Measuring cardiac output remains central to the management of critically ill patients. For decades, this role was filled by the pulmonary artery catheter (PAC), which served as both the gold standard and a source of arrhythmias, infections, and mechanical complications, with an impact on mortality that was never fully demonstrated. Hence the long and still ongoing search for less invasive alternatives. At the same time, critical care cardiology turned its attention back to the periphery. Skin mottling and capillary refill time regained prominence as markers of perfusion, and the ANDROMEDA-SHOCK studies and their follow-up demonstrated that they can even guide resuscitation—with much of this work conducted by Latin American groups. (1-4) Once again, the skin serves as a mirror of circulation.

The work by Atamañuk and colleagues fits into this line of research, posing a question so simple that it’s surprising it hadn’t been asked be : Can the face predict low cardiac output when examined with an infrared camera and some artificial intelligence (AI)? The physiology supports this. When cardiac output drops, the sympathetic response diverts blood to vital organs and sacrifices the skin; skin temperature drops, and it does so early on. The idea is not new. As early as 1969, Joly and Weil had already shown that the temperature of the big toe indicated the severity of shock. (5) What changes here is the chosen area: the face, which is always visible in hospitalized patients, well-perfused, and easy to photograph.

The study design has strengths worth acknowledging. The thermal images were taken simultaneously with CAP measurements—that is, against a reference standard and in real time—using a standardized protocol, regional facial readings by blinded evaluators, and machine learning processing. I am not aware of any previous report that combines these three elements at once: facial thermography, concurrent invasive monitoring, and AI. The results are consistent with the hypothesis. Several regions and gradients correlated significantly with cardiac index, with the thermal gradient showing the best performance (r = –0.45), and a neural network achieved an area under the curve (AUC) of 0.75 for detecting low cardiac output. There is a signal.

That said, an editorial that merely offers praise isn’t very helpful. It’s worth pausing to consider the limitations, several of which are acknowledged by the authors. The sample size is small: 35 monitoring sessions across 18 patients, with nine of them providing repeated measurements. This dependency among observations, if not accounted for using mixed-effects models, inflates statistical significance. More fundamental is the pharmacological issue. Nearly all patients were receiving vasoactive drugs, and more than half were receiving two or more. Norepinephrine constricts the skin bed, while dobutamine dilates it; added to this are fever, sedation, mechanical ventilation, and room temperature. The face ends up reflecting this combination, not just cardiac output. Separating the hemodynamic signal from the noise will be the truly difficult part. (6)

Furthermore, discriminatory power is still modest. An AUC of 0.75 is encouraging, but a negative predictive value of 50% means that, at present, the method does not allow us to rule out low cardiac output with confidence—which is precisely what one would expect from a noninvasive monitor. And training a neural network with only a few dozen cases invites overfitting: the result is only as good as its validation—first internal and then, above all, external. The TRIPOD+AI guidelines are a good reference for reporting these models transparently. (7) The low-resolution, consumer-grade camera is sufficient for a proof of concept but not for a clinical product.

None of this invalidates the proposal; rather, it highlights what is missing (Table 1). The study is part of an exciting trend—that of digital biomarkers and contactless monitoring—where AI promises to extract useful information from noisy data, with applications ranging from telemedicine to the remote monitoring of heart failure. Thermography itself, combined with AI, has already succeeded in distinguishing acute decompensated heart failure from other conditions. (8) And there’s one detail that’s been on my mind: in some patients, thermal improvement paralleled that of the cardiac index over time. If this is confirmed, thermography would cease to be a snapshot and become a moving image—a trend monitor.

That said, we must not confuse promise with proof. Medical AI has produced brilliant models in a single center that later fail to withstand external validation. To make it to the patient’s bedside, we’ll need what we’ve always needed—which is also the most costly: larger, less heterogeneous, multicenter cohorts; prospective designs that track thermal patterns when interventions are made; statistical analysis that respects the data structure; and standardized reporting. Thermography is not going to replace echocardiography or catheterization. Its place, if it has one, will be as a physiological, inexpensive, and noninvasive adjunct.

The authors dared to propose a risky hypothesis and tested it honestly, without hiding its limitations. They showed that there is something to be gleaned from the face of a patient in shock. It remains to be seen whether that “something” can translate into a clinical decision—and that is a collective effort, the kind that cardiologists in the region are capable of when they organize themselves. For now, the patient’s face barely hints at how their heart is beating. That is no small thing, and it deserves our continued attention.