Interpreting CSF, PET, and Neurofilament Biomarkers in Progression Models

From Disease States to Measured Biomarkers

Progression models describe latent biological quantities such as aggregate burden, synaptic integrity, and neuronal injury. CSF assays, PET scans, and neurofilament measurements do not observe those quantities directly. Each measurement reflects a particular molecular species or imaging signal, filtered through assay properties, biological production and clearance, and measurement error. A useful model therefore separates the disease-state model from the biomarker observation model.

This distinction matters when longitudinal signals disagree. Aggregate burden can increase while a soluble CSF species decreases; a PET signal can change little when its response is near saturation; and neurofilament can remain elevated after an injury episode has subsided. Such patterns need not imply that one result is wrong. They should prompt examination of what each biomarker measures and how quickly it responds.

Definition: Let \(A(t)\) denote latent aggregate burden, \(S(t)\) latent synaptic integrity, and \(J(t)\) the rate of neuronal injury at time \(t\). These are biological states or processes, not assay results. A biomarker observation is a measured quantity linked to one or more of these states by an observation function, which may include production, clearance, delay, saturation, and measurement error.

The earlier material on alpha-synuclein and tau aggregation kinetics, linking protein burden to neuronal injury and loss, and neuronal reserve and nonlinear functional decline supplies the disease-state context. Here the emphasis is on connecting those states to assay measurements. In particular, \(J(t)\), an injury rate, should not be confused with accumulated neuronal loss: a biomarker released during injury may respond to the rate of damage rather than to the total amount of damage already sustained.

CSF Measurements: The Measured Species Determines the Direction

A CSF assay may quantify a soluble protein, an oligomeric or modified species, or a synaptic or injury-associated analyte. Its concentration is governed by the balance between appearance in the measured compartment and removal from it. A basic one-compartment model is

\frac{dC}{dt}=q_C(t)-k_C C(t), \qquad k_C>0,

where \(C(t)\) is the measured CSF concentration, \(q_C(t)\) is its net input rate, and \(k_C\) is a first-order clearance rate. If \(q_C\) is constant, the equilibrium concentration is \(C_{\mathrm{eq}}=q_C/k_C\). The units are consistent: concentration per time divided by inverse time gives concentration.

The biological source term must be chosen for the specific analyte. For example, a soluble species consumed into aggregates could have a source that decreases as \(A(t)\) rises. Another analyte could be released in response to aggregation or injury, so its source might rise with \(A(t)\), \(J(t)\), or both. Therefore, “CSF biomarker rises with disease” is not a general rule. The direction is a property of the analyte and its proposed mechanism, not of the fact that it was measured in CSF.

Worked Example: A Soluble CSF Species Falls as Aggregation Increases

Suppose, hypothetically, that a soluble CSF species is consumed as aggregate burden increases, with source rate \(q_C=12-0.8A\) concentration units per day and clearance \(k_C=0.5\) per day. Assume aggregate burden rises from \(A=2\) to \(A=5\) units, rapidly enough that the source can be approximated as changing immediately. The initial CSF concentration is at equilibrium for \(A=2\).

At \(A=2\), the source is \(q_C=12-0.8(2)=10.4\) concentration units per day, so the initial equilibrium is \(C(0)=10.4/0.5=20.8\) concentration units. After the increase to \(A=5\), the source is \(q_C=12-0.8(5)=8.0\) concentration units per day, giving a new equilibrium of \(C_{\mathrm{eq}}=8.0/0.5=16.0\) concentration units.

For a constant source after the change, the solution is \(C(t)=C_{\mathrm{eq}}+[C(0)-C_{\mathrm{eq}}]e^{-k_Ct}\). At 3 days, \(C(3)=16.0+(20.8-16.0)e^{-0.5(3)} =16.0+4.8e^{-1.5}\approx17.07\) concentration units.

Thus aggregate burden increases while this particular soluble CSF measure falls toward its new equilibrium. This conclusion follows from the stipulated source mechanism; it is not a universal prediction for CSF biomarkers.

