When a patient with schizophrenia starts an antipsychotic and their brain scans begin to look more like those of healthy people, two things could be happening. The illness could be receding, or the drug could be pushing brain activity around in ways that only partly overlap with the illness. Current longitudinal studies cannot tell these apart. They track "dynamic connectivity" (how the pattern of correlations between brain regions changes second by second during a resting scan) using one broadband signal, and they call any movement toward the healthy group "normalization." This proposal splits that movement in two, separately for each frequency band of the slow brain signal measured by functional MRI (fMRI, which tracks blood oxygenation). One part moves patients along the axis that separated them from controls before any drug was given. The other part moves them in some other direction. We predict that genuine normalization and drug effects occupy different frequency bands, and that the known partial normalizations reported in the literature are explained by this separation.
The motivating anomaly is concrete. In a prospective cohort of unmedicated patients, Risperidone Effects on Brain Dynamic Connectivity found that six weeks of risperidone normalized how long patients stayed in a "sparse" connectivity state (p=0.0115), but did not normalize the total fraction of time they spent there. A single broadband analysis offers no explanation for why one statistic recovers and its close sibling does not. Our hypothesis is that the two statistics are dominated by different frequency bands, and that one of those bands carries drug-induced change orthogonal to the disease difference.
Core Claim
We claim that treatment-related change in dynamic connectivity can be decomposed, band by band, into an on-axis component that reverses the baseline disease difference and an off-axis residual. The residual tracks drug exposure rather than clinical improvement, and it concentrates in bands that broadband methods mix together, especially the fast BOLD band above 0.1 Hz. Here BOLD is the blood-oxygen signal that fMRI records.
To make this precise, let $f_b$ be a vector of dynamic-connectivity features in band $b$. We define the disease axis $d_b = \bar f_b^{HC} - \bar f_b^{SZ,pre}$, the difference between the healthy-control mean and the unmedicated-patient mean at baseline. We define $\Delta_b$ as each patient's post-minus-pre change, minus the controls' own test–retest change over the same interval. The decomposition is
$$ \alpha_b = \frac{\langle \Delta_b, d_b\rangle}{\|d_b\|^2}, \qquad \rho_b = \frac{\|\Delta_b - \alpha_b d_b\|}{\|d_b\|}, $$
where $\alpha_b$ is the fraction of the disease gap closed in band $b$ and $\rho_b$ is the size of the change that points elsewhere. The claim is falsifiable through a double dissociation. Across patients, $\alpha_b$ should correlate with symptom improvement but not with dose. $\rho_b$ should correlate with dose, measured as chlorpromazine equivalents, but not with symptom improvement. The bands where each holds should differ.
The closest existing work does pieces of this. It does not separate the two components, and it does not resolve by frequency. The risperidone study and the first-episode follow-up of Dynamic functional connectivity and coupling analysis of triple networks and white matter functional networks track normalization longitudinally. That follow-up reported State-2 fractional window p=0.020 and dwell time p=0.019 after about five months. Both studies are broadband, and both count any shift toward controls as recovery. A joint time-frequency analysis of resting-state functional connectivity showed that schizophrenia-related dynamic coupling is confined to mid-range frequencies, but only in a cross-sectional cohort of chronically medicated patients, so drug and disease stay fused. Ada-FCN showed that learned sub-bands, including the fast BOLD components usually discarded as noise, carry diagnostic signal. It is temporally static within each band and ignores medication. That gap matters because Pascoal et al. showed a psychoactive drug alters cerebral blood flow enough to confound a PET marker, which makes a vascular origin for fast-band "diagnostic" signal plausible.
Research Program
The pipeline has three parts.
Band-resolved dynamic features. We first decompose each region's BOLD time series into fixed wavelet bands. The conventional slow-5 and slow-4 ranges (roughly 0.01–0.027 and 0.027–0.073 Hz) form the lower bands. A slow-3 band and a fast band above 0.1 Hz, up to the scanner's sampling limit, form the upper bands. We use fixed bands rather than Ada-FCN's learned ones so that the axes are comparable across cohorts and interpretable. As a secondary analysis we check whether learned bands change the conclusions.
