Toward a Dynamical Taxonomy of Insomnia: A Multiaxial Framework for Sleep-State Transitions and Architectural Failure

Insomnia disorder is defined at the syndrome level, yet similar complaints can arise from different abnormalities in sleep regulation, state transition, state stabilization, spatial recruitment, architectural sequencing, and state perception. We propose a multiaxial dynamical framework whose principal contribution is organizational: a candidate profile is specified by the dynamical operation that fails, the sleep stage or boundary at which it fails, and its causal status. Objective sleep duration, age, circadian phase, comorbidity, medication exposure, and night-to-night variability are modifier/covariate dimensions rather than additional mechanistic classes. A local Landau-Ginzburg formalism, adopted from prior cortical and sleep-dynamics work, supplies a phenomenological language for nested hypotheses. Its relaxational form applies only to boundary-local dynamics under an approximate gradient description; non-gradient escape requires an action or quasipotential treatment, and whole-night REM-NREM sequencing requires reactive or oscillatory dynamics. Routine polysomnography usually identifies effective combinations rather than curvature, escape action, bias, noise, and relaxation separately. The strongest boundary-level evidence concerns sleep onset, where published results are consistent with bifurcationlike or bistable dynamics but do not yet exclude a driven smooth transition produced by the homeostatic-circadian ramp. The remaining operation classes are hypothesis-generating extensions. The taxonomy is judged by pragmatic utility, including improved communication, stratification, and prediction; the scalar-field implementation and specific dynamical profiles are separately falsifiable. The framework is a phenomenological organizing model rather than a new diagnosis, validated biomarker, or treatment-selection system.

The central hypothesis is that insomnia complaints may arise from separable abnormalities in slow control trajectories, state-boundary crossing, post-transition stabilization, spatial recruitment, architectural sequencing, and the mapping between physiological and experienced sleep. A low-dimensional nonequilibrium framework can organize these abnormalities and generate testable distinctions that are not captured by symptom labels or conventional stage totals alone.

The framework makes two principal affirmative predictions. First, local recovery from small perturbations and escape under larger perturbations will dissociate reproducibly across patients and will predict perturbation vulnerability or outcome beyond a single undifferentiated instability measure. Second, rounded-bifurcation or bistable/escape models will predict subject-held-out sleep-onset trajectories better than a Process-S/Process-C-driven smooth sigmoid. Failure of either prediction would materially weaken the dynamical implementation, while leaving the descriptive ontology to be judged by its independent utility criteria.

The taxonomy is organized along three independent descriptive axes. Axis I (dynamical operation) contains six classes: control-path progression; boundary crossing; post-transition stabilization; spatial recruitment; architectural sequencing; and observation mapping. These operation classes are the operational taxonomic level. Wake-state curvature, barrier height, bias, or noise are competing mechanistic explanations within boundary crossing rather than parallel diagnoses. Axis II (stage or boundary) specifies Wake → N1, N1 → N2, N2 → N3/SWS, NREM ↔ REM, sleep → wake, or the whole-night transition network. The location does not determine the mechanism. Axis III (causal status) assigns every dynamical signature a potential causal role: primary defect, secondary architectural consequence, maintaining mechanism, associated biomarker, or undetermined. Cross-sectional PSG usually supports only the undetermined category, together with exclusion of obvious event-linked secondary causes.

The framework distinguishes five properties that must remain distinct: local curvature or stiffness μ = V''(ψeq) controls the response to sufficiently small perturbations; under the gradient approximation, ΔV = V(saddle) −V(minimum) is the local potential barrier; in a non-gradient system, the corresponding escape quantity is the minimum action or quasipotential difference ΔW; relative well depth is meaningful only within an approximately gradient representation; the field h tilts a local gradient model, while Γ converts local curvature into a recovery timescale. A state can therefore be locally stiff yet easy to escape from, or locally soft yet protected by a large escape action. Spontaneous transition rates identify only combinations such as ΔV/Deff in the gradient case or ΔW/ε in a general small-noise action formulation. Small perturbations primarily estimate displacement and recovery, whereas larger perturbations probe empirical escape. The resulting dissociation is central: rapid local recovery can coexist with frequent full state escape, and slow local recovery can coexist with infrequent escape.

