Is There a Feedback Loop in Alzheimer's Biomarker Trajectories? A Falsifiable Test with Structured ODEs

Biologists increasingly describe neurodegeneration as a vicious circle rather than a one-way cascade. Protein deposits switch on brain immune cells. Inflamed and dying cells then disable the brain's waste-clearance machinery, so more protein piles up. The progression models clinicians and machine-learning researchers actually fit to patient data assume the opposite. They treat disease as a one-way sequence (amyloid, then tau, then atrophy, then cognitive decline) along which each patient slides at a personal, fixed speed. These two pictures make different predictions about how a single patient's biomarkers should bend over time. Nobody has checked which one the longitudinal human data support.

We propose to build that check. We will fit small, interpretable differential-equation models to longitudinal Alzheimer's cohorts. In these models, how fast each biomarker changes is a learned function of only the biomarkers biology says should drive it. We then compare a strictly one-way version against a version with one hidden "clearance capacity" variable that downstream injury wears down and that in turn controls how fast amyloid is removed. The two are compared on held-out future visits and on a sharper within-patient signature. The one-way model says that once a patient's personal speed is accounted for, amyloid accumulation should not depend on how fast that same patient's tau and atrophy are worsening. The loop model says it should, and that the dependence should look like a threshold. If the loop is real, the fitted model yields something no current progression model provides: a per-patient estimate of distance to a tipping point, past which clearance cannot recover.

Core Claim

Our central claim is that self-reinforcing neuroinflammatory loops leave a detectable, falsifiable trace in routinely collected biomarker trajectories, even though inflammation itself is not measured. The trace has two parts.

The first is reverse coupling. Within a person, acceleration of amyloid accumulation is predicted by that person's worsening tau or neurodegeneration, beyond what a constant personal progression speed explains.

The second is a threshold. In the fitted dynamics, the hidden clearance variable collapses abruptly once injury passes a critical level. A collapse of this kind is a fold bifurcation: the point where a stable "compensated" state disappears and the system runs away.

A one-way cascade with person-specific speeds predicts neither part. That is what makes the comparison decisive rather than a contest of curve fitting.

Where This Sits in Prior Work

The continuous-progression literature supplies the method and the baselines, and also the assumption we test.

Two related lines of work encode the one-way, fixed-speed picture. Scelsi et al. fuse amyloid PET and hippocampal MRI into a scalar Disease Progression Score. This score is a per-subject time shift onto one population curve, and it was powerful enough to find a progression locus in LCORL that single markers missed. The EuroPOND Event-Based Model, validated by Wittens et al., orders nine markers into a fixed population-wide sequence. Both assume the very thing we test: a fixed ordering and a constant individual rate. Neither can express a downstream marker speeding up an upstream one.

Neural differential-equation models remove the fixed-curve restriction but keep the dynamics as a black box. BrainODE learns a neural ordinary differential equation (a neural network that outputs the rate of change of the state, integrated forward in time) on brain-surface shapes. It conditions only on age and cognition, and it explicitly excludes mild cognitive impairment (MCI) as diagnostically ambiguous. CNODE adds per-patient onset shifts and progression speeds for Parkinson's cortical thickness. ImageFlowNet learns image-level flows with a stochastic variant for uncertainty. These papers motivate our use of continuous time, because they show recurrent models struggle with irregular visits. Their unstructured vector fields are the strongest competitor on forecasting, but they cannot answer a structural question. TD-HNODE adds graph structure to a neural ODE, but only from expert-defined clinical pathways in health records. Our structure comes instead from cellular biology, and the specific edges are what we test.

The neuroinflammation literature supplies the loop topology but no quantitative model. Zou et al. show that protein oligomers disrupt astrocyte water channels and block meningeal lymphatics, reducing clearance. Zhang et al. show drainage bottlenecks cause ApoE to build up and further block metabolite exit. Wu et al. and Fang et al. describe microglial pyroptosis, an inflammatory cell death that releases alarm signals, triggers neighbouring microglia, and permanently removes phagocytes. Ma shows autophagy failure sustaining inflammasome signalling. Together these justify one structural commitment: a clearance resource that injury depletes, that recovers slowly or not at all, and that limits amyloid removal.

All of this evidence comes from animals, cell culture or postmortem tissue. We do not claim to observe pyroptosis. We claim to test whether human trajectories carry the signature such a loop would produce. The hidden variable is a hypothesis-shaped slot, not a measured inflammation marker.

Two observational findings motivate the threshold part of the claim. Kim et al. show that "amyloid-negative" people with subthreshold accumulation still decline. Jung et al. show that amyloid positivity accelerates cortical thinning in amnestic MCI. However, Jung et al. model amyloid only as an input covariate to a recurrent network and never test the reverse direction.

