Beyond the Result | Architecting the Future of Diagnostic Intelligence, Global Scale & Precision Care

Session Focus

Walt framed how far the category has moved. Diagnostics has evolved well beyond a single test confirming a diagnosis, now spanning screening, genomic profiling, molecular testing, imaging, surveillance tools and AI that predicts, detects and tracks disease.

The panel paired a pathologist advancing novel diagnostics in the clinic, a diagnostics industry leader focused on the data foundations that make AI useful, and a cardiologist who has spent his career translating biomarkers into treatment decisions.

Data as the Precondition

Gianluca Pettiti opened with an analogy from outside medicine. Self-driving cars demonstrate how far technology, digital systems and AI have advanced. He noted as an Italian that he still likes driving his own car, while observing that nobody would now buy a car without a collision sensor, and that in a few years the same may be true of self-driving capability.

He expects convergence to move healthcare toward what he called a self-driving healthcare ecosystem, with substantial impact on quality of care, given that studies estimate 20 to 30 percent of US healthcare spending is waste.

The parallel he drew is that sensors, data and AI enabled self-driving vehicles, and the same combination will be required in healthcare. It begins with data: instrumentation and workflow creating data at scale.

He announced a large partnership with Mayo Clinic, called Pre-Cure, targeting testing of a million biospecimens, alongside close work with Mass General Brigham.

His qualification was important. Data alone is not enough, but it is the foundation that must exist: good, quality, well-annotated and organized data, so that AI can be used at scale.

A Taxonomy for AI in Diagnostics

John Iafrate MD, PhD described the current moment as exciting and also overwhelming, noting how hard it is even for someone doing technology development and using these tools clinically to maintain a global view of the landscape.

He organized AI in diagnostics into three categories.

Productivity tools, which he expects to arrive fast and furious, drawing on large language models already working well. His pathology examples were laborious tasks clinicians do not enjoy: counting mitotic activity, which a computer can be trained to do better than a human, and scoring immunohistochemistry such as HER2, which remains the gold standard in breast cancer worldwide while being genuinely difficult for pathologists to score accurately and reproducibly across sites.

Risk-based tests, which he characterized as more black box outcome predictors built on multimodal data combining clinical information with imaging and other technologies. He sees this as a very exciting area where statistical performance can keep improving, because there is almost too much in the multimodal space for human researchers to compete with what computers can now do.

Definitive diagnosis, where he was directly skeptical of the prevailing narrative.

Why Pathology and Radiology Are Not Being Replaced

Iafrate noted that pathologists hear essentially daily that replacement is a matter of time, and said the field is really not there.

On standard H&E slide diagnosis, despite impressive foundation model work, those models remain foundational and their performance is not clinical grade.

His argument about expectations was the sharpest part of the session. Patients and clinicians ordering a test expect it to be right 100 percent of the time. That is not true of any test, but it is the expectation, particularly for a layperson.

He grounded that in real validation numbers. For detecting ALK fusions in lung cancer, a subtype he worked on for many years, validation in an academic laboratory across roughly 100 known positives and negatives produced 100 percent sensitivity, 99 percent specificity and 95 percent positive predictive value. That is the standard clinicians hold certain tests to.

Asked whether AI could replicate that by predicting the same result from an H&E image, his answer was that the field is not even close at this point. His framing was deliberately balanced: AI is making huge strides in productivity and will help everyone, and the current hype that it can do everything should not be believed.

Companion Diagnostics in Cardiovascular Medicine

Paul Ridker MD contrasted his field with oncology, which has long had companion diagnostics matching a genetic alteration to a specific drug, and which has subdivided what used to be simply leukemia and lymphoma into many distinct diseases.

Cardiovascular medicine has not done that. It still has heart attacks and strokes, and has not moved meaningfully beyond LDL cholesterol as its fundamental companion measure.

The gap that motivated his career is stark: half of patients who have heart attacks and strokes have completely normal lipid levels. The question is what else is wrong, and a substantial part of the answer is the inflammatory cascade.

As an inflammation biologist he has worked on C-reactive protein, IL-6 and ICAM-1 to determine who is at risk and why, and he emphasized that the why matters more than the who.

The CANTOS trial demonstrated the principle. Using a monoclonal antibody specifically targeting interleukin-1 beta, the study achieved risk reduction comparable to aggressive PCSK9 inhibition without touching ApoB or LDL at all, working purely through the NLRP3 to IL-1 to IL-6 cascade. He called that a significant opening, because it validated an entire therapeutic direction based on biology and biomarker data identifying risk in patients who otherwise appear not to be at risk.

What Has and Has Not Worked

Ridker was candid that not everything in the pathway works, and gave the current scorecard.

Three positive trials exist in the inflammation space: canakinumab, and two for colchicine, a very inexpensive drug he noted almost nobody uses because there is no marketing behind it, despite being very effective as an anti-inflammatory in that setting.

Against that, a recent trial of an IL-6 inhibitor was not effective, with data to be presented shortly.

On lipoprotein(a), commonly measured and genetically mediated, he recommends measuring once to identify families at high risk who do not otherwise appear to be. But the only information in the public domain, ahead of a November presentation, is that the first trial of a drug lowering it appears to have been neutral.

His summary of the field’s position: cardiovascular medicine is trying to move past LDL and ApoB, where interventions reliably work, into a space where other targets may or may not, and it is tricky.

Key Takeaways

1. Data foundations precede AI value. Well-annotated, organized, quality data at scale is the prerequisite, not the model.

2. 20 to 30 percent of US healthcare spending is estimated waste, which is the size of the opportunity.

3. Three distinct categories of diagnostic AI carry different evidence bars: productivity tools, risk prediction and definitive diagnosis.

4. Foundation models are not clinical grade for primary diagnosis. Validated molecular assays reach 100 percent sensitivity and 99 percent specificity, and image-based prediction is not close.

5. Half of heart attacks and strokes occur in patients with normal lipids, which is the gap inflammation biomarkers address.

6. Inflammation targeting works selectively. Three positive trials, one negative IL-6 trial, and an apparently neutral first lipoprotein(a) outcome trial.