The evidence base is sharply asymmetric. Sleep onset provides the strongest current boundary-level evidence. Li and colleagues identified a low-dimensional coordinate with a predictable bifurcation-like trajectory and passive early-warning signals in two large cohorts. Hu and colleagues fitted a stochastic bistable model that reproduced wake-sleep flickering and related drift/noise parameters to sleepiness. These results are strongly consistent with bifurcation-like or bistable dynamics. They do not yet distinguish intrinsic critical slowing from slowing induced by the deterministic Process-S/Process-C ramp, establish small-field bistability, demonstrate hysteresis, or show enrichment of a particular mechanism in insomnia. Fold language is therefore reserved for analyses that outperform a driven smooth null and identify branch loss or matched-control bistability.

A multicenter CBT-I study found dissociable changes in the NREM delta/beta ratio and transition-matrix stability, consistent with the framework's separation of cortical activation from macrostate stability, but it does not identify curvature, barrier height, bias, or noise. An auditory perturbation study found greater evoked complexity and P300 responses, especially in N3, with altered intermediate-window coupling, compatible with reduced perturbation resistance or containment, but graded perturbations are required to distinguish a soft state from a low barrier. Reduced 0.5-1 Hz slow-oscillation power in older adults with insomnia and reduced spindle and slow-oscillation density in a broader adult sample motivate N2/SWS organization domains, but they do not distinguish reduced local ordering, weak spatial recruitment, or low state stability. Large-sample REM findings - increased REM arousals and awakenings, REM density, and phasic REM proportion - motivate treatment of REM microstructure as a distinct domain, but they are compatible with REM instability without uniquely identifying reduced REM curvature, lowered barrier height, altered bias, or multivariate noise. Sleep-state misperception studies show altered NREM spectral activity, evoked processing, reproducible multi-night discrepancy patterns, and associations with N3 proportion, motivating an observation-layer model but not establishing whether the dominant source is local wake-like physiology, sensory access, memory encoding, expectations, or objective transition instability.

Several findings constrain broad claims. Some carefully characterized insomnia samples show little or no group-level waking or NREM spectral difference, and meta-analytic PSG effects are heterogeneous rather than diagnostic at the individual level. Slow-oscillation, spindle, and conventional N3 abnormalities are absent in subsets, so N2/SWS profiles cannot be universal. Misperception studies identify multiple discrepancy patterns rather than one observation-layer physiology. REM microstructural differences can coexist with near-normal macroarchitecture. No study has yet demonstrated stable operation-location profiles across multiple nights or shown that the proposed parameters outperform simpler model-free measures.

The most decisive currently feasible test is a preregistered reanalysis of an existing large sleep-onset cohort. The same continuous onset trajectories should be fitted with (i) a driven smooth sigmoid in which Process S and circadian phase supply the ramp, (ii) a rounded bifurcation model with an estimated bias/rounding parameter, (iii) stochastic bistable or escape models, and (iv) hidden-state/change-point baselines. Model selection should use subject-held-out likelihood and calibration, with the transition-localizing features separated from the testing features. If the driven sigmoid predicts as well as or better than the rounded or bistable models, the manuscript's sleep-onset claims must be downgraded to descriptive state change and the top evidence tier remains empty. A reproducible predictive advantage for the dynamical alternatives would justify targeted perturbation studies and sharpen the framework's strongest empirical foothold.

The three-axis taxonomy is an ontology rather than a physical law. It earns its place if trained raters can apply operation-location-causal labels reliably, if the labels improve communication or trial stratification, or if they add held-out prediction beyond conventional onset/maintenance/terminal-awakening/nonrestorative subtyping and standard covariates. If it produces poor inter-rater agreement, excessive unresolvable ambiguity, or no incremental predictive utility, the ontology should be simplified or retired even when an individual dynamical model remains mathematically adequate. The scalar-field implementation is rejected if ϕ lacks partial invariance, discriminant validity, and transport, or if vector and simpler models consistently predict better. A particular mechanistic profile is rejected when its required transition or perturbation signature is absent after confounds are modeled. The Landau-Ginzburg implementation is rejected for a boundary when a driven sigmoid, HMM/semi-Markov model, change-point model, or model-free perturbation description consistently wins on subject-held-out prediction.

