Psychedelics and schizophrenia: Distinct alterations to Bayesian inference

Rajpal and colleagues situate this study within two converging lines of interest: the recent scientific and clinical resurgence in research on classical serotonergic psychedelics, and the longstanding idea that some psychoactive drugs can produce “psychotomimetic” effects that resemble aspects of schizophrenia. Previous empirical work shows that both psychedelics (notably LSD) and dissociatives (notably ketamine) alter large-scale neural dynamics, including increases in measures of signal diversity and changes in inter-regional information transfer, and EEG work has reported similar increases in signal diversity in schizophrenia. However, a parsimonious account that explains both the commonalities and the divergences between drug-induced states and schizophrenia at a neurophysiological level remains lacking. This paper aims to compare resting-state neuroimaging data from healthy subjects under LSD and ketamine with EEG data from patients with schizophrenia, and to interpret the empirical patterns through a computational model grounded in the predictive processing (PP) framework. Specifically, the investigators test whether observed changes in Lempel–Ziv complexity (LZ, a measure of signal diversity) and transfer entropy (TE, a directed measure of information transfer) can be reproduced by perturbations to a single Bayesian state-space model, instantiated as changes in the precision of priors versus sensory inputs. The overarching goal is to refine psychotomimetic models of psychosis and to clarify how different disruptions to Bayesian inference map onto distinct neural signatures in resting-state activity.

The empirical component combined three datasets. Schizophrenia data were obtained from the BSNIP database and comprised 64-channel EEG recordings from 29 patients and 38 age-matched healthy controls, sampled at 1000 Hz during eyes-closed rest; demographic metadata, PANSS symptom scores, and number of antipsychotics per patient (used as a proxy for medication load) were available. Drug datasets came from prior studies: 17 subjects for LSD and 19 for ketamine, each scanned with a CTF 275-channel MEG system at 600 Hz in two eyes-closed sessions (drug and placebo). Subjects were selected to match an age range of 20–40 years. Preprocessing steps were harmonised across datasets when possible. Continuous recordings were epoched into 2 s segments, visually screened to remove gross artefacts, and cleaned using independent component analysis to remove muscle and ocular artefacts. Source reconstruction used a linearly constrained minimum variance (LCMV) beamformer to project sensor data onto centroids of Automated Anatomical Labeling (AAL) regions. Source time series were bandpass filtered between 1 and 100 Hz and downsampled (to 250 Hz for EEG and 300 Hz for MEG). AAL sources were grouped into five regions of interest (ROIs): frontal, parietal, occipital, temporal and sensorimotor. Two principal outcome metrics were computed. Lempel–Ziv complexity (LZ) was calculated on binarised source time series (values above the epoch mean mapped to 1, below to 0), resulting in an entropy-rate estimate per AAL source and then averaged per ROI. Transfer entropy (TE), estimated via state-space Gaussian models implemented with the MVGC toolbox, provided a conditional, directed 5×5 TE network between the ROIs, with each ROI treated as a multivariate vector of its constituent AAL sources. Statistical testing used within-subject one-sample t-tests for the drug datasets (drug versus placebo differences per subject) and linear models for the schizophrenia dataset, with LZ or TE as dependent variables and predictors including condition, age, gender and number of antipsychotics; quadratic age terms were considered and selected when a likelihood-ratio test favoured them. Multiple comparisons across TE edges were controlled using a directed-network adaptation of the Network-Based Statistic (NBS) cluster-permutation approach. A complementary computational model was implemented to interpret empirical findings. Building on predictive processing, the authors constructed a linear stochastic Bayesian state-space model comprising a low-level (sensory) variable s_t, a higher-level internal state h_t, top-down predictions ŝ_t = b h_t, prediction errors ξ_t, and two precision parameters: λ_p for prior precision and λ_s for sensory precision. The control model was calibrated using placebo visual cortex data via expectation–maximisation to estimate baseline parameters (a_con, b_con, λ_con^p, λ_con^s). Schizophrenia was modelled by increasing sensory precision (λ_s scaled by a factor η > 1), yielding stronger bottom-up influence and a more precise posterior; the drug condition was modelled by reducing prior precision (λ_p scaled down), consistent with the REBUS hypothesis. After parameter adjustments, a and b were retrained, and LZ of ξ_t and top-down TE (ŝ_t → ξ_t) were computed from the model-generated time series to compare with empirical results. Additional control variations of precision parameters were also explored.

