragmented care or the absence of a primary provider makes the “usual provider will handle it” assumption unreliable, which argues for assigning responsibility explicitly to radiology, the ED, or a centralized navigation function for certain populations.
Designing follow-up so that every patient can easily complete an order is, in effect, a design choice that lets disadvantaged patients fall out of the system.
Define True Closure
Much of the follow-up problem stems from an impoverished definition of success. Sending an alert is not closing the loop. Closed-loop communication frameworks hold that a loop is closed only when the recommendation is received, acknowledged, and acted upon—and when the necessary follow-up is either completed or explicitly documented as declined or not indicated (Royal College of Radiologists, 2012; Fu et al., 2023).
The data make clear how wide the gap is between “notified” and “closed.” In the Mannix et al. (2021) notification program, of 177 overdue patients whose providers were notified, only 36 (20.3%) went on to complete follow-up, 34 (19.2%) were documented by a provider as not indicated, and 107 (60.5%) were still lost to follow-up after notification. In other words, even an active notification left the majority of overdue cases unresolved—vivid evidence that issuing an alert cannot be equated with case closure.
Operationally, “closure” should be defined as one of three documented end states:
- Completed—recommended imaging (or an appropriate alternative) was performed within a clinically suitable timeframe.
- Appropriately declined / not indicated—a clinician documents that follow-up is not warranted (e.g., because of comorbidities, patient preference, or updated clinical information).
- Transferred—responsibility explicitly reassigned to another provider or institution, with documented, accepted handover.
Anything else—an alert sent but unacknowledged, an order placed but never scheduled—is an open loop and should be visible as such.
Measure What Happens Next
Stratified follow-up only works if the organization measures completion rather than detection. The reassuring news is that these outcomes are measurable, and existing studies model how they are measured. Reported adherence to follow-up imaging recommendations varies widely across settings—on the order of roughly one-third to two-thirds of cases in the published range summarized by Ángel-González Calvillo et al. (2022)—which means that measuring your own rate is both feasible and likely to reveal substantial room for improvement. Notification studies further quantify impact: Mannix et al. (2021) reduced the incomplete-follow-up rate from 26.0% to 20.7% (p = .002) and, when non-indicated studies were included, resolved 39.5% of overdue cases, yielding four clinically important diagnoses, including a biopsy-proven malignancy. Moreover, disparity analyses show that stratifying completion by insurance and social needs surfaces inequities that an aggregate number would hide (Ángel-González Calvillo et al., 2022; Kapoor et al., 2023).
A concrete measurement set for a stratified program might include:
- Time from recommendation to first action (order placed or outreach initiated) by risk tier: Detects whether higher-risk tiers actually move faster.
- Completion rate by risk tier (immediate, high-risk, routine, ambiguous): Shows where the pathway succeeds or breaks at the resolution that matters.
- Proportion of cases unresolved past defined thresholds, and the share eventually closed via escalation or navigation: Quantifies the “open loop” backlog and whether escalation works.
- Disparities in completion by insurance type, neighborhood socioeconomic status, race/ethnicity, and measured social-needs variables make equity a monitored outcome rather than an afterthought.
- Measuring completion by tier and by subgroup turns follow-up from an article of faith into a managed process.
Radiology AI Plus Stratified Follow-Up
Production AI in radiology has demonstrated that recommendations can be extracted from free-text reports with high accuracy and that patients who would otherwise be lost can be flagged when detection is embedded in a real workflow (Mabotuwana et al., 2019; Fu et al., 2023). However, the adherence, disparities, and overdue-recommendation literature is equally clear that detection alone does not produce complete care. Adherence is far from universal (Ángel-González Calvillo et al., 2022), notification leaves most overdue cases unresolved without further action (Mannix et al., 2021), and social and structural barriers systematically pull disadvantaged patients out of the pathway (Kapoor et al., 2023).
The next stage of radiology AI, then, is not primarily about finding more lesions. It is about orchestration—stratifying findings by risk and urgency, matching each to the right intensity of automation, outreach, navigation, or escalation, accounting for the barriers a specific patient faces, and defining success as a documented, closed loop rather than a sent alert. An AI system that detects a critical finding has done something valuable; a system that ensures that finding reliably results in completed, equitable care has done what actually protects the patient.
You can read more about what happens after detection in our recent article in Electronic Health Reporter: After the Scan: Where Radiology AI Falls Short
References
Ángel-González Calvillo, A., Kodaverdian, L. C., Garcia, R., Lichtensztajn, D. Y., & Bucknor, M. D. (2022). Patient-level factors influencing adherence to follow-up imaging recommendations. Clinical Imaging, 90, 5–10. https://doi.org/10.1016/j.clinimag.2022.07.006
Cochon, L. R., Kapoor, N., Carrodeguas, E., Ip, I. K., Lacson, R., Boland, G., & Khorasani, R. (2019). Variation in follow-up imaging recommendations in radiology reports: Patient, modality, and radiologist predictors. Radiology, 291(3), 700–707. https://doi.org/10.1148/radiol.2019182826
Fu, T., Berlin, S., Gupta, A., Plecha, D., Sunshine, J., & Sommer, J. (2023). Implementing a streamlined radiology workflow to close the loop on incidental imaging findings in the emergency department. Journal of Digital Imaging, 36(3), 776–786. https://doi.org/10.1007/s10278-022-00773-x
Kapoor, N., Lynch, E., Lacson, R., Flash, M. J. E., Hammer, M. M., Boland, G. W., & Khorasani, R. (2023). Predictors of completion of clinically necessary radiologist-recommended follow-up imaging: Assessment using an automated closed-loop communication and tracking tool. American Journal of Roentgenology, 220(3), 429–440. https://doi.org/10.2214/AJR.22.28378
Lee, H. Y., Jeon, J. Y., & Park, S. H. (2025). Critical radiology results in thoracic imaging: Categorization by urgency and clinical outcome. Clinical Imaging, 126, 110594. https://doi.org/10.1016/j.clinimag.2025.110594
Mabotuwana, T., Hall, C. S., Hombal, V., Pai, P., Raghavan, U. N., Regis, S., McKee, B., Dalal, S., Wald, C., & Gunn, M. L. (2019). Automated tracking of follow-up imaging recommendations. American Journal of Roentgenology, 212(6), 1287–1294. https://doi.org/10.2214/AJR.18.20586
Mannix, J., LaVoye, J., Wasserman, M., Lada, N. E., Onoue, K., Hassan, K., Sarangi, R., Haroon, S., Gaffar, A., Qureshi, M. M., & Gupta, A. (2021). A notification system for overdue radiology recommendations improves follow-up and diagnostic rates. American Journal of Roentgenology, 217(2), 515–520. https://doi.org/10.2214/AJR.20.23173
Royal College of Radiologists. (2012). Standards for the communication of critical, urgent, and unexpected significant radiological findings (2nd ed.). The Royal College of Radiologists.
