Flare–Remission Cycles in Relapsing Autoimmunity

From a disease-activity state to flare–remission dynamics

A relapsing disease is not described fully by its activity at one time point. The same measured activity may be rising toward a flare or falling toward remission, depending on the preceding inflammatory response and the state of counter-regulation. A useful disease-activity model must therefore represent both activation and a response that accumulates or changes more slowly.

The earlier material on representing disease activity as a mechanistic state and mapping cytokine dynamics to composite disease activity scores establishes the role of latent activity and its clinical readout. Here, the focus is the structure of a minimal dynamical system: inflammatory self-amplification, delayed resolution feedback, and external triggers. The model is a deliberately reduced representation, not a claim that one pair of states captures every autoimmune mechanism.

A minimal activation–resolution model

Let \(x(t)\) be a dimensionless latent activity coordinate and \(y(t)\) a dimensionless, delayed counter-regulatory state. Larger \(x\) maps to greater clinical activity; \(y\) represents the net suppressive influence of resolution processes, expressed as a deviation from a reference level. Let \(I(t)\) denote an external inflammatory trigger. One simple model is

The term \(x-x^3/3\) is a nonlinear activation relationship. Its positive slope around intermediate activity allows amplification, while the negative cubic term prevents unbounded growth. The \(-y\) term makes the resolution state suppress activity. The second equation makes \(y\) respond gradually to activity: when \(x>-a\), \(y\) increases; when \(x<-a\), it decreases. The parameter \(0<\varepsilon\ll1\) makes this feedback slower than the activity response, and \(a\) sets the activity level at which the feedback state stops changing.

The clinical score need not equal \(x\). For example, a bounded readout could be \(D(t)=D_{\min}+(D_{\max}-D_{\min})/[1+\exp(-x(t))]\), where \(D_{\min}\) and \(D_{\max}\) are the score bounds. This monotone mapping preserves the ordering of activity while keeping the score in range. The equations describe a latent mechanism; the readout equation describes how it is compared with observations.

This model assumes that the chosen states are sufficient for the question, that their parameters are constant over the modeled interval, that activation and feedback can be represented continuously, and that the trigger is specified separately. It also assumes no explicit delay: slow feedback is represented by its own state. These assumptions can be useful for reasoning about qualitative regimes, but they require scrutiny before interpreting fitted parameters biologically.

Equilibrium and local stability

First consider the untriggered system, \(I(t)=0\). An equilibrium requires both derivatives to be zero. From \(dy/dt=0\), \(x^*=-a\). Substituting this into \(dx/dt=0\) gives \(y^*=x^*-(x^*)^3/3\). Thus

To determine whether small deviations decay or grow, linearize the equations around this equilibrium. The Jacobian, the matrix of partial derivatives of the two rates with respect to the two states, is

For a two-dimensional linear system, local asymptotic stability requires a negative trace and a positive determinant. Here, \(\operatorname{tr}(J)=1-a^2\) and \(\det(J)=\varepsilon\). Since \(\varepsilon>0\), the determinant is positive. Therefore, the equilibrium is locally stable when \(a>1\), and linearly unstable when \(0<a<1\). At \(a=1\), the trace is zero and this linear stability test is inconclusive about nonlinear behavior.

This is a local result: it describes trajectories sufficiently close to the equilibrium. It does not, by itself, prove that every large trigger causes a flare or that every unstable equilibrium produces a particular cycle. The nonlinear trajectories must also be examined.

Three regimes and how triggers act

A remission regime has a stable low-activity equilibrium under baseline conditions. A transient trigger can displace the state, but if the system returns to equilibrium after the trigger ends, the flare is isolated rather than recurrent. The trigger must be large enough to move the trajectory into a substantial excursion; a stable equilibrium alone does not specify the trigger threshold.

An isolated-flare regime therefore combines a stable baseline with a sufficiently strong or sustained finite trigger. After the input ends, the slow feedback state continues to evolve and can help drive activity back down. Repeated flares in this regime require repeated triggers, or some other time-varying input; they are not autonomous cycles of the untriggered equations.

A recurrent-cycle regime is possible when the baseline equilibrium is unstable and the nonlinear dynamics keep states bounded. In the model above, \(0<a<1\) gives a positive trace and positive determinant, so nearby deviations grow rather than decay. For parameter values in the oscillatory region, the nonlinear system settles into a repeating trajectory: activity rises, delayed feedback builds, activity falls, and the feedback state later relaxes. The stability calculation identifies a candidate regime; numerical integration or additional nonlinear analysis is needed to establish the cycle and characterize its amplitude and period for a specific parameter set.

The distinction matters clinically and mechanistically. Regularly recurring activity does not by itself show that triggers recur: it may be generated by internal feedback. Conversely, a sequence of flares does not establish an autonomous oscillator if each flare could follow an external insult. Trigger histories and longitudinal state information are needed to distinguish these explanations.

Justification, diagnostics, and alternatives

The stability criterion follows directly from the characteristic polynomial of the equilibrium Jacobian. For a two-by-two matrix, that polynomial is \(\lambda^2-\operatorname{tr}(J)\lambda+\det(J)=0\). When the trace is negative and determinant positive, both eigenvalues have negative real parts; perturbations decay locally. When the determinant is positive but the trace is positive, the sum of the eigenvalues is positive, so both cannot have negative real parts. In this model the positive-trace equilibrium is unstable. This derivation supports local regime classification, but nonlinear integration is still necessary to assess flare size, cycle existence, and transient duration.

Several checks help identify when a fitted or simulated model is going wrong:

• Inspect the untriggered trajectory. After setting \(I(t)=0\), a model labeled as an isolated flare should return to its stable baseline; one labeled as an autonomous cycle should continue oscillating without further input.
• Check the feedback timescale. If \(\varepsilon\) is not small relative to the activity timescale, describing \(y\) as delayed feedback may be misleading. The equations may still be solvable, but the intended mechanism is less clear.
• Check for trigger–feedback confounding. A delayed rise in measured activity could be attributed to a new input, slow internal feedback, or an observation lag. The earlier material on time lags between cytokines and clinical activity is relevant when interpreting such timing.
• Check numerical robustness. Reduce the integration step size or tighten solver tolerances. A cycle that disappears under modest numerical refinement may be a numerical artifact.
• Check what the data can identify. Sparse observations may not distinguish an autonomous oscillation from several transient flares. A good fit to a composite activity score alone does not uniquely establish the underlying mechanism.

A static threshold model can classify activity as flare or remission at each time, but it cannot represent how the current state depends on prior activity. A trigger-only model can represent repeated external insults, but it cannot generate recurrent cycles when its input is constant. Adding an explicit time delay instead of a feedback state is another option when a discrete lag is mechanistically justified; it changes the mathematical problem and requires evidence to support that added structure. The two-state model is a compact middle ground, not a universally preferred model.

Important limitations remain. The cubic activation function is phenomenological, the feedback state aggregates potentially distinct biological mechanisms, and the trigger is not assigned to a specific cytokine or cell population. The model has no patient heterogeneity, measurement noise, or changing treatment and disease context. Its dimensionless parameters should not be interpreted as measured biological rates without an explicit scaling and estimation strategy. These limitations are reasons to test model structure against data, not reasons to treat a simulated cycle as a clinical prediction.