The framework is phenomenological. Its coefficients summarize low-dimensional state-space geometry and do not identify a unique circuit, neurotransmitter, or molecular mechanism. Different biological causes may produce similar effective geometry, and the same cause may alter several parameters. The model narrows hypotheses but does not replace mechanistic neuroscience. The relaxational quasipotential is local and conditional on approximate gradient dynamics. A global scalar potential need not exist for the full sleep cycle because non-gradient probability currents, adaptation, delays, REM-on/REM-off oscillations, and Hopf-like components contribute to whole-night dynamics. In those regimes, escape must be described through a minimum-action quasipotential or model-free hazard, and sequencing requires reactive or oscillator terms. Process S, Process C, and neuromodulatory variables remain external controls to the local models rather than autonomously generated state variables. A scalar NREM-ordering coordinate is a deliberate reduction and may fail when spindles, K-complexes, local slow waves, and high-frequency activation vary independently. REM and wake require a multidimensional vector. Measurement invariance, lead-field effects, finite-size rounding, colored noise, first-night effects, and stage-scoring rules can all manufacture or obscure apparent transition signatures.

The immediate value of the framework is clinical interpretability and research stratification, not actionability. It suggests augmenting a standard assessment by asking whether the dominant complaint is failure to become sleepy, failure to cross sleep onset, repeated early-stage reversals, nocturnal escape from established sleep, failure of N2 or SWS organization, REM fragmentation, premature terminal arousal, whole-night dysregulation, or a discrepancy between experienced and measured sleep. These patterns can guide differential measurement, but prospective evidence is required before they guide treatment selection. CBT-I can alter several dimensions simultaneously: homeostatic pressure, conditioned wake bias, circadian regularity, and arousal. Pharmacologic agents that produce similar total sleep time may likewise differ in their effects on boundary crossing, local recovery, escape, N2 event probability, or REM organization. These are hypotheses for mechanism studies, not current prescribing rules. Closed-loop neuromodulation is included solely as an illustrative example of hypothesis generation, with no validated target, product rationale, or dosing rule.

A useful dynamical taxonomy would complement diagnosis by asking: at which operation and boundary does the sleep process appear to fail, what causal role remains plausible, and does that profile improve communication, stratification, or prediction beyond conventional measures? A successful dynamical model must answer the additional question of whether its inferred geometry predicts perturbation response and outcome better than driven smooth, hidden-state, and model-free alternatives.

Improvements to AI Systems Based on This Paper:

1. Hierarchical Causal-Profile Classifier for Sleep Disorders

• Improvement: Build a multi-label AI classifier that assigns each patient a three-axis profile (dynamical operation, stage/boundary, causal status) from raw polysomnography (PSG) and actigraphy data, rather than outputting a single insomnia subtype.

• Capability: The AI can distinguish between, e.g., boundary-crossing failure at Wake→N1 with primary defect vs. post-transition stabilization failure at N1→N2 with maintaining mechanism, enabling more precise patient stratification for clinical trials and differential diagnosis.

2. Model-Selection Engine for Sleep-Onset Dynamics

• Improvement: Implement a Bayesian model-comparison framework that fits four competing generative models (driven smooth sigmoid, rounded bifurcation, stochastic bistable/escape, hidden-state change-point) to continuous sleep-onset trajectories, using subject-held-out likelihood and calibration.

• Capability: The AI can automatically determine whether a patient's sleep-onset data is better explained by a deterministic homeostatic ramp or by genuine bistable/escape dynamics, providing a reproducible test of the paper's core hypothesis and flagging which patients warrant perturbation-based follow-up.

3. Dissociation Predictor for Local Recovery vs. Global Escape

• Improvement: Train a neural network on perturbation-response data (small vs. large stimuli) to estimate two independent parameters: local curvature (μ) and escape action (ΔW), rather than a single instability score.