Research Program

The model tracks four observed quantities per person: amyloid burden $A$ from PET, cerebrospinal-fluid phosphorylated tau $T$, neurodegeneration $N$ from hippocampal and temporal volumes, and cognition as a read-out of $T$ and $N$. It also tracks one hidden clearance capacity $C \in [0,1]$. The loop model is

$$ \frac{dA}{dt} = p_\theta(A) - k\,C\,A, \qquad \frac{dC}{dt} = -\lambda\, g_\theta(T, N)\, C + \rho\,(1 - C), $$

with $T$ and $N$ following the usual cascade, $dT/dt = f_\theta(A, T)$ and $dN/dt = h_\theta(T, N)$. The terms are as follows. Production of amyloid is $p_\theta$. The clearance rate constant is $k$. Injury-driven depletion of clearance is $g_\theta$, a small monotone neural network. The loop gain is $\lambda$, the parameter we care most about. The slow recovery rate is $\rho$. Every $f$, $g$, $h$ and $p$ is a small network constrained to be monotone in its inputs and allowed to depend only on its graph parents.

When $\lambda$ is large relative to $\rho$, the $C$ equation has a fold. Below a critical injury level it holds near 1. Above that level it collapses toward 0, and amyloid then accelerates. Setting $\lambda = 0$ recovers a pure one-way cascade with the same networks.

Each person gets their own initial state. They also get either a personal progression speed (in the cascade model) or a personal loop gain (in the loop model), estimated by amortized inference from the first visits. We fit by maximizing likelihood of all observed biomarker values at irregular visit times, with a held-out final visit per person. Missing modalities are simply dropped from the likelihood, which matters for real-world data. Unlike BrainODE, we keep MCI participants, because MCI is where the loop model predicts the most heterogeneity: people at similar biomarker levels but on opposite sides of the threshold.

Three analyses come out of the fitted models. The first is the forecasting comparison described below. The second is a direct within-person reverse-coupling estimate. Amyloid acceleration between visits is regressed on change in $T$ and $N$, conditioning on the person's estimated constant speed. This is done both inside the ODE and as a model-free lagged mixed-effects regression. The third is threshold validation. For each person, the loop model estimates the time until $C$ crosses half its range. People predicted to cross between baseline and last visit should show amyloid curvature that turns upward at later visits, while matched people predicted not to cross should not.

As a secondary, hypothesis-generating analysis, we re-run the Scelsi et al. genetic association on two endpoints separately: the loop gain $\lambda$ and the onset shift. The question is whether the LCORL signal, and APOE ε4 given ApoE's role in Zhang et al., attaches to how strongly the loop amplifies or merely to when disease starts.

Before touching real data, we build a semi-synthetic identifiability benchmark. Trajectories are simulated from both model classes, with known loop gains, at ADNI's actual visit schedules and measurement noise levels. We then check that the fitting procedure recovers $\lambda$, distinguishes loop from no-loop at realistic sample sizes, and does not hallucinate a loop when trajectories are generated by a cascade with heterogeneous speeds. This is the main technical risk. A hidden variable plus flexible networks can absorb many patterns, so we need evidence of power and of a controlled false-positive rate before reading any real-data result.

Baselines

The naive reference is per-person linear extrapolation of each biomarker from that person's previous visits, together with last-observation-carried-forward. Any model that cannot beat straight lines at two-to-four-year horizons is not worth interpreting.

The fixed-curve time-shift model is our reimplementation of the GRACE progression score from Scelsi et al.. Each person is a time shift along one population curve, here extended to all four observed markers. It is the cleanest statement of the one-way, constant-rate hypothesis, so beating it on within-person acceleration is the core result.

The strongest forecasting competitor is an unstructured conditional neural ODE in the style of CNODE, with per-person onset shift and speed. To make the comparison fair we give it one extra free latent dimension, so its state size and parameter count match the loop model. CNODE reports outperforming GRU, Transformer and plain neural ODE baselines on PPMI. The file gives no scalar number, and there is no reported result found on Lacuna for any ODE forecasting metric on ADNI biomarkers; BrainODE's reported figures are hippocampal shape errors of 0.365 mm on LBC1936 and 0.406 mm on AIBL, on a different target. We will therefore run this baseline ourselves.

We also include a GRU with time decay, following Jung et al., who report hippocampus $R^2 = 0.924$ on their Samsung MCI cohort. That figure is on different data and cannot be compared directly, so we retrain the architecture on ADNI.

For MCI-to-AD conversion as a downstream check, the reference is HOPE, which reports 72.0% accuracy on ADNI and 67.6% on NACC. We re-run HOPE on our exact ADNI split, and only compare on the same subjects.

Two ablations isolate the key ingredient. The cascade-only ablation sets $\lambda = 0$ while keeping every network and the personal-speed term. The frozen-clearance ablation keeps $C$ as a learned per-person constant that never depletes. This separates "an extra person-level latent helps" from "a depleting, injury-driven latent helps."

All models share the same subjects, splits, input modalities, and a matched hyperparameter search budget of equal trials and GPU hours. They are tuned only on validation visits and scored once on held-out final visits.

Beating the baselines means three things. The loop model must improve held-out forecasting error for amyloid at horizons of at least three years over the cascade ablation and the time-shift model, with confidence intervals from subject-level bootstrap excluding zero. It must be within noise of, or better than, the unstructured neural ODE. And it must yield a reverse-coupling effect that survives the synthetic false-positive calibration.

Why not just…?

