AI triage shows no impact on lung cancer diagnostic timelines

Artificial intelligence (AI) is increasingly positioned as a potential solution to delays in lung cancer diagnosis—but prospective evidence suggests that implementation strategy may be more important than the technology itself.

In a large, multicentre randomised trial published in Nature Medicine, investigators evaluated whether AI-driven prioritisation of chest X-rays (CXRs) in primary care could accelerate time to CT imaging and lung cancer diagnosis. Despite strong performance in retrospective studies, the intervention did not translate into measurable clinical benefit.

A prospective, multicentre randomised trial assessed whether AI-based prioritisation of abnormal CXRs could reduce time to CT imaging and lung cancer diagnosis in patients referred from primary care.

Patients undergoing CXR were randomised to either the standard radiology workflow or AI-assisted prioritisation, in which scans flagged as suspicious by the algorithm were moved up the reporting queue.

The key findings indicated no improvement in diagnostic timelines – median time to CT was 53 days in both groups, with no significant difference in time to diagnosis (44 vs 46 days). There were no significant differences in time to specialist referral, treatment initiation, or stage at diagnosis.  AI and radiology reports differed in 30.3% of CXRs. Expert review identified actionable findings in 23.9% of these discordant cases.

Featured Event

Why didn’t AI make a difference?

At face value, the findings are counterintuitive. If AI can identify suspicious CXRs earlier, why does this not translate into faster diagnosis? The answer likely lies in where delays occur.

Prioritising image reporting addresses only one step in a multi-stage pathway. Downstream processes—such as access to CT imaging, specialist review, and biopsy—remain unchanged. If these stages represent the true bottlenecks, earlier reporting alone may have a limited effect.

Importantly, the study was not designed to formally test which component of the pathway drives delay. However, the absence of any signal across multiple endpoints suggests that reordering radiology queues in isolation may be insufficient to shift overall timelines.

One of the more striking findings was the level of discordance between AI and radiologist interpretation.

Nearly one in three CXRs showed disagreement, and in almost a quarter of these cases, expert review identified findings considered actionable.

This raises two competing interpretations:

• AI as a safety net: The algorithm may identify subtle abnormalities missed on initial read, supporting a complementary role alongside radiologists.
• AI as a source of noise: Discordance may also introduce uncertainty, particularly if flagged findings do not trigger immediate clinical action.

In practice, the value of this signal depends on how it is used. In this study, AI outputs influenced prioritisation—but did not mandate escalation, additional imaging, or clinical review. Cases flagged as abnormal by both AI and radiologists were associated with shorter diagnostic intervals, while those missed by both experienced the longest delays. This pattern suggests that concordant detection—rather than prioritisation alone—may be more relevant to improving outcomes.

However, these findings are hypothesis-generating and should be interpreted cautiously. The study was not powered to evaluate subgroup effects, and causality cannot be inferred.

A key strength of the study is its real-world design.

Unlike many AI evaluations, which focus on retrospective accuracy in curated datasets, this trial embedded the algorithm within a routine clinical workflow. As such, it captures the complexity of implementation, including competing demands, variable reporting practices, and system-level constraints.

At the same time, this design introduces limitations:

• The relatively low incidence of lung cancer in the study population may dilute measurable effects.
• The intervention was limited to prioritisation, without integration into broader clinical decision-making.
• The healthcare setting and workflow structure may not generalise across systems with different capacities or referral pathways.

Implications for practice

The findings do not suggest that AI lacks value in thoracic imaging. Rather, they highlight a critical distinction between diagnostic performance and clinical impact.

AI tools that perform well in detecting abnormalities may still fail to improve outcomes if they are not embedded within systems that can act on those findings.

This has several practical implications:

• Workflow integration is key: AI outputs may need to trigger defined actions (e.g., same-day CT, rapid-access clinics) to influence outcomes.
• System capacity matters: Without sufficient downstream resources, earlier detection may not translate into earlier treatment.
• Evaluation must move beyond accuracy: Prospective, pathway-level studies are essential to understand real-world impact.

Delays in lung cancer diagnosis remain a persistent challenge, with outcomes closely tied to stage at presentation. While screening programmes and rapid diagnostic pathways have shown benefit, implementation remains uneven.

AI has been widely proposed as a scalable solution—particularly in primary care settings where early signs may be subtle. However, this study underscores that technology alone cannot compensate for structural limitations in care delivery.

Instead, the greatest opportunity may lie in combining AI detection with pathway redesign—ensuring that high-risk findings trigger immediate and coordinated action.

Paper: Woznitza, N., Smith, L., Rawlinson, J. et al. AI-based chest X-ray prioritization in the lung cancer diagnostic pathway: the LungIMPACT randomized controlled trial. Nat Med (2026). Access online here.