PET: A Tracer-Dependent, Potentially Saturating Readout

A PET measure is an imaging signal, not a direct count of aggregates. Its relationship to latent burden depends on tracer binding, nonspecific signal, acquisition, and image processing. A simple saturable observation function is

Y_P(t)=b_P+B_{\max}\frac{A(t)}{K_P+A(t)}+\varepsilon_P(t),

where \(b_P\) is a baseline signal, \(B_{\max}\) is the maximum modeled burden-related signal, \(K_P\) is the burden at half of that maximum, and \(\varepsilon_P(t)\) is measurement error. This function is increasing for nonnegative \(A\), but its sensitivity to changes in \(A\) declines as burden grows: its derivative is \(B_{\max}K_P/[K_P+A(t)]^2\). The formula is a useful possible observation model, not a claim that every tracer has this exact response.

Worked Example: Recognizing PET Signal Saturation

For illustration, let \(b_P=5\) signal units, \(B_{\max}=100\) signal units, and \(K_P=2\) burden units. The burden-related component is \(100A/(2+A)\).

At \(A=1\), the total predicted signal is \(5+100(1)/(2+1)=38.33\) units. At \(A=2\), it is \(5+100(2)/(2+2)=55.00\) units. At \(A=8\), it is \(5+100(8)/(2+8)=85.00\) units.

Increasing burden from 1 to 2 units raises the predicted signal by \(16.67\) units, while increasing burden from 2 to 8 units raises it by \(30.00\) units across a six-unit burden change. The average signal change per burden unit is therefore much smaller over the latter interval. A relatively flat PET trajectory at high signal may reflect limited sensitivity in the saturated range rather than stable aggregate burden.

Neurofilament: Injury-Related Release and Clearance

Neurofilament light (NfL) is often modeled as a marker related to axonal injury. It is not specific to one disease mechanism, and its measured concentration depends on release, transport into the sampled compartment, and clearance. A simple model for a CSF or blood concentration \(F(t)\) is

\frac{dF}{dt}=q_0+q_J J(t)-k_FF(t), \qquad k_F>0,

Here \(q_0\) is background input, \(q_J\) scales the injury-related input, and \(k_F\) is the elimination rate. The exact parameterization depends on the sampled compartment and should not be assumed identical for CSF and blood. If \(J(t)\) changes faster than \(F(t)\) can clear, the measured concentration can lag behind the injury process.

Derivation: If the total input \(q\) is constant, the model \(dF/dt=q-k_FF\) has equilibrium \(F_{\mathrm{eq}}=q/k_F\). Subtracting this equilibrium gives \(d(F-F_{\mathrm{eq}})/dt=-k_F(F-F_{\mathrm{eq}})\). Integrating yields \(F(t)=F_{\mathrm{eq}}+[F(0)-F_{\mathrm{eq}}]e^{-k_Ft}\). Thus the concentration approaches the new equilibrium exponentially, with half-life \(\ln(2)/k_F\), provided the constant-input, first-order-clearance assumptions hold.

Worked Example: NfL Remains Elevated After an Injury Episode

Consider a hypothetical NfL concentration with background input \(q_0=2\) ng/L/day and clearance \(k_F=0.25\) per day. Before an injury episode, the concentration is at background equilibrium: \(F(0)=q_0/k_F=2/0.25=8.00\) ng/L. For the next 4 days, the injury-related input adds 3 ng/L/day, so the total input is 5 ng/L/day and the new equilibrium is \(5/0.25=20.0\) ng/L.

During those 4 days, \(F(t)=20.0+(8.00-20.0)e^{-0.25t}\). At day 4, \(F(4)=20.0-12.0e^{-1}\approx15.59\) ng/L.

After day 4, suppose the injury-related input stops and only the background input remains. The concentration then approaches 8.00 ng/L. Four days after the episode ends, \(F(8)=8.00+[15.59-8.00]e^{-0.25(4)} \approx8.00+7.59e^{-1}\approx10.79\) ng/L.

The modeled injury-related input has ended, but NfL remains above baseline because concentration declines according to its clearance rate. In this example its half-life is \(\ln(2)/0.25\approx2.77\) days.

Interpreting Discordant Longitudinal Signals

When biomarkers disagree, first compare what each measures and its response timescale. A PET signal may track deposited aggregate burden; a soluble CSF analyte may reflect the balance of production, sequestration, and clearance; NfL may respond to injury flux and then decay over time. A synaptic integrity state may deteriorate without an immediate proportional change in any one of these assays. These are different observation channels, not interchangeable measurements of a single latent quantity.