Why not just fit an unstructured neural ODE with a spare latent dimension? This is the real threat. A black-box vector field with an extra state can probably forecast as well as our structured model, and it may forecast better. We accept that outcome as possible and do not stake the proposal on forecasting superiority. The unstructured model cannot say whether amyloid's rate depends on downstream injury, nor where a threshold sits. If it matches us on error while our structural tests come back positive, the structure is a free explanatory gain. If it clearly beats us, that tells us the biologically specified loop is the wrong shape, and we will report it.

Why not just give each patient a personal progression speed, as the progression-score and CNODE models do? Personal speed already explains much of the heterogeneity that might look like feedback. That is why it is the null hypothesis rather than a strawman. The two models disagree on a specific quantity: personal speed is constant in disease time, while a loop makes the same person accelerate as their own tau and atrophy worsen. The within-person reverse-coupling estimate, calibrated on synthetic cascades with heterogeneous speeds, is exactly the experiment that separates them.

Why not just regress amyloid change on lagged tau and atrophy in a mixed-effects model? This is a cheap, honest version of our second analysis, and we run it as a baseline. If it finds no reverse coupling, the ODE machinery is unlikely to rescue the claim. Its weaknesses are practical: irregular visits, missing modalities, PET signal saturation at high burden, and regression to the mean all bias lagged differences. What the ODE adds beyond handling these is the threshold. A linear regression cannot predict who will cross a tipping point between visits, and our validation analysis tests exactly that.

Why not just use the Event-Based Model, which already stages patients without cut-offs? Wittens et al. show the model generalizes even to messy clinic data. However, it assumes one invariant ordering and outputs integer stages with no rate, so a loop is inexpressible by construction. We use it only as a sanity reference for ordering. We also note that its finding that non-amnestic MCI patients rarely reach advanced stages is the kind of off-sequence heterogeneity our model should explain.

Proposed Experiments

The primary dataset is ADNI, which contains longitudinal amyloid PET, cerebrospinal-fluid tau, MRI volumes, cognition and genotypes, with MCI participants included. External validation uses AIBL. The file lists it as longitudinal imaging and biomarker data but does not describe its fluid markers, so we will use whatever overlap of modalities it shares with ADNI. The semi-synthetic identifiability benchmark comes first and gates everything else.

Metrics are as follows. For forecasting, mean absolute error per biomarker at one, three and five-year horizons. For reverse coupling, the within-person coefficient with bootstrap intervals. For the threshold, recovery error of $\lambda$ and the bifurcation point on synthetic data, and on real data the difference in later-visit amyloid curvature between predicted crossers and matched non-crossers. For conversion, MCI-to-AD accuracy against HOPE on identical subjects.

The decisive checks are three. First, the synthetic benchmark must show the method recovers loops at ADNI-scale sampling with false-positive rates near nominal, and does not confuse speed heterogeneity with feedback. Second, on ADNI, the full loop model must beat the cascade-only and frozen-clearance ablations on long-horizon amyloid forecasting while producing a positive within-person reverse-coupling estimate, and this estimate must replicate in the model-free lagged regression. Third, the predicted tipping-point crossers must show upward-bending amyloid trajectories on visits the model never saw, with the effect replicating in AIBL.

A null result on all three would be informative in its own right. It would say that, at the resolution of current clinical biomarkers, human Alzheimer's trajectories are indistinguishable from a one-way cascade, and that the loop biology described in animal and tissue studies either acts on scales these measurements cannot see or does not dominate progression.

References

  1. Genetic study of multimodal imaging Alzheimer’s disease progression score implicates novel loci
  2. Independent validation and outlier analysis of EuroPOND Alzheimer’s disease staging model using ADNI and real-world clinical data
  3. BrainODE: Neural Shape Dynamics for Age- and Disease-aware Brain Trajectories
  4. Conditional Neural ODE for Longitudinal Parkinson's Disease Progression Forecasting
  5. ImageFlowNet: Forecasting Multiscale Trajectories of Disease Progression with Irregularly-Sampled Longitudinal Medical Images
  6. Temporally Detailed Hypergraph Neural ODE for Disease Progression Modeling
  7. Blocking meningeal lymphatic drainage aggravates Parkinson’s disease-like pathology in mice overexpressing mutated α-synuclein
  8. Microglia prevent beta-amyloid plaque formation in the early stage of an Alzheimer’s disease mouse model with suppression of glymphatic clearance
  9. Microglia Pyroptosis: A Candidate Target for Neurological Diseases Treatment
  10. The role and mechanism of NLRP3 inflammasome-mediated astrocyte activation in dehydrocorydaline against CUMS-induced depression
  11. The emerging role of autophagy and mitophagy in tauopathies: From pathogenesis to translational implications in Alzheimer’s disease
  12. The impact of subthreshold levels of amyloid deposition on conversion to dementia in patients with amyloid-negative amnestic mild cognitive impairment
  13. Deep learning model for individualized trajectory prediction of clinical outcomes in mild cognitive impairment
  14. HOPE: Hybrid-Granularity Ordinal Prototype Learning for Progression Prediction of Mild Cognitive Impairment