Cardiology's averages hide the patients who benefit and the risk they miss

The average patient is disappearing

In a preprint, researchers re-examined the SAVE trial, whose overall result on CPAP therapy for obstructive sleep apnoea had been neutral, by applying machine learning to 2,676 participants. Gaussian mixture models and causal survival forest analysis identified five sleep apnoea prototypes with different cardiovascular risk profiles and different predicted responses to CPAP. Only one prototype, Prototype 5, was predicted to benefit, while the others showed no benefit or potential harm. m:…9448

The implication the authors draw is about trial design as much as treatment: a neutral average can hide a subgroup that benefits and subgroups that do not, and precision phenotyping could improve both trial enrichment and clinical decisions about cardiovascular risk reduction in sleep apnoea patients. The result is a re-analysis of existing trial data, not a new randomised trial, so it points at who should be studied next rather than settling who should be treated. m:…9448

Three ways the average misleads

The same pattern appears in imaging. A large registry study of 35,391 patients undergoing PET myocardial perfusion imaging finds that regional myocardial flow reserve adds prognostic value for major adverse cardiac events beyond conventional global flow reserve and stress total perfusion deficit. The refinement matters most for patients who look low-risk on the global measure but carry regional risk that the standard summary hides. m:…9048

It appears again in wearable data. In 8,292 participants from the All of Us Research Program, higher wearable-derived daily step counts were associated with lower odds of a first recorded cardiovascular diagnosis, but the association was stronger for acute-care recordings than for outpatient ones. The authors caution that researchers and clinicians must account for clinical setting and recording completeness when interpreting wearable data linked to electronic health records, because where a diagnosis is recorded changes what the numbers mean. m:…2270

Across the three preprints, the failure mode is the same: a single summary number, whether a trial average, a global perfusion score, or a step-count association, smooths over structure that changes the decision. The useful signal lives one level down, in the subgroup, the region, or the recording setting. m:…9448 m:…9048 m:…2270

What changes in practice

None of these preprints changes a guideline on its own. The sleep apnoea prototypes need prospective validation before CPAP decisions follow them, the regional flow reserve finding needs to show that acting on regional risk improves outcomes rather than only predicting them, and the wearable result is observational, so steps may mark healthier people rather than cause the difference. m:…9448 m:…9048 m:…2270

What they do change is what a careful reader should ask of the next cardiology result: which patients, which region, and which setting. A result reported only as an average, without those splits, is now the result to be suspicious of. m:…9448 m:…9048 m:…2270

For a reader of the next cardiology preprint, the three studies suggest a short checklist: does the trial report subgroups or only an average, does the imaging score report regions or only a global value, and does the wearable result report the setting in which each diagnosis was recorded. A study that answers all three is making a stronger claim than one that reports a single polished number. m:…9448 m:…9048 m:…2270