AI-Enabled Opportunistic Screening

Session Focus

Singal, a computer scientist who trained in AI before becoming a physician and spending a decade at a cancer diagnostics company, opened with a deliberately basic question: is under-diagnosis actually a problem worth solving? The panel worked through the scale of the gap in oncology, cardiology and chronic inflammatory disease, what has changed technically to make it addressable, and the operational barriers that remain on the far side of a flag.

The Scale of Under-Diagnosis in Cancer

Bea Bakshi MD, a primary care physician by background, framed the structural problem. Most patients are diagnosed through the physician front door rather than through a screening program. Screening picks up roughly 5 to 15 percent of cancers across the Western world.

The specifics behind that: between 86 and 95 percent of cancers are diagnosed through primary care and first contact settings, where more than 100 cancer types present. Screening programs cover three to five cancers, meaning roughly two-thirds of cancers have no screening program at all.

She pointed to the pattern that makes this actionable. About a third of lung cancers show sequential visits where the signal could have been recognized earlier, and the same holds for pancreatic and other hard-to-diagnose cancers.

Her diagnosis of the cause was not physician failure but arithmetic. A clinician has 10 to 15 minutes to triage a complex disease with complex presentations. The opportunity for AI is to compute the complexity sitting in years of medical records and surface it inside that appointment, so the physician can recognize that a pancreatic cancer is in front of them and route the patient correctly.

Pattern Recognition in Cardiology

Alexandra Gonçalves MD, PhD, a cardiologist who spent more than a decade in practice and cardiovascular imaging before moving to industry, described work spanning common and rare conditions.

On the common end, her company began an atrial fibrillation detection program more than six years ago, work she characterized as the early days of what consumer devices have since made familiar.

The more interesting development is what AI extracts from the electrocardiogram. The technology recognizes patterns a human eye does not, including patterns that expert readers cannot detect, which opens a different category of screening.

Her team has deployed AI pattern recognition for hypertrophic cardiomyopathy at scale, with more than 100 US hospitals now having access. Any patient receiving an ECG at an implemented site gets that level of triage.

She framed the larger ambition around risk rather than diagnosis. Every ECG could indicate both what a patient has now and what they are at risk for in future, which is where personalized medicine becomes risk adjustment for an individual at a moment in time.

Redefining the Gap Beyond Diagnosis

Nandini Nayar argued the under-diagnosis framing is too narrow. There is an equally large problem of longitudinal control assessment, where patients are correctly diagnosed but cycle for years through suboptimal therapies, often suffering in silence.

The cause is structural. Care models are episodic and reactive, and a 10-minute appointment does not get to the heart of managing a chronic inflammatory disease.

Her program flags patients who are diagnosed but uncontrolled and therefore at high risk, drawing signals from the EHR and elsewhere. She was emphatic about the second half: you do not just send off a bunch of flags. The program builds an end-to-end care pathway with triage so a flagged patient actually sees someone rather than being told to wait six months.

What Changed Technically

Faisal Mahmood PhD located the shift in what models can now represent.

Multi-modal learning has been underway for a decade, but recent capability allows very high-capacity models that combine representations from many modalities into a patient-level representation. His group’s work produces a single vector representing a patient at a given moment, one that continuously evolves. The model is called Apollo, and other groups are developing along similar lines with likely convergence toward very large models in the space.

The consequence is that the patient journey becomes computable. Once that representation exists, it can be channeled toward many downstream clinical tasks, including diagnosing disease early, predicting risk and screening.

Asked what specifically makes this more solvable than three or five years ago, he contrasted it with supervised learning. The previous approach defined a problem, gathered data, built a cohort and supervised it with available labels. That never captured the whole patient and confined each model to the single disease it was built for.

The same class of technology behind large language models can build models that represent data extremely well. Computing on top of those representations is far easier, which makes it newly practical to build exhaustive models that screen for everything at once.

False Positives That Are Not

Gonçalves offered the most instructive deployment story. Her team’s hypertrophic cardiomyopathy program drew complaints about false positives until they examined the flagged patients more closely.

Those patients did not have hypertrophic cardiomyopathy. Many of them had aortic stenosis or hypertension-related hypertrophy. They were not false positives for having a heart problem.

The episode illustrates a definitional problem in evaluating these systems: a model trained for one condition may be surfacing genuine pathology outside its label.

The Downstream Bottleneck

The recurring concern across the panel was what happens after the flag.

Mahmood argued the binding constraint is now how risk information reaches physicians and patients without every flag triggering a referral. The detection capability is ahead of the communication and triage design.

Nayar named the same gap from the industry side. Models identify high-risk patients reliably, but downstream capacity to act on them has not kept pace.

Key Takeaways

1. Screening reaches a small fraction of cancers. Roughly two-thirds of cancer types have no screening program, and most diagnoses come through primary care.

2. The constraint is appointment arithmetic, not physician skill. Years of record complexity cannot be triaged in ten minutes without help.

3. AI reads ECG patterns experts cannot. That opens rare disease screening on a test already performed at scale.

4. Under-diagnosis is only half the gap. Diagnosed but uncontrolled patients represent a comparable population.

5. Patient-level representations changed what is possible. A single evolving vector supports screening for many conditions at once rather than one model per disease.

6. A flag without a pathway is not a solution. Downstream triage capacity is the current bottleneck across every program discussed.