The coronary artery calcium (CAC) score is the most widely used imaging metric for cardiovascular risk stratification, and a score of zero has long been treated as an approximate 'low-risk pass.' However, the paper opens by stating a fact: the population holding this 'pass' is so large that the absolute number of cardiovascular events occurring within it is not small. The root cause lies in the fact that calcification measures historical plaque deposition, whereas inflammation reflects what is actively happening in the vessel wall right now. On November 11, 2026, JAMA Cardiology published online—ahead of print—a cohort study by the ORFAN Consortium: using a novel AI algorithm to measure the 'fat attenuation index' (FAI), an imaging biomarker quantifying coronary inflammation, on non-contrast cardiac CT (NCCT) scans without iodinated contrast agents—and validating its ability to stratify risk for cardiovascular death in an independent external cohort of 19,110 individuals. The study was led by the University of Oxford, with authors spanning Harvard Medical School, Mayo Clinic, Cleveland Clinic, and University College London.
## How the Validation in 19,110 Individuals Was Designed
The design follows a standard 'derivation–external validation' structure. The population comprised consecutive patients who underwent both coronary CT angiography (CCTA) and non-contrast calcium scoring for clinical indications between 2010 and 2021. It was multicenter and longitudinal, with a median follow-up of 6.7 years (IQR 4.6–10.0). The derivation cohort included 6,818 individuals (median age 61 years; 52.0% female); the external validation outcome cohort included 19,110 individuals (median age 56 years; 48.6% female). During follow-up, there were 536 cardiovascular deaths and 1,727 major adverse cardiovascular events (MACE: myocardial infarction, new-onset heart failure, or cardiovascular death).
The algorithm measured the FAI score on non-contrast CT (NC FAI, a novel biomarker), directly compared against the already-established FAI measured on contrast-enhanced CT (CCTA FAI, an existing biomarker). This design simultaneously addresses two questions: Can coronary inflammation stratify cardiovascular risk independently of the CAC score? And can this information be extracted from non-contrast CT images—i.e., without requiring contrast agents?
## Core Findings: The Strongest Association Occurs Precisely Where the CAC Score Is Blind
First, the FAI score independently stratified cardiovascular death risk across every CAC stratum—and showed the strongest association specifically among those with CAC = 0, exactly where the CAC score provides no information.
Second, adding either CCTA FAI or NC FAI to the CAC score significantly improved risk discrimination: the AUC increased from 0.76 to 0.81 (CCTA FAI) and 0.80 (NC FAI), respectively (P < 0.001); improvements in reclassification and net benefit held across all risk thresholds.
Third, examining only the 'low-risk' group: Among the 89 cardiovascular deaths in the CAC = 0 subgroup, 44.9% had NC FAI values above the 75th percentile; among the 138 deaths in the CAC 1–99 subgroup, this proportion was 44.2%. Within the CAC = 0 group, individuals with NC FAI above the 75th percentile had a cardiovascular death rate of 4.0 per 1,000 person-years, versus 1.4 per 1,000 person-years for the rest—an approximately 2.9-fold increase (calculated herein); the MACE rate rose from 5.9 to 13.3 per 1,000 person-years, an approximately 2.3-fold increase (calculated herein). Importantly, we must clarify: the absolute risk remains low in the CAC = 0 group (1.4–4.0 events per 1,000 person-years); this 'risk enhancement' is relative in nature.
Fourth, the prognostic performance of NC FAI was comparable to that of CCTA FAI, which requires contrast enhancement. This implies that any routine non-contrast CT scan performed for calcium scoring could potentially be analyzed by AI to extract inflammatory risk—without additional scanning or contrast agents. This is our interpretation: the original text does not use the word 'conveniently,' but the fact that both scans were performed on the same individuals and the two FAI metrics performed comparably is explicitly stated in the study design and conclusions.
## How the Authors Position It: A Risk Enhancer, Not a Diagnostic Certificate
The authors’ framing of the conclusion is measured: an elevated FAI should be interpreted as a 'risk enhancer,' prompting intensified implementation of evidence-based preventive measures. The paper further states that FAI has the potential to serve as a 'gatekeeper' for future targeted anti-inflammatory therapies. Note the phrasing: the original text says 'a potential future gatekeeper'—this is forward-looking speculation, not current evidence. No trial to date has demonstrated that initiating anti-inflammatory therapy guided by FAI improves clinical outcomes.
## Parallel Case: When an ECG Begins Speaking for a Lab Test
In the same month’s Journal of Cardiac Failure, the Houston Methodist team demonstrated another instance of this paradigm: using a modified VGG16 convolutional neural network to directly classify B-type natriuretic peptide (BNP) and NT-proBNP levels from single-lead (Lead I) electrocardiograms compatible with wearable devices. The model was trained to categorize BNP into three clinically relevant tiers: ≥100, ≥400, and ≥900 pg/mL. These are key blood biomarkers for diagnosing and managing heart failure—typically requiring venipuncture. This work represents another line of evidence for using AI to identify high-risk heart failure populations from ECGs—and its randomized trial has been covered elsewhere. Here, we focus on a deeper layer—not identifying patients, but inferring the numerical value of a blood test directly from an ECG image.
The scale is substantial: 528,654 Lead I ECGs (from 237,686 patients, each paired with a BNP measurement taken within 24 hours), split into training, validation, and test sets by patient to prevent data leakage; external testing used the MIMIC-IV-ECG database (18,549 patients, 42,377 ECGs), with comparisons against NT-proBNP thresholds of ≥125/≥450/≥900 pg/mL. Results: AUROC for BNP classification across the three thresholds ranged from 0.84 to 0.86; for NT-proBNP in external testing, 0.79–0.85; sensitivity and specificity for severe elevation (≥900 pg/mL) were 0.87 and 0.70, respectively; overall accuracy distinguishing normal (<100 pg/mL) from elevated levels was 0.76; performance remained consistent across sex and age groups. Prognostic associations were also observed: among 706,715 patients, higher baseline ECG-inferred BNP levels correlated more strongly with incident heart failure diagnosis over up to 10 years of follow-up—and the closer the baseline ECG was temporally to the eventual diagnosis, the stronger the association.
Limitations are evident in the numbers: an overall accuracy of 0.76 for distinguishing normal from elevated BNP is moderate; at present, this tool is a candidate for screening and risk stratification—not a replacement for venous blood testing. The prognostic analysis is retrospective. The authors themselves conclude only that the findings 'support further research.' Additionally, the abstract of this paper lacks a conflict-of-interest statement in its cached version; this article makes no assertions regarding that omission.
Together, these two studies point toward a shared direction—this is our synthesis, not a claim made in either original paper: medical AI is evolving from 'replacing the human eye in image interpretation' to 'extracting biologically invisible information—such as inflammation, molecular biomarkers, and future risk—from routine clinical signals.'
## Governance Reading: Interpreting Effect Sizes Against the Conflict-of-Interest Landscape
The conflict-of-interest disclosures in the ORFAN study merit close, layered scrutiny: Caristo-related disclosures alone span over a dozen authors. First layer: commercial entities are directly present—the commercial company Caristo Diagnostics (Oxford) appears in the affiliations of multiple authors. That is, company employees are co-authors of this paper. Second layer: patent structure—multiple granted and pending patents have been licensed from the University of Oxford to Caristo. Third layer: senior author roles—several senior Oxford team members are founders or shareholders of Caristo; multiple senior Harvard-affiliated authors serve as scientific advisors to Caristo (one unpaid); other authors disclosed advisory fees, personal payments, stock options, and patent royalties from Caristo.
This does not invalidate the study itself—its independent external validation, multicenter prospective cohort design, and median 6.7-year follow-up are robust methodological facts; the paper underwent peer review and was published in JAMA Cardiology. But it exemplifies how mature readers should approach such literature: effect sizes must be read against the conflict-of-interest map. For biomarker studies deeply embedded with commercial interests, the conclusion that 'the new biomarker is superior' inherently demands independent replication and prospective trials for balance. The transparency of disclosure itself is exemplary. The issue has never been whether conflicts exist—but whether readers incorporate them into their assessment of evidentiary weight.
## My Three Judgments
These are synthesized judgments drawn from reading both studies side-by-side—not conclusions stated in either original paper.
Judgment One: FAI’s value lies in 'filling blind spots,' not replacing the CAC score. Its strongest association occurs precisely at CAC = 0—indicating a complementary, not competitive, relationship. Yet three caveats must be read together: this is an observational cohort study—association does not imply intervention benefit; the 75th percentile is a distribution-based threshold, lacking standardized, cross-population cut points; and the absolute risk remains low in the CAC = 0 group—fold increases occur on a low baseline.
Judgment Two: The engineering leverage of non-contrast reading may outweigh its statistical incremental gain. The AUC improvement—from 0.76 to 0.80–0.81—represents a modest increment of 0.04–0.05 (calculated herein), not striking. Yet 'no additional scan, no contrast agent' means marginal cost approaches zero: every existing non-contrast CT performed for calcium scoring becomes a potential platform for inflammation assessment. This is our inference: for population-level screening programs, the cost of acquiring incremental information may matter more for feasibility than the magnitude of the increment itself. The ECG-to-BNP study follows the same logic—the value of single-lead ECG lies in its wearable-grade scalability.
Judgment Three: The 'gatekeeper' hypothesis now has a clearly defined trial target. If elevated FAI truly marks residual inflammatory risk, the next step is not further observational validation—but randomizing FAI-high-risk individuals to targeted anti-inflammatory therapy versus standard prevention, and measuring hard cardiovascular endpoints. Until then, 'gatekeeper' remains confined to the 'future outlook' section. The original paper is internally consistent on this point—it explicitly states 'potential future.'
## What You Can Take Away
If you are a cardiologist or work in a health screening center: CAC = 0 no longer automatically equals 'all clear'—if AI-derived inflammation readings are elevated on the same non-contrast CT, this provides imaging-supported justification to intensify existing primary prevention measures, without requiring additional tests.
If you are a general practitioner: when explaining 'low calcium score' to patients, you may acknowledge residual inflammatory risk—but also honestly state that absolute risk remains low, avoiding unnecessary anxiety.
If you are a payer or guideline developer: the cost-effectiveness of incorporating FAI as a 'risk enhancer' into decision pathways depends on whether its reclassification of preventive medication eligibility translates into reduced event rates—and that requires prospective evidence, which currently does not exist.
If you are a researcher: the gatekeeper hypothesis offers a ready-made enrollment tool for anti-inflammatory trials; the ECG-to-molecular-biomarker direction demonstrates a scalable research framework—'extracting invisible biology from routine signals.'
If you are a general reader: the next time you see a headline like 'AI Sees Inflammation in Non-Contrast CT,' ask three questions—does the study demonstrate association or benefit? What is the commercial background of the research? And just how large is the absolute risk? Answers to these three questions usually matter more than the AUC number itself. Feel free to share in the comments: Have you or a family member undergone coronary artery calcium scoring? Did the report mention 'inflammation'?
Sources: JAMA Cardiology, online ahead of print on November 11, 2026 (PMID 43013289, https://doi.org/10.1001/jamacardio.2026.5006); Journal of Cardiac Failure, November 2026 (PMID 43009245).