Biograph scanned a group of asymptomatic, low-risk adults and found elevated coronary inflammation in 42.9% of them. Cholesterol markers missed it. HbA1c missed it. hs-CRP missed it. The AHA PREVENT-ASCVD risk calculator missed it. What found the signal was Caristo Diagnostics’ CaRi-Heart technology, running AI analysis on the exact same CT scans those patients were already getting. No new test, no new hardware, no extra radiation. Just a different read of data that was already sitting there.

That pattern should sound familiar if you run a plant or a quality function. You are already collecting far more data than you analyse: inspection images, sensor logs, batch records, maintenance tickets. This article breaks down what the Biograph study actually proves about AI on existing data, and how to spot the same opportunity in your own operation.

Every Standard Test Said These Patients Were Fine

Nearly half of the participants with no detectable plaque at all still showed elevated perivascular inflammation. 17.9% had very high inflammation in at least one major artery. These were people a cardiologist would have sent home. The inflammation was real, measurable, and sitting in imaging data that had already been captured and stored.

What makes this uncomfortable is not the diagnostic miss. It is that nothing was broken. The bloodwork ran correctly. The risk calculator computed what it was designed to compute. Every process worked as specified and still produced a wrong answer, because none of those instruments were pointed at the tissue where the problem lived.

Your inspection images, vibration logs, and audit records have the same property. They are complete, they are retained, and almost nobody is extracting a second answer from them.

What Biograph and Caristo Actually Measured

The study setup was straightforward. Biograph ran high-resolution coronary computed tomography angiography (CCTA) plus cardiometabolic profiling on clinically eligible Biograph Black tier members, then fed the imaging through Caristo Diagnostics’ CaRi-Heart and CaRi-Plaque technologies. Two layers: the scan itself, and an algorithmic read of what the scan contained.

No participant did anything extra. The CCTA was already part of the protocol. Caristo’s contribution was analysis, not acquisition, which is the whole reason this study is worth your attention as an operations problem rather than a medical one.

The numbers: 42.9% elevated, 17.9% very high

42.9% of participants showed elevated coronary inflammation. Within that group, 17.9% had very high inflammation in at least one major heart artery. That is close to one in five people flagged at the top of the scale, in a population selected for being low risk.

The figure that should bother you most is a different one. Roughly 47% of participants with no detectable coronary plaque still showed elevated perivascular inflammation. Nearly half the people who passed the structural check were carrying a signal the structural check was never designed to see. The test was not wrong. It was answering a narrower question than anyone assumed.

Reading the tissue around the artery, not the blockage

Conventional cardiac CT analysis looks at the artery: how much plaque, how much narrowing, how bad is the blockage. CaRi-Heart looks at the fat tissue surrounding the coronary arteries and reads its biology. Inflamed vessel walls change the composition of that surrounding tissue, and those changes show up in CT attenuation patterns a human reader is not going to eyeball reliably.

So the AI is not finding a better blockage. It is extracting a completely different variable from the same pixels, one that sits upstream of the thing everyone was already measuring. The scan contained both signals the entire time. Only one of them was ever being read.

The Real Lesson: AI Found Signal in Data Already Being Collected

The tissue around the coronary arteries was in every one of those scans. Radiologists looked past it because it wasn’t what they were trained to read. Caristo didn’t add a sensor, it added a variable derived from pixels that were already captured, stored, and functionally thrown away.

That distinction matters more than the clinical result. A new measurement usually means new capital, new validation, new training. A new read of an existing measurement means software and a labelled dataset.

Why new sensors are rarely the bottleneck

Most plants are not short on data. They are short on anyone reading it. Ask how many of your installed sensors feed a model, a trend, or even a chart someone actually opens. The honest answer is usually a small fraction.

Adding instrumentation feels like progress because it has a line item and a delivery date. It also adds integration work, calibration schedules, and more unread data. If a signal already exists in your historian and nobody has looked for it, buying another sensor does not fix the underlying problem, which is analysis capacity.

