PhronesisThe claim being checked
The framework says a system has a ground state, that everything else is displacement, and that displacement costs something to hold. Its one measurable quantity is return cost — how expensive it is to get back.
In the autonomic nervous system that quantity has an instrument. Heart rate variability indexes the return, and the rate at which the return slows is a validated early warning: critical slowing down in day-to-day affect precedes the onset of depression. Sympathetic away, parasympathetic back, and a number for the coming-back.
In cancer it has no instrument, and Autodisplasia said so. Return cost there is a re-description of stage. This note went looking for the strongest counter-argument available.
The strongest counter-argument
There is one, and it is serious. Dynamical network biomarkers — Chen, Liu and Aihara — propose that a complex disease does not deteriorate smoothly but passes through a pre-disease state: a critical point where a small group of molecules starts to fluctuate together, hard, while decoupling from everything else. Detect that group and you have an early warning before the transition, not after it. The original paper reports it for lung injury, liver cancer and lymphoma; a 2021 review collects a decade of it.
This is the framework’s own claim in someone else’s vocabulary. A system in a basin, a restoring force weakening, fluctuations growing as the return gets slower — that is displacement and return, and it is being measured in tumours. If it works, Autodisplasia’s admission was premature.
What it actually does
So read the method rather than the abstract. The single-sample version exists precisely because the original could not be run on people: “evaluating DNB at the critical state required the data of multiple samples on each individual, which are generally not available.” That sentence is the whole problem, stated by the authors, in the paper that works around it.
The workaround is to stop needing a series. Score one tumour sample against a reference set of adjacent normal tissue, then group the scores by clinical stage. Applied to three cancers in TCGA, the index peaks at stage IIIB for lung and stomach adenocarcinoma and stage III for thyroid — one stage before metastasis.
The tipping point is found in a cohort that was sorted by stage before the method ran.
Nothing here is wrong, and the mechanistic reading is genuinely interesting: the stage before spread looks, molecularly, like a system losing its restoring force. But notice what it cannot do. It is retrospective, on samples already labelled. It reports a cohort average, not a person’s trajectory. And the stage it identifies as critical is a stage the pathologist has already written down. To act on it you must first know the thing it tells you.
So the admission stands. Not because the work is weak — because it is measuring across people where the framework’s claim is about one system over time.
The actual difference, which is not biology
Set the two cases side by side and the asymmetry is not about tumours versus nerves.
HRV is continuous, non-invasive, and effectively free once a strap is on. You get thousands of beats an hour, indefinitely, from one person. A rolling window of variance and lag-1 autocorrelation has something to roll over.
Cancer’s equivalent signal is circulating tumour DNA, and it is a good signal. Serial post-operative ctDNA precedes radiographic recurrence by a median of around six months in gastric cancer, and a systematic review across resectable cancers finds longitudinal testing beats single landmark tests on sensitivity. It is per-person, quantitative, and repeatable — everything the cohort methods are not.
It is drawn every three to six months.
Two to four points a year cannot support a rolling-window statistic. Our own detector for this — the one built for the autonomic study — needs at least twenty points per person before it will report a trend at all, and it was working on nightly data. At quarterly sampling, twenty points is five years, over which the underlying system is not the same system.
A domain admits this framework’s number exactly when its return signal can be sampled faster than the system moves.
That is a general claim, it is checkable, and it explains the pattern rather than restating it. Affect: daily diaries, and critical slowing down works. Autonomic: continuous, and it works. Cancer: quarterly, and it does not. The reach is set by the sampling rate, not by the disease.
Where this could still be wrong
Two ways, and the second is worse for the idea.
Density might arrive. ctDNA assays get cheaper on the usual curve, and nothing prevents monthly or weekly draws in a trial that wants them. If that happens the test becomes available and this note becomes a prediction rather than a limit.
Density might not be enough. Critical slowing down is a statement about a system sitting in a basin while the walls flatten. Growing tumour burden may not be that shape at all — it may be escape from a basin already lost, or plain exponential growth, and a quantity with no restoring force cannot have a slowing return. If so, denser sampling would produce a clean null and the framework’s reach into cancer closes for a deeper reason than logistics. That would be the more useful result of the two, and it is the one this note would bet on if forced.
What would settle it
One study, stated so it can fail:
Take a resected cohort on ctDNA surveillance. Sample monthly for eighteen months. Compute windowed variance and lag-1 autocorrelation of each person’s ctDNA trajectory, detrended. Then ask whether those indicators rise before radiographic relapse in the people who relapse, and stay flat in the people who do not — scored before anyone knows which group is which.
If they rise, return cost has an instrument in cancer and Autodisplasia’s admission is overturned by measurement. If they do not, the framework does not reach here, and that should be written down as plainly as this sentence.
Neither outcome needs a new theory. It needs somebody to draw blood twelve times instead of three.
References
Dynamical network biomarkers, original statement: Chen L, Liu R, Liu Z-P, Li M, Aihara K, “Detecting early-warning signals for sudden deterioration of complex diseases by dynamical network biomarkers,” Scientific Reports 2:342, 2012 — full text. Decade review: Liu R, Chen P, Chen L, “Dynamical network biomarkers: theory and applications,” Gene 2021 — record.
The single-sample method, its stated motivation, and the stage-IIIB result quoted above: Yu X, Zhang J, Sun S, et al., “Quantifying critical states of complex diseases using single-sample dynamic network biomarkers,” PLOS Computational Biology 13(7):e1005633, 2017 — full text, doi:10.1371/journal.pcbi.1005633.
ctDNA lead time and the longitudinal-versus-landmark comparison: Yang J, Gong Y, Lam VK, et al., gastric cancer molecular residual disease, 2020 — full text; and the systematic review and meta-analysis of ctDNA-based residual disease detection in resectable cancers, 2024 — full text.
The autonomic side of the comparison, including the depression early-warning result it rests on, is in Autonomic Displasia. The detector referred to, and the twenty-point floor, are from the autonomic study’s Phase A.
On the links. The PLOS copy of the single-sample paper answered 502, then 200, then not at all across three consecutive requests, so it is given as a DOI rather than a hyperlink and the stable PMC mirror carries the link instead — this site does not carry a link it cannot stand behind. Every hyperlink above returns 200.