Check the analyte and assay.
Confirm which molecular species or PET tracer is measured, the sampled compartment, units, and whether higher values are expected to indicate more or less of the relevant process.

Check timing and kinetics.
Compare sampling intervals with plausible production and clearance times. A lagging concentration should not be aligned automatically with the biological state on the same date.

Check the observation function.
Consider saturation, baseline signal, assay precision, and whether the measured species is expected to track burden, injury rate, or another state.

Check repeatability and model fit.
Review assay quality, missing or mistimed samples, and residual patterns. A systematic mismatch may indicate an inadequate source function, unmodeled delay, or a biological pathway absent from the model.

A mechanistic observation model is not the only option. A simple empirical model can describe association between a biomarker and time or clinical status with fewer assumptions, but it may not explain lag or predict changes under altered production and clearance. A latent-state model can combine several measurements, but its conclusions depend on the assumed links between states and assays and may be weakly identified when data are sparse. A direct one-to-one proxy assumption is simplest, yet is often inadequate when kinetics, saturation, or assay specificity matter.

These approaches have limits. A good fit does not establish that the proposed biological mechanism is correct. Several combinations of production, clearance, and state effects can generate similar concentration profiles. PET binding can be affected by tracer properties and technical factors; NfL can rise with injury from different causes; and CSF concentrations can be sensitive to sampling and compartment processes. Interpretation should therefore distinguish measured signal, model-derived latent state, and biological explanation.

Common Pitfalls and Model Checks

Treating every marker as a direct disease-burden measure. The correction is to specify the assay-to-state relationship and its expected direction before interpreting a trajectory.
Assuming a falling CSF concentration proves improvement. A soluble analyte can fall if it is sequestered or its source declines, even while aggregate burden rises.
Reading a flat PET trajectory as proof of stable pathology. Check whether the signal is near saturation and whether the change expected from the model exceeds measurement variability.
Equating NfL concentration with instantaneous injury. Release and clearance create a time-dependent concentration; use the kinetic model to account for persistence after injury input changes.
Using an equilibrium formula during a transient. \(C_{\mathrm{eq}}=q/k\) applies to constant input at equilibrium, not as a replacement for the observed starting concentration. For a changing input, use the trajectory with the correct initial condition.

Useful diagnostics include plotting each observed biomarker and its model-predicted trajectory on its own scale, checking residuals for systematic trends or time-dependent spread, and examining whether inferred kinetic parameters are plausible and estimable. If a marker consistently changes before the modeled driver, a simple downstream-release model may be inadequate. If a biomarker is near its assay range limit, apparent flattening may be technical rather than biological. Discordance is informative when it tests a model’s assumptions; it is not resolved by forcing all measures to move in the same direction.

Check Your Understanding

Answer each question using the stated model assumptions. Full solutions follow the questions.

A CSF analyte has constant input \(q=6\) concentration units per day and clearance \(k=0.3\) per day. What is its equilibrium concentration?
In the saturable PET model, what happens to sensitivity to aggregate burden as \(A\) becomes large? Give the mathematical reason.
An NfL model has \(q_0=1.5\) ng/L/day, \(k_F=0.3\) per day, and no injury-related input at baseline. What baseline equilibrium concentration is predicted?
Why might PET increase only slightly while aggregate burden continues to rise?
A soluble CSF measure decreases while NfL remains elevated after an injury episode. Give one kinetic or mechanistic explanation for each observation.

Solutions. (1) At equilibrium, \(C_{\mathrm{eq}}=q/k=6/0.3=20\) concentration units. (2) The derivative is \(B_{\max}K_P/(K_P+A)^2\), which approaches zero as \(A\) becomes large; the signal is less sensitive to additional burden. (3) \(F_{\mathrm{eq}}=q_0/k_F=1.5/0.3=5.00\) ng/L. (4) The PET observation function may be near saturation, so its slope with respect to burden is small. (5) A soluble species may decline if its source falls or it is consumed into aggregates; NfL can remain elevated because its concentration decays over time after injury-related input ends.