Lempel–Ziv complexity increased robustly across all three datasets. Rajpal and colleagues report significant and widespread LZ increases for LSD and ketamine, and an increase in the schizophrenia cohort relative to controls that was most pronounced in frontal and parietal regions. A simple unadjusted comparison of patient versus control LZ showed no significant difference (t = -0.38, p = 0.70), but when linear models corrected for age, gender and number of antipsychotics, the number of antipsychotics had a negative association with LZ (β = -0.016, t = -2.3, p = 0.021). Using those covariate-corrected LZ values, a two-sample t-test between patients and controls revealed a significant difference (t = 3.4, p = 0.001). The authors emphasise that the schizophrenia LZ result is sensitive to medication adjustment and to the proxy used for dose, so it should be considered preliminary. Transfer entropy showed divergent patterns across conditions. Both LSD and ketamine produced widespread decreases in TE between most ROI pairs compared with placebo. By contrast, schizophrenia patients exhibited localised increases in TE relative to healthy controls and no decreases; these increases were predominantly frontal in origin and strongest in frontal→occipital connections, i.e. a front-to-back pattern. The negative correlation between antipsychotic count and some TE edges was small and did not survive multiple-comparison correction. Within the schizophrenia cohort, neither LZ nor TE differences correlated with PANSS symptom scores in the extracted analyses. The computational model reproduced the empirical pattern when precision parameters were perturbed in specific ways. Increasing sensory precision (model variant for schizophrenia) or decreasing prior precision (model variant for the drug state) both led to increased LZ in the simulated prediction-error signals. However, these two perturbations produced opposite effects on top-down TE: increased sensory precision yielded an increase in top-down TE, whereas decreased prior precision yielded a decrease. Control variations that moved precision parameters in the alternative directions (reducing sensory precision or increasing prior precision) did not reproduce the empirical findings. The authors therefore highlight that similar apparent strengthening of bottom-up influence can be implemented by different mechanistic parameter changes, and that TE changes alone do not unambiguously specify which precision parameter is altered.

The authors interpret their results as showing a shared increase in spontaneous neural signal diversity across psychotomimetic drug states (LSD and ketamine) and schizophrenia, accompanied by opposite changes in directed inter-regional information flow: decreases in TE under both drugs versus increases in TE in schizophrenia. Using a parsimonious predictive processing model, Rajpal and colleagues argue that both classes of state can be framed as favouring increased bottom-up influence over top-down priors, but realised through different mechanistic perturbations: psychedelic drugs via reduced precision-weighting of priors (consistent with the REBUS hypothesis), and schizophrenia via increased precision of sensory inputs. This asymmetric implementation explains how the same high-level bias (relative strengthening of sensory information) can produce divergent signatures in TE while yielding similar increases in signal diversity. They caution against interpreting TE changes as straightforward evidence for increased top-down control, because TE quantifies directed information flow but is agnostic about functional role. The modelling results show that strengthening bottom-up influence can produce either increases or decreases in top-down TE depending on whether prior or sensory precision is altered. The authors therefore recommend more nuanced, multidimensional approaches to characterise top-down versus bottom-up dynamics; decomposing TE into finer information modes is suggested as a promising direction. Several limitations are acknowledged. The computational model assumes weakened priors for schizophrenia, whereas other task-based work has argued for stronger priors to explain some hallucinations; the present model is a resting-state formulation and may not generalise to task contexts. Empirically, differences in imaging modality (MEG for drugs, EEG for schizophrenia), sampling rates, and experimental design (within-subject for drugs, between-subject for schizophrenia) complicate direct comparisons. Spatial resolution was limited to 60 AAL sources across five ROIs, which may mask finer-scale PP effects. Crucially, antipsychotic dosage data were unavailable, so the number of antipsychotics was used as an imperfect proxy; this proxy correlated negatively with PANSS positive scores and with some neural metrics, but confounding cannot be disentangled without detailed medication information. The authors also note that their statistical models were linear and might miss non-linear relationships between medication load and neural dynamics. For future work, the investigators propose richer hierarchical modelling that could bridge resting-state and task-based findings, finer-grained analyses of directed connectivity (including band-limited approaches), decomposition of TE into constituent information modes, and studies that link neural dynamics to symptom subtypes and medication dosage. They emphasise that more detailed datasets—including dosage and temporally resolved symptom measures—are needed to clarify how antipsychotics modulate neural dynamics and to characterise heterogeneity within the schizophrenia spectrum.