• Capability: The AI can predict whether a patient will show rapid local recovery but frequent full-state escapes (e.g., brief arousals that escalate to full awakenings) or slow local recovery with rare escapes, enabling personalized vulnerability assessment and targeted intervention (e.g., stabilizing barriers vs. increasing local stiffness).

4. Ontology-Aware Annotation Assistant for Sleep Researchers

• Improvement: Develop a semi-supervised NLP and rule-based tool that parses sleep research literature and clinical notes to automatically assign Axis I–III labels (operation, location, causal status) to described findings, with confidence scores and inter-rater reliability metrics.

• Capability: The AI can accelerate systematic reviews and meta-analyses by tagging thousands of studies with the proposed taxonomy, identifying which operation-location combinations have strong evidence vs. remain hypothesis-generating, and flagging ambiguous or contradictory findings.

5. Perturbation-Response Simulator for Closed-Loop Neuromodulation

• Improvement: Create a generative AI model that simulates sleep-state trajectories under different perturbation protocols (e.g., auditory stimuli, transcranial stimulation) using the Landau-Ginzburg formalism, with separate parameters for curvature, barrier height, bias, and noise.

• Capability: The AI can pre-test closed-loop stimulation strategies in silico—e.g., determining whether a low-amplitude continuous stimulus or a high-amplitude brief pulse is more effective for a given patient's inferred dynamical profile—before any human trial, reducing risk and cost.

6. Multi-Night Stability Tracker for Dynamical Signatures

• Improvement: Implement a longitudinal AI model that tracks a patient's inferred dynamical parameters (e.g., μ, ΔV, ΔW, bias) across multiple nights, using hierarchical Bayesian methods to separate true trait-level stability from night-to-night noise and first-night effects.

• Capability: The AI can identify whether a patient's boundary-crossing failure is a stable trait or a transient state, and can predict which patients are likely to show reproducible profiles—directly addressing the paper's call for demonstrating stable operation-location profiles across nights.

7. Misperception-Source Disentangler

• Improvement: Build a multimodal AI that combines PSG, EEG spectral features, evoked-response potentials, and self-reported sleep quality to decompose sleep-state misperception into four potential sources: local wake-like physiology, sensory access, memory encoding, and transition instability.

• Capability: The AI can output a probabilistic attribution (e.g., 70% local wake-like EEG, 20% memory encoding, 10% transition instability) for each patient, enabling targeted cognitive or pharmacological interventions rather than a one-size-fits-all approach to paradoxical insomnia.

8. Causal-Role Inference Engine from Cross-Sectional Data

• Improvement: Use causal discovery algorithms (e.g., PC, FCI) on large PSG datasets to infer the likely causal status (primary defect vs. secondary consequence vs. maintaining mechanism) of each dynamical signature, while explicitly modeling confounds like age, circadian phase, and medication.

• Capability: The AI can generate testable causal hypotheses—e.g., reduced REM curvature is a secondary consequence of NREM instability, not a primary defect—that can guide future longitudinal or intervention studies, moving beyond the current undetermined category.

9. Adaptive Sleep-Stage Transition Predictor

• Improvement: Develop a recurrent neural network that predicts the probability of imminent state transitions (e.g., Wake→N1, NREM→REM) in real time, using the paper's distinction between local stiffness and escape action to weight different EEG features (e.g., slow-oscillation power vs. spindle density) dynamically.

• Capability: The AI can provide early-warning alerts for impending arousals or awakenings in clinical monitoring, and can adapt its prediction strategy per patient based on their inferred dominant operation class (e.g., boundary-crossing vs. stabilization failure).

10. Taxonomy Utility Scorer

• Improvement: Create an AI evaluation framework that quantitatively assesses whether the three-axis taxonomy improves communication, stratification, or prediction over conventional insomnia subtypes (onset/maintenance/terminal/nonrestorative), using metrics like held-out predictive accuracy, inter-rater agreement, and incremental information gain.

• Capability: The AI can automatically decide, based on empirical data, whether the proposed ontology should be retained, simplified, or retired—providing a rigorous, data-driven validation of the framework itself, rather than assuming its utility.