Insights from CODEX Director Sumant Ranji, MD, SFHM: "AI might get you a quicker diagnosis, but it could delay everyone else’s."

AI might get you a quicker diagnosis, but it could delay everyone else’s.

How could that be possible? We have ample evidence that large language models can diagnose as accurately as clinicians in simulated settings; many (maybe most) physicians are using some form of AI to help with diagnosis; and AI-based diagnostic applications are widely used in image-based fields, with studies of EHR-embedded LLMs in real-world settings likely coming soon. With all that momentum and a growing amount of evidence, how could AI actually worsen diagnosis?

The answer lies in an important recent study that had nothing to do with AI. Last month in JAMA, Sean Mann and colleagues at RAND examined the effect of a large-scale randomized trial of a cancer screening program using multi-cancer early detection testing (MCED) testing in the United Kingdom’s National Health Service. MCED tests look for residual cancer DNA in the bloodstream and may help diagnose certain types of cancer at an earlier (and potentially more treatable) stage. The MCED trial was conducted in eight NHS regions, in which nearly 2 million patients underwent workup for suspected cancer (of the head and neck, lung, or upper GI tract) during the trial period. The study did not assess the effectiveness of the MCED test itself – rather, the investigators examined the overall rate of delayed diagnosis of these cancers in the NHS regions who participated in the trial. The rates of delayed cancer diagnosis in the trial regions were compared to delayed diagnosis rates in the 13 NHS districts that did not participate in the expanded screening trial. By doing so, the investigators tried to determine how expanding access to a new diagnostic test (MCED) affected diagnostic timeliness for all patients, whether they had access to the test or not.

Here is what they found, using sophisticated statistical methods that I won’t attempt to explain: patients in the MCED regions were more likely to experience a delayed cancer diagnosis in the first year after MCED was implemented, compared to regions that did not have access to MCED testing. This finding held true after adjustment for various health system factors, including staffing and bed capacity amid the COVID-19 pandemic.

This seems counterintuitive – how could more testing for cancer result in more delayed diagnoses? The answer lies in capacity constraints. Multi-cancer early detection tests can provide a signal that a patient may have lung cancer, for example, but the patient is still going to need further testing (such as CT scans), evaluation by a specialist, and ultimately a biopsy. But CT scanners, pulmonary specialists, and interventional radiologists are a fixed resource. Identifying more patients that might have cancer doesn’t change the fact that all of them will need definitive testing to figure out if they do or do not have cancer. (In fact, more than 90 percent of patients referred for cancer workups in the UK actually do not have cancer.) Without more testing or biopsy availability, the patients whose MCED tests indicated possible cancer just wait in a longer line – a line which includes all the patients who also might have cancer and are being worked up through traditional diagnostic pathways. The result: more delayed diagnoses, due to constrained system diagnostic capacity.

I’m afraid that widespread use of AI for diagnosis – by clinicians and patients – could result in the same problem. We don’t know if AI-assisted diagnosis results in more diagnostic testing, but it seems quite plausible (in my own clinical experience, " That additional testing probably will speed up the diagnostic process for some patients, but the overall effect could just add more strain to an already beleaguered system. This concern is also true for consumer-directed diagnostics like full-body MRI or MCED, and novel technologies like Midjourney Medicine’s full-body ultrasound scanner – any abnormal findings will lead to standard diagnostic imaging or invasive tests, but who will do those additional biopsies or read the additional PET scans? Patients and their doctors may think they are diagnosing quicker, but they’ll find they’re just queuing for longer.

I’m cautiously optimistic about the possibilities for diagnostic AI, and I use it regularly in my own practice. But the JAMA paper should remind us that the outcome of every clinical encounter depends in large part on the very flawed health care system clinicians work in and patients are subjected to. Using AI might advance the diagnostic process for an individual patient – at the cost of impeding diagnosis for everyone.

__

View the post and join the conversation on LinkedIn. 

__

Mann S, Nascimento de Lima P, Eagan J, Ulyte A, Griffin BA. Cancer Diagnostic Delay Rates Associated With a Population-Based Screening Trial Evaluating a Cell-Free DNA Multicancer Early Detection Test. JAMA. Published online May 30, 2026. doi:10.1001/jama.2026.6803