Within each band we estimate time-varying coupling using a window whose length scales with that band's slowest period. This detail matters. A 44-second window, as used in the risperidone study, cannot represent coupling at 0.01 Hz, whose period is 100 seconds. Broadband dwell times from short windows are therefore implicitly dominated by the faster bands. Scaling the window per band addresses the window-sensitivity problem raised in the static-limitations literature.
A continuous state description. Rather than forcing k-means states, we project windowed connectivity matrices into a shared log-Euclidean tangent space. This is the same trick Latent Factor Gaussian Processes use to treat covariance matrices as vectors; we take the projection but not their full Gaussian-process model, which does not scale to hundreds of regions. From the resulting trajectories we compute occupancy density, persistence (a continuous analogue of dwell time), and trajectory speed. For comparability with prior work we also compute the classic k-means (k=3) fractional window and dwell time per band.
The decomposition, protected against the obvious artifact. Patients recruited when acutely ill will drift toward controls by regression to the mean even without treatment. To limit this, we estimate $d_b$ on one random half of patients and controls and compute $\Delta_b$ on the other half, repeating across splits. We subtract control retest drift, and we assess $\alpha_b$ and $\rho_b$ by permutation of pre/post labels. The off-axis direction estimated in one cohort is then tested as a "drug signature" in the others.
Data. The primary data are the two longitudinal cohorts named above. The first is the prospective risperidone monotherapy cohort: 34 unmedicated patients, 24 at six weeks, and 35 controls with 19 retested. The second is the first-episode cohort: 93 patients, 92 controls, and 39 followed on naturalistic antipsychotics. Neither is listed in the file as public, so obtaining access through the original authors is a prerequisite; an equivalent drug-naïve inception cohort would substitute. The chronically medicated 151-patient cohort from the joint time-frequency study serves as an out-of-sample test. HCP-Rest supplies healthy data for simulation, and ADNI supplies the data for reproducing Ada-FCN.
Baselines
Naive reference. The do-nothing reference is the prior studies' own readout. Broadband sliding-window correlation with k-means, as in the risperidone study, counts a patient as normalized when a state statistic moves toward controls. Reproduced on the same cohort, it should recover the published pattern: dwell time in the sparse state normalizes at p=0.0115 while fractional occupancy does not. Our method must explain that split, not merely restate it.
Linear dose covariate. The standard confound correction applies the same broadband analysis with chlorpromazine-equivalent dose as a linear covariate. We run it on identical preprocessing and identical subjects. It "loses" if its dose-adjusted normalization estimate still correlates with dose in held-out splits, or if it cannot separate symptom-linked from dose-linked change.
Strongest learned models. These are two classifiers used as normalization meters: a model's predicted probability of "patient" is tracked from pre to post treatment. The first is STAGIN, a dynamic graph network with spatial and temporal attention. It reports 88.20% accuracy (AUROC 0.9296) on HCP-Rest gender decoding, against 81.34% for a static graph network. The second is Ada-FCN, which reports 79.68% accuracy on ADNI.
We first reproduce Ada-FCN's ADNI number with the released protocol as a sanity check. We then train both models on baseline unmedicated patients versus controls with the same cross-validation folds and the same hyperparameter-search budget as our feature pipeline. We apply them to post-treatment scans. No reported result was found on Lacuna for any classifier used as a longitudinal normalization index, or for any frequency-resolved normalization metric. Beating these models therefore means one of two things. Either their score change correlates with dose after controlling for symptom change while our $\alpha_b$ does not, or our $\alpha_b$ predicts symptom improvement in held-out patients better than their score change.
Ablations. Two ablations isolate the key ingredients. Collapsing all bands into one tests whether frequency resolution is necessary. Replacing dynamic features with band-limited static connectivity tests whether dynamics are necessary.
Why not just…?
Why not just regress out chlorpromazine equivalents? A linear dose term assumes the drug adds a fixed, dose-proportional shift to every connectivity feature. The pharmacological literature in the file argues against this: antidepressants alter physiological coupling independently of illness severity (Zhao et al.), and sedation biases EEG connectivity (Choi et al.). More basically, in a longitudinal design dose is collinear with time on treatment, so regressing it out also removes true recovery. The decomposition does not need dose to estimate its components; dose is used only afterward to validate them. That makes it testable where regression is not. The dose-covariate baseline is run directly, so if it achieves the same dissociation, that will show.
Why not just compare band-limited static connectivity before and after? This is the real threat. It is simpler, and Choi et al. found drug-naïve disease effects with static band-wise connectivity alone: lower-alpha wPLI of 0.241 versus 0.276, p=0.024. Against it, the risperidone study showed that static abnormalities across seven component pairs were in fact confined to one transient state, and the dwell-time-versus-occupancy split is by definition invisible to static features. The static ablation settles the question. If static band features give the same $\alpha_b$/$\rho_b$ dissociation, we will report that the frequency axis, not dynamics, carries the result, and simplify the method accordingly.
Why not just track band power (fALFF or spectral density) instead of connectivity? Drug effects on vascular tone could appear purely as amplitude changes, and power is trivial to compute. Two findings cut against relying on power alone. Choi et al. found no power difference where connectivity differed, and Miller et al. showed that opposite-signed changes across bands cancel when bands are averaged. We include per-band amplitude as a covariate and as a competing feature set. If $\rho_b$ in the fast band is fully explained by amplitude change, that is itself evidence for a vascular drug artifact rather than neuronal coupling, which is the answer to the file's fourth gap.
Why not just use a trained classifier's patient score? It is the most convenient single number, but it is trained only to separate groups, so any drug-induced movement that happens to lower the score counts as recovery. We run this directly as the strongest baseline, as described above.
Proposed Experiments and Decisive Checks
Simulation with known ground truth. On HCP-Rest scans we plant a "disease" coupling perturbation in one band and a "drug" perturbation in another, with a controllable angle between them, and we simulate pre/post sessions. The decomposition should recover planted $\alpha_b$ and $\rho_b$ with bias shrinking as the angle widens. Broadband k-means should blur the two. This establishes when the method can and cannot work: we expect it to fail when drug and disease directions are nearly parallel.
Main longitudinal test. In the risperidone cohort, we test whether sparse-state dwell time and fractional occupancy load on different bands. We also test whether the non-normalizing occupancy corresponds to large $\rho_b$ in some band. The double dissociation, $\alpha_b$ with symptom change and $\rho_b$ with dose, is the decisive result.
We are candid about power: with 24 patients at six weeks, only large effects are detectable. We therefore pre-register the band-level hypotheses. The first is that fast-band change is mostly off-axis. The second is that mid-range change, where the joint time-frequency study localized disease coupling, is mostly on-axis. We then replicate the direction of both in the 39-patient first-episode follow-up.
Transfer test. The off-axis drug direction learned in one cohort is projected onto the chronically medicated 151-patient cohort. If part of that cohort's patient–control difference aligns with the drug direction, its "disease" signature is partly pharmacological. That would give a concrete explanation for the static-literature inconsistency reported by Jensen et al..
Failure conditions. The proposal fails if $\alpha_b$ and $\rho_b$ show no band dependence, or if the dose-covariate or static baselines reproduce the dissociation equally well. Either outcome is informative and would be reported as such.
References
- Risperidone Effects on Brain Dynamic Connectivity—A Prospective Resting-State fMRI Study in Schizophrenia
- Dynamic functional connectivity and coupling analysis of triple networks and white matter functional networks in first-episode schizophrenia patients: mechanisms revealed by follow-up studies
- A joint time-frequency analysis of resting-state functional connectivity reveals novel patterns of connectivity shared between or unique to schizophrenia patients and healthy controls
- Ada-FCN: Adaptive Frequency-Coupled Network for fMRI-Based Brain Disorder Classification
- Monoamine oxidase B inhibitor, selegiline, reduces 18F-THK5351 uptake in the human brain
- Modeling Dynamic Functional Connectivity with Latent Factor Gaussian Processes
- Learning Dynamic Graph Representation of Brain Connectome with Spatio-Temporal Attention
- Cardiorespiratory Coupling Analysis Based on Entropy and Cross-Entropy in Distinguishing Different Depression Stages
- Electroencephalographic Resting-State Functional Connectivity of Benign Epilepsy with Centrotemporal Spikes
- Spectral Changes in Cortical Surface Potentials during Motor Movement
- Aberrant cortical–subcortical-cerebellar connectivity in resting-state fMRI as an imaging marker of schizophrenia and psychosis: a systematic review of data-driven whole-brain functional connectivity analyses