Defusion as a Half-Life: Timing How Long Distorted Thoughts Keep Their Grip in Sparse Diaries

Therapies such as Acceptance and Commitment Therapy target cognitive fusion, which is the tendency to treat a thought like "I will always fail" as a fact rather than as a passing mental event. The goal is defusion: the thought may still come, but it stops steering mood and behavior. This theory makes a claim about time. A fused person's distorted thought keeps affecting them for hours or days. A defused person has the same thought and recovers quickly. Current language models of mental health ignore that claim. They detect cognitive distortions, which are recognizable reasoning errors such as fortune telling or "should" statements, in isolated snippets of text, and then correlate how often distortions occur with symptom scores.

We propose to measure fusion directly as a per-person decay rate. For each person, we estimate how quickly the mood impact of a distorted diary entry fades with real elapsed time. We compare that to how quickly the impact of an equally negative entry without a distortion fades. The key difficulty is the data. Diaries arrive irregularly and voluntarily, sometimes hours apart and sometimes days apart. Mood prompts are frequently skipped. The field's usual way of handling this is to treat consecutive entries as equally spaced steps, and that choice erases exactly the time information the construct depends on.

Core Claim

Fusion has a measurable signature in sparse, irregularly timed text-plus-mood records. We define it as a person-specific distortion half-life: the elapsed time over which the mood perturbation that follows a distorted entry falls by half. We predict three things. First, this half-life can be recovered reliably despite realistic sparsity. Second, it is longer than the half-life of matched non-distorted negative entries in more severe individuals. Third, it predicts later severity beyond what distortion frequency predicts.

The contribution is a measurement, not a large new architecture. We deliberately fit the simplest continuous-time model that can express the construct. We then check whether heavier irregular-time models add anything.

Where This Sits in Prior Work

Three papers motivate the gap by coming close to it while each discarding a different piece.

Linking Language-based Distortion Detection to Mental Health Outcomes trains RoBERTa distortion detectors on the Thinking Traps and Patient Queries datasets and applies them to the DS4UD daily diaries of service workers. It shows that distortion rates track depression and emotional drain. However, it collapses the diaries into aggregate correlations, so it measures how often distortions happen, not how long they last. We reuse its detector and ask the temporal question.

Acute suicidal ideation in context follows 268 outpatients with depression for 90 days, with three prompts a day and 5,938 diary entries. It finds sub-daily linguistic markers of escalating and de-escalating suicidal ideation. It organizes the data into "triplicates" of consecutive responses, which ignores whether two responses were four hours or five days apart, and it drops incomplete sequences.

Emotion-based Modeling of Mental Disorders on Social Media shows that healthy Reddit users quickly return to a neutral emotional state after a perturbation, while clinical users stay stuck. This is the closest existing analogue of our construct. It is estimated with Markov transitions inside a fixed 30-minute window, over coarse emotions rather than distortions, so it cannot separate "stuck because of the thought" from "stuck in general."

Our proposal keeps Guo's recovery idea, uses Varadarajan's detectors, and applies both to the Lekkas-style data in real time.

The method comes from the irregular time-series literature. Temporal Belief Memory lets confidence in an old observation decay exponentially with elapsed time. That design beat mean substitution for septic shock prediction, with AUC 0.7840 versus 0.7078. We borrow its exponential decay form but make the decay rate the scientific quantity of interest rather than a nuisance parameter. Neural Controlled Differential Equations are neural models whose hidden state evolves continuously in time, driven by the incoming observations. They supply two things here: a heavier continuous-time competitor, and the trick of adding an observation-count channel so that the rate of logging becomes informative.

One result constrains the design. Computational network models for forecasting and control of mental health trajectories found that linear vector autoregression and Kalman filters often failed to beat a person's historical mean on EMA forecasting (EMA, or ecological momentary assessment, means short smartphone surveys during daily life). A nonlinear state-space model, the piecewise-linear recurrent network (PLRNN), did better. Our core model is linear in its decay kernels, so this is a genuine warning. It is the reason we benchmark against the PLRNN and a Neural CDE rather than assuming a parsimonious model is adequate.

The Model

Let $y_i(t)$ be person $i$'s self-reported negative mood at time $t$, observed only when they answer a prompt. Let $s_k$ be the timestamps of their diary entries. Each entry receives two scores:

  • a distortion score $d_k$, the detector probability, residualized on negativity;
  • a negativity score $n_k$, from a sentiment engine.

We model

$$ y_i(t) = \mu_i + \sum_{k:\, s_k < t} \left( \gamma_i\, n_k\, e^{-\lambda_i (t - s_k)} + \beta_i\, d_k\, e^{-\kappa_i (t - s_k)} \right) + u_i(t) + \varepsilon, $$

where the terms are defined as follows:

  • $\mu_i$ is the person's baseline mood.
  • $\gamma_i$ and $\lambda_i$ are the size and decay rate of a generic negative-affect perturbation.
  • $\beta_i$ and $\kappa_i$ are the size and decay rate of the distortion-specific perturbation.
  • $u_i(t)$ is a continuous-time mean-reverting noise process whose variance grows naturally across long gaps.
  • $\varepsilon$ is measurement noise.

The distortion half-life is $\ln 2/\kappa_i$. The fusion index is $\lambda_i/\kappa_i$. Values above one mean distorted content lingers longer than equally negative undistorted content.

Residualizing $d_k$ on $n_k$ is what makes $\kappa_i$ interpretable as a property of the thought's structure rather than its tone. Person-level parameters are partially pooled in a hierarchical Bayesian fit, because many individuals contribute only a few dozen entries. Because the noise process $u_i$ spreads over real time, long silences automatically widen uncertainty. We therefore do not need a separate evidential layer such as EDICT to express "we have not heard from this person in three days."

