A Sixty-Second Morning Reading Before the Crash: What It Can and Cannot Tell You

Start with the number that ought to give any digital-biomarker claim pause. In this study, a model that knew nothing about a person except whether they had crashed the day before could already tell good evenings from bad ones with an area under the receiver-operating curve of 0.78 — where 0.5 is a coin toss and 1.0 is perfect. Yesterday, in other words, is a strong forecast of today. The interesting question is what a morning measurement adds on top of that, and for people living with Long COVID, the stakes behind it are not academic.

What they fear is the crash: a good stretch broken by a sudden, days-long collapse after exertion that felt harmless at the time. Clinicians call it post-exertional symptom exacerbation. It makes ordinary planning impossible, because the bill for activity arrives hours or a day late. So the humane version of the question is whether a sixty-second reading from a phone camera each morning could warn someone that today is a day to hold back.

Aitken and colleagues, writing in npj Digital Medicine, brought unusually rich data to it: 4,244 people using a symptom-tracking app for Long COVID, myalgic encephalomyelitis / chronic fatigue syndrome (ME/CFS) and other energy-limiting conditions, contributing on average 125 mornings of data each. Every morning, a sixty-second photoplethysmography reading — the optical pulse signal a phone or wearable derives from skin blood flow — captured resting heart rate and heart-rate variability (HRV, the beat-to-beat variation that tracks autonomic balance). Every evening, people rated crash, fatigue and brain fog.

The distinction that decides everything

This is a retrospective observational cohort assembled from real-world app use between 2022 and 2024, not a trial; no one was asked to act on a reading. Its care sits in one analytic choice. The authors separated between-person from within-person effects. The between-person question asks whether people with lower HRV are, on average, sicker than people with higher HRV — a comparison contaminated by everything that differs between two human beings. The within-person question is the only one that could ever help an individual: on the mornings when your own HRV runs below your own usual level, are you more likely to crash that evening? They modelled that signal and then tested prediction by walk-forward validation, training on each person's past to forecast their own future rather than borrowing strength across strangers.

A real signal, and a small one

The within-person association is consistent and points the expected way. On a person's own worse-than-baseline mornings — higher resting heart rate, lower HRV — crash, fatigue and brain fog were all more likely that evening, across all three symptoms. As evidence that the body registers autonomic strain before the person consciously does, this is a genuine, carefully analysed finding.

Then the size of it. Against the 0.78 that yesterday's symptom alone delivered for crash, adding the morning biometrics lifted discrimination to 0.81. For fatigue the move was 0.73 to 0.74; for brain fog, 0.83 to 0.85. Every increment was statistically significant, which is unsurprising across hundreds of thousands of observations — and every increment was a point or two of AUROC. The strongest single predictor of a bad evening remained, by some distance, a bad day before.

What it cannot say

An association measured after the fact is not a tested intervention. No one in this cohort saw a warning, rested, and was then shown to crash less often. That step — the one that would justify telling a patient to rearrange a day on the strength of a reading — needs a prospective study in which the alert actually exists and its effect on outcomes is measured. The present work does not reach it, and the authors do not pretend it does; they close by calling for exactly that prospective validation.

Two limits temper the read further. Diagnoses were self-reported through the app — Long COVID against the WHO definition, but no standardised ME/CFS criteria, alongside other energy-limiting conditions — so the cohort is mixed and self-selected, which the authors acknowledge by describing it as 'complex chronic illness' rather than any one disease. And phone-derived photoplethysmography is a coarser instrument than a research electrocardiogram, with no guarantee every morning reading was taken truly at rest; the authors flag this too. There is also a disclosed conflict to weigh: of the eleven authors, three are current or former employees of Visible Health Inc., the company behind the app that produced the data, and the first author reports consulting fees from it, declared as unrelated to this manuscript. The disclosure is clear, and a reader simply factors it in.

Why it is worth reading anyway

The quiet value of this paper is methodological, and it travels well past Long COVID. Most consumer-wearable claims rest on the between-person comparison — the kind that yields a striking population average and tells an individual nothing. By insisting on the within-person question and validating forward in time, this study models how digital-biomarker evidence ought to be read before it is believed. For European systems weighing whether to fold such signals into chronic-illness care, the lesson is not that the morning reading is worthless. It is that the bar is a demonstrated benefit over the cheapest predictor available — and on that bar, the work is honest enough to say it is not yet cleared. The signal is real. The proof that acting on it helps is the study still to be done.

Source: Aitken A, Sawyer A, Iwasaki A, et al. Digital physiological biomarkers predict within-person symptom changes in complex chronic illness. npj Digital Medicine 2026;9:257. A retrospective, observational analysis of real-world app data with within-person modelling and walk-forward validation — strong on association, but with no prospective test that acting on the signal reduces crashes, and with disclosed ties between several authors and the app's manufacturer.