The Cybernetic Neuromorphic Framework: an Artificial Peripheral Thalamus for the Stroke Continuum

A fourteen-module multi-scale model, built only from public data, that specifies what the device must do — and what it cannot be expected to achieve

Authors. Tunjay Akbarli¹, Parvin Akbarov²

• ¹ Chief Architect, Omnira CJSC, Baku, Azerbaijan
• ² MSc Healthcare Administration, Johns Hopkins University, Baltimore, MD, USA

Correspondence. T.A.: tunjay.akbarli@theomnira.com

Background. The device this paper is built around is an artificial peripheral thalamus: a cybernetic neuromorphic relay that stands in for the damaged afferent pathway after stroke, regulating the timing and fidelity of sensory feedback during active movement. A device of that kind is only as good as the answer to three questions — how much benefit is available to it at all, how fast its loop must close, and when in the recovery course its sessions should be spent. None of the three can be answered inside the rehabilitation literature, because all three are set upstream: by the flow field that decided how much cortex survived, by the cascade that decided how fast that decision was made, and by the kinetics of the sensitive period that decides what any training can still reach. We therefore built the disease model the device needs, and held both to the same standard of evidence.

Methods. We assembled a provenance-tagged public evidence base spanning the whole stroke continuum — NCBI (Gene, dbSNP via the Ensembl mirror, GEO, PubChem BioAssay, UniProt/RCSB), Data.gov and CDC (NVSS mortality, PLACES, BRFSS, NHANES), GBD, and the primary literature — in which every value carries a numeric source key and a primary/secondary/UNVERIFIED flag, and unverified cells raise an exception rather than being imputed. On that base we built fourteen coupled modules sharing one state vector: K social and access geography, A genomic risk architecture, F a curated multi-layer molecular network, B perfusion and tissue fate, C a fourteen-state ictal cascade, J a spatially resolved stochastic reaction–diffusion model of infarct growth with spreading depolarisation, G haemorrhage and haematoma expansion, H small-vessel disease, post-stroke cognition and aphasia, I a therapeutic layer that asks what effect each failed neuroprotection trial could have attained, D a two-timescale recovery law, E the closed-loop interface, M a stochastic dynamic program for the device's own schedule, and L/N the inferential machinery — Bayesian posteriors, bootstrap intervals, profile-likelihood identifiability, variance-based sensitivity and multiplicity control. Free constants were solved inside the model against named verified observations; every other comparison was reserved as an out-of-sample test.

Results. Eighty-four comparisons with the published literature, fifty-two of them out-of-sample; fifteen carry a published interval from which a standard error can be recovered, five disagree uncorrected, and four survive Benjamini–Hochberg control of the false-discovery rate. The device layer is specified rather than asserted: at matched dose the entire achievable benefit of better feedback is a standardised mean difference of 0.29 at 10 ms loop latency — bracketing the published robot-training effect of 0.32 — falling to zero from about 80 ms; and the stochastic dynamic program shows that at a 20-hour budget the same hours spread evenly across the year fall below the verified dose threshold of 1.72 times usual care and buy nothing, while concentrated they yield 6.6 Fugl-Meyer points and clear the minimal clinically important difference. At the verified repetition rates, the dose threshold is reached at 43 minutes a week of therapist-delivered practice and 4 minutes a week through the device, so the device's value is not a multiplier on an effect that already exists but the fact that it reaches the threshold at a tenth of the contact time. Upstream, the model reproduces the published median infarct growth rate, time to complete infarction and collateral–growth correlation with one calibrated parameter; a spatially resolved reaction–diffusion model on a cortical sheet meets 12 of 17 spatial benchmarks and shows that recurrent depolarisations recruit only 0.13% of the final infarct volume while advancing its completion; the haemorrhage layer attributes 55% of the effect of intensive blood-pressure lowering to haematoma growth; and the therapeutic layer finds that 15 of 15 neuroprotection trials were powered to detect an effect larger than the largest effect attainable at their own treatment window, a conclusion that survives replacing the framework's salvage curve with the steeper one the published reperfusion ladder itself implies. Three independent benchmarks — the thrombolysis time gradient, the penumbral share of the final infarct, and mismatch prevalence at 18 hours — fail in the same direction, which locates a specific defect in the framework's tolerance curve rather than a diffuse one. The motor recovery law does not transfer to language: all four aphasia benchmarks fail. Turning the negative therapeutic result into a specification, a delivery-physics layer shows that interstitial penetration falls only as the inverse square root of carrier radius — so size is a weak lever — while the penumbra receives 38% of the delivery normally perfused cortex receives, and that a sustained selective enrichment of about sevenfold in penumbral tissue, which no improvement in potency can substitute for, is what would convert the attainable effect into a detectable one. Where someone lives is not a footnote to the genome: the modifiable risk-factor burden is 7 times the liability variance the polygenic score explains, and the documented racial gap in thrombectomy access is worth at least half an hour of additional transport time on the most conservative calibration of the time–benefit curve.

Conclusions. A single framework can carry state from a postcode and a genotype through a spatially resolved infarct to a rehabilitation schedule, using public data alone, with every value traceable to its source. Its value is less prediction than arbitration and specification: it identifies published quantities that cannot simultaneously be true, it says what a closed-loop interface can and cannot buy before the trial is run, and it reports its own failures — eight of which are converted into pre-registered predictions with numeric falsification thresholds.

Keywords. stroke; artificial peripheral thalamus; closed-loop rehabilitation; multi-scale modelling; spreading depolarisation; reaction–diffusion; polygenic risk; proportional recovery; stochastic optimal control; neuroprotection; reproducibility

Source Code: https://github.com/hekimoghlu/Thalamus