Logging is not random, and entries become more verbose near peaks of ideation. We handle this in two ways. We include the recent entry rate as a covariate on $\beta_i$. We also run a sensitivity analysis that jointly models entry times as a self-exciting point process, so that we can see whether the half-life estimates move.

Research Program

The work proceeds in three stages.

Stage 1: simulation. We simulate individuals with known $\kappa_i$ and $\lambda_i$, using entry schedules resampled from real diary timestamps. This tells us how many distorted and non-distorted entries per person are needed to recover the fusion index. It also quantifies how badly two common shortcuts bias the estimates: triplicate-style equal spacing and daily binning.

Stage 2: fitting and validation. We fit the model to DS4UD and, subject to data access, to the Lekkas cohort, then test the construct three ways:

  • Reliability: split-half agreement of $\kappa_i$ between the first and second halves of each person's record.
  • Convergent validity: association of half-life with baseline severity.
  • Incremental validity: whether half-life predicts later severity after controlling for distortion rate and negativity rate.

Stage 3: replication at a different timescale. We replicate the qualitative contrast on the Reddit histories from Guo et al.. There, the "mood" outcome is the negativity of subsequent posts and the contrast is between diagnostic groups and controls. This replication is weaker, because the outcome is text-derived, but it spans years rather than weeks.

Detector choice. We use the Thinking Traps plus Patient Queries RoBERTa detector, which reports span F1 0.833 in Varadarajan et al.. We prefer it over few-shot LLMs, which reached only F1 0.51 on C2D2. Detector error attenuates $\beta_i$ but should not systematically distort $\kappa_i$. We verify this in the simulation by injecting label noise at the reported error rate.

Baselines

Naive reference. Each person's historical mean mood, and last-observation-carried-forward. The network-models paper showed that linear models frequently failed to beat the mean, so any model must clear this bar on held-out mood prompts.

Distortion-frequency model. This is the aggregate approach of Varadarajan et al.: a mixed-effects regression of later mood and severity on each person's distortion rate and negativity rate. It is the strongest competitor on the validity question, because if frequency already carries everything, half-life is redundant. It uses the same detector and the same person splits. Beating it means a significant increment in explained variance in later severity from adding half-life.

Discrete-step model. A lagged mixed-effects model in which "time since entry" is replaced by entry index, mimicking the triplicate approach. This serves as the ablation that isolates real elapsed time as the key ingredient.

PLRNN. The strongest forecaster from Computational network models, which reports MAE 0.831 and 0.795 on its second and third micro-randomized trial samples. Those numbers come from different data and are not directly comparable. We found no reported result on Lacuna for forecasting on DS4UD or the Lekkas cohort, so we will retrain it ourselves. We give it the same diary scores as extra inputs and the same hyperparameter search budget as our model.

Neural CDE. A Neural CDE with the observation-intensity channel, run on the same inputs with a one-sided (causal) interpolation so that it cannot see the future.

Shared-decay ablation. Our model with $\kappa_i = \lambda_i$ forced. This isolates the distortion-specific decay.

On forecasting, our model does not need to win. It needs to be within the PLRNN's and Neural CDE's error range, specifically within their bootstrap confidence intervals of held-out MAE. Beating the frequency model on incremental validity and beating the shared-decay ablation on held-out likelihood are the results that matter.

Why not just…?

Why not just feed diary embeddings into a Neural CDE or ContiFormer and let it learn the dynamics? It may forecast mood slightly better, and we will report that honestly. But neither model produces a per-person quantity that clinicians or trials can use. ContiFormer's demonstrated advantages are also on dense benchmarks, not on a few thousand entries spread over hundreds of people. If the Neural CDE clearly beats our model on held-out MAE, meaning outside the bootstrap interval, we will conclude that our kernels are misspecified. In that case we will extract an analogous decay estimate from the Neural CDE's impulse responses, in the spirit of the relative cumulative impulse response tracing used with the PLRNN, rather than defend the parametric form.

Why not just use distortion frequency, which already correlates with outcomes? This is the most serious threat. The incremental-validity test answers it directly. If half-life adds nothing beyond frequency on DS4UD and the Lekkas data, the temporal reading of fusion is unsupported at these sampling densities, and we will report a negative result.

Why not just add elapsed time as a feature in an ordinary lagged mixed-effects model? Done carefully, with an elapsed-time interaction per content type, such a model essentially becomes our model under a different name, and we would welcome that. The contribution is the specific contrast between distortion-driven and negativity-driven decay, together with the evidence that it is identifiable and reliable. The simulation stage will show whether the simpler parameterization recovers the same fusion index.

Why not just use Guo's neutral-reset Markov chain? Its fixed 30-minute window cannot represent recovery across gaps of hours to days, which is the typical spacing in diaries. It also models general emotion rather than distortion content. We will fit it on the Reddit replication, compare its reset rates with our $\lambda_i$, and check whether our distortion-specific $\kappa_i$ explains group differences beyond those reset rates.

Decisive Checks

The proposal succeeds only if all four of the following hold:

  • Simulations show the fusion index is recoverable at realistic entry densities, and that equal-spacing methods bias it.
  • Split-half reliability of half-life is at least moderate.
  • The model with separate decay rates beats the shared-decay ablation on held-out likelihood.
  • Half-life predicts later severity beyond distortion frequency.

Failure of any one of these narrows or retires the claim that defusion can be read off time in naturalistic text.

Two limitations are known in advance. Access to the Lekkas cohort is uncertain, and DS4UD's daily cadence may be too coarse to resolve sub-daily half-lives. If so, the Lekkas data are necessary for the sub-daily claim, and DS4UD supports only day-scale conclusions.