Start where acquisition is already paid for. The payback window on an analytical layer over existing data is measured in months, not capital cycles.

What ‘unread data’ looks like in a plant or lab

• CMM output: pass/fail gets recorded, the full point cloud and deviation pattern usually does not get mined for tool wear signatures.
• Vision inspection frames: images are captured, judged against a threshold, then deleted or archived without ever being clustered by defect type.
• Thermal imaging: used for spot checks, almost never trended across shifts or compared batch to batch.
• Machine telemetry: vibration, current draw, and cycle time sit in the PLC or historian at full resolution and get summarised into averages.
• CAPA and complaint records: years of free-text root cause narratives that nobody has ever read as a corpus.

Each of these is a scan with unexamined pixels. The signal is sitting in storage you are already paying for, waiting on someone to ask a different question of it.

Where This Kind of Detection Wins, and Where It Doesn’t Yet

The strengths here are real. The analysis is non-invasive, it sits upstream of irreversible damage, and it is independent of markers that demonstrably failed to correlate with anything. That last point is the one worth stealing. A signal that correlates with your existing metrics tells you nothing new. A signal that moves independently of them is where the value lives.

The limits are equally real, and the study authors say so themselves. The cohort was drawn from a premium preventive clinic, which is about as self-selected as a population gets. And the paper’s own language is cautious: “More research is needed to validate this finding.”

Detection without a response protocol is just anxiety

Suppose you find elevated perivascular inflammation in a patient with clean coronary anatomy. Then what? There is no established pathway, no agreed intervention threshold, no outcome data showing that acting on the finding changes the trajectory. The detection is ahead of the medicine.

The same trap kills manufacturing AI pilots. You deploy a model that flags drift in a bonding process three shifts before the defect appears, and the line supervisor has no authority, no procedure, and no spare capacity to act on it. The alert becomes noise within a month. Define the response before you deploy the detector, or you have built an expensive worry machine.

Validation is the gate, not the model

Getting a model to produce a number is the easy part. Proving the number means something, consistently, across operators, shifts, product variants, and seasons, is the part that takes six months and kills most projects. Biograph and Caristo ran a study. That is the right order of operations.

Treat your first deployment as a study, not a rollout. Run the model in shadow mode against known outcomes. Measure false positives honestly, because an inspection team that stops trusting the flag is worse than no flag at all. Scale after the evidence, not before.

The Pattern Worth Copying Before Your Competitors Do

For a decade, most industrial AI spend went into automating things people already did: sorting, counting, flagging known defect types. That work pays, but it is bounded. The ceiling is whatever a trained human could do, faster.

The shift underway is different. AI is now being pointed at measurements nobody was reading, to produce variables that did not previously exist. Caristo’s work on perivascular tissue is one version of that. Your version is sitting in a historian database nobody has queried since the integrator left.

Three questions to ask about your existing data

• What do we capture and never look at?: Vibration traces, thermal images, vision system frames that passed inspection, torque curves, batch logs. Anything retained for compliance but never analysed is a candidate.
• Where do our current metrics have known blind spots?: Ask your quality engineers which failures surprise them. Surprise means your instrumentation is measuring the wrong thing, or the right thing too late.
• Can we test the hypothesis on history before buying anything?: If you have twelve months of stored data and a record of what failed, you have a labelled dataset. That is a pilot, not a capital request.

Run those in order. Most teams skip straight to vendor demos, which is how you end up with new sensors measuring the same things you already measured badly. The expensive mistake is adding acquisition when the gap is analysis.

The ROI logic is the same one the Biograph study rests on. Elevated inflammation appeared in 42.9% of people with no symptoms, and roughly 47% of participants with no detectable plaque still showed it. Catching the upstream condition costs a consult. Managing the downstream consequence costs a stent, or in your case a recall.

Field failures, scrap, unplanned downtime: all late consequences of a process that drifted while your dashboard stayed green. The data showing the drift was almost certainly recorded. Go read it.

Source: biograph.com