“Incidentalomas” in the Age of Artificial Intelligence

The Hippocratic principle of “First, do no harm” applies just as much to diagnostic tests as to any side-effect prone drug or surgery. Yet, the sheer volume of data produced by modern diagnostics makes abiding by this principle ever more difficult. While physicians might well be parsimonious in choosing individual laboratory tests to maximize benefit and minimize harm, data-rich diagnostics such as MRIs or whole genome sequencing have proven more challenging.

“Incidentaloma” refers to an unsought after finding of uncertain clinical significance, for example a renal cyst on a hepatic ultrasound. As imaging frequency and complexity increase¹, so have the likelihood of these incidental findings. Different imaging modalities produce them at widely different rates, often without clear consensus on appropriate management². While they sometimes lead to earlier diagnosis and treatment, incidental findings have historically had low diagnostic yields³. Misguided subsequent testing may even harm patients due to unnecessary invasive procedures like biopsies.

The response of the medical community has mostly been focused on reducing unnecessary tests. While laudable, this ignores the basic challenge—some incidental findings are helpful, most are useless, and a few are harmful. Rather than accept this uncertainty, we should confront it directly and create actionable solutions that empower more precise care.

The same advanced information technology that has produced this flood of data may hold the key to channeling it towards clinical insight and patient benefit. Leveraged artificial intelligence (AI) already improves diagnostic imaging insights on intended clinical questions^{4, 5}. Recent studies have applied AI to separate actionable incidental findings from those with little clinical import⁶. Two common incidental findings on non-ECG-gated, non-contrast, chest computed tomography (CT) scans, pulmonary nodules and coronary artery calcium (CAC), serve as examples for further investigation.

Pulmonary nodules are incidentally found in up to 53% of CT scans,⁷ with a wide range of benign and malignant etiologies⁸. The Fleischner Society Guidelines recommend follow-up intervals to ascertain growth based on nodule-specific factors like size and underlying probability of disease⁹. Still only half of patients receive guideline-aligned care⁹. As many as 150,000 individuals get tests with no yield on the one hand, and 3000 have harmful diagnoses missed annually⁹. AI predictive models could help improve these results. Consider how an algorithm trained and validated on both imaging data and the electronic medical record might both identify pulmonary nodules and better estimate their malignant potential. With these improved models, we can more sensitively empower clinicians and patients’ shared decision-making on how much follow-up is needed.

Coronary artery calcium can also be measured on CT scans, but unlike pulmonary nodules, is not routinely noted even when present¹⁰. CAC is documented in atherosclerosis and is a strong predictor of subsequent heart disease in diverse groups. It adds information to traditional risk factors, with guidelines recommending considering treatment when present^11–13. While intentional CAC scanning uses cardiac gating, incidental CAC seen on routine chest CTs has similar prognostic significance¹⁴. AI-based incidental CAC identification already exists¹⁵ and the risks of potential treatments are minimal and have well-established safety profiles. Because CAC-specific scans are less likely to be ordered for those with worse health care coverage, algorithms may actually improve equity of atherosclerotic detection¹⁶ and provide more precise care to at-risk individuals.

These examples illustrate how future research could help AI to fulfill its promise in separating the proverbial wheat from the chaff of incidentalomas. That research should focus not just on predictive effectiveness, but also on how the AI will be implemented. Just because we can predict future harm does not mean we have effective interventions to prevent it. Even if effective, AI implementation may be limited by (1) cost, including that of downstream testing; (2) data flow and workflow constraints impeding widescale uptake; (3) concerns about patient data sharing and skepticisms of a newly validated tool; and (4) uncertainty regarding performance in diverse populations.

The history of incidentalomas is a checkered one at best. As in other areas of modern life and medicine, AI has enormous potential to change that history. It can help patients and doctors harvest the most benefit from the onslaught of incidental data generated by diagnostic testing from imaging to wearables to genetic testing. And, like in other areas, we must first apply a skeptical scientific lens to ensure that this AI bounty does not itself come with unintended harmful consequences.