Authors: Maya L. Foster (1)Department of Biomedical Engineering, Yale University), Milana Khaitova (2)Department of Radiology & Biomedical Imaging, Yale School of Medicine), Saloni Mehta (2)Department of Radiology & Biomedical Imaging, Yale School of Medicine), Jean Ye (3)Interdepartmental Neuroscience Program, Yale School of Medicine), Dustin Scheinost (1)Department of Biomedical Engineering, Yale University; 2)Department of Radiology & Biomedical Imaging, Yale School of Medicine; 3)Interdepartmental Neuroscience Program, Yale School of Medicine; 4)Department of Statistics & Data Science, Yale University; 5)Yale Child Study Center, Yale School of Medicine; 6)Yale Biomedical Imaging Institute, Yale School of Medicine)
Categories: Article
Source: medRxiv
Authors: Maya L. Foster, Milana Khaitova, Saloni Mehta, Jean Ye, Dustin Scheinost
Conduct a mega-analysis of two complementary measures of resting-state functional magnetic resonance imaging (rsfMRI) dynamics—amplitude of low-frequency fluctuation (ALFF) and low-frequency spectral entropy (lfSE)—in mood and psychosis-spectrum disorders to evaluate group differences and clinical symptom associations.
ALFF and lfSE were calculated at the node-level by filtering data from 0.01 Hz to 0.08 Hz, regressing demographic variables, and harmonizing sites. Group differences were assessed using the Wilcoxon signed test. Symptom associations were evaluated with Spearman’s rho. Analyses were conducted at both whole-brain and network levels, with sensitivity analyses to evaluate the impact of frequency brands.
Four independent open-source case-control datasets with resting-state functional magnetic resonance imaging were the Center for Biomedical Research Excellence, the Human Connectome Project for Early Psychosis, the Strategic Research Program for Brain Sciences, and the UCLA Consortium for Neuropsychiatric Phenomics.
Included participants had a mood disorder (bipolar, dysthymia, or major depressive disorder, n=228, aged 38.31 ± 12.56 years), a psychosis-spectrum disorder (early psychosis, schizophrenia spectrum disorder, or mood disorder with psychotic symptoms, n=318, aged 29.8 ± 13.21 years), or a healthy control (n=535, aged 39.89 ± 15.3 years).
To identify group differences and symptom associations in mood and psychosis-spectrum disorders using ALFF and lfSE.
ALFF in psychosis-spectrum was significantly lower than mood disorders and controls (q’s<0.001) at the whole-brain and network levels. lfSE in controls was significantly lower than both psychosis-spectrum and mood disorders at the whole-brain and network levels (q’s<0.001). Whole-brain ALFF is positively associated with mood symptoms (rho=0.27, p<0.05). Whole-brain lfSE is negatively associated with positive (rho=−0.13, p<0.05) and mood (rho=−0.38, p<0.01) symptoms. A greater sensitivity of group differences and symptom associations to frequency ranges was observed in mood disorders. ALFF is sensitive to medication.
Widespread, global differences in ALFF and lfSE underly psychosis-spectrum and mood disorders. lfSE may be applicable for wider use in fMRI. Differences in spectral measures of brain dynamics may represent shared and distinct markers of mental health.
Mood (e.g., major depressive disorder and bipolar disorder) and psychosis-spectrum (e.g., first-episode psychosis and schizophrenia) disorders cause significant disability globally^1–4^ and current treatment outcomes remain suboptimal. These disorders also have shared and unique genetic risks^5–11^. Thus, continuing to advance knowledge of their shared and unique mechanisms will be important for improving differential diagnosis^11,12^ and symptom management for affected individuals. Widely associated with cognition and mental health^13–16^, changes in brain dynamics may be linked to the pathophysiology of mood and psychosis-spectrum disorders. However, while applications of brain dynamics methods are becoming increasingly popular, there is a scarcity of work exploring their simultaneous application to mood and psychosis-spectrum disorders, raising questions of whether linked dynamics signatures are shared, distinct, or both.
Spectral measures succinctly characterize brain dynamics across frequency bands^17–19^. One widely used metric is the amplitude of low-frequency fluctuations (ALFF), which quantifies the signal strength in the 0.01-0.08 Hz range and is well-established in functional neuroimaging^20–24^. A comparable measure is low-frequency spectral entropy (lfSE)—a frequency-bounded derivative of spectral entropy that quantifies the complexity or unpredictability of fMRI signal fluctuations in the 0.01–0.08 Hz range. Spectral entropy has been broadly used in EEG^24^, but not in fMRI studies. ALFF and lfSE provide distinct, but complementary insights (amplitude versus complexity) into intrinsic brain dynamics that may help biologically differentiate disorders and mediate individual differences in symptom severity.
Prior work has identified ALFF differences across major functional networks in mood^25–27^ and psychosis-spectrum^28–30,27^ disorders. In schizophrenia, studies found reduced ALFF in the default mode (DMN), salience (SAL), sensorimotor, parietal, and visual networks compared to healthy controls^31^. Early-stage schizophrenia has been marked by increased ALFF in subcortical and visual networks and decreased ALFF in the DMN and parietal networks^31^. In contrast, chronic schizophrenia includes widespread increases in ALFF across frontal, SAL, temporal, and limbic networks, alongside reductions in sensorimotor, parietal, and occipital areas^31^. In depression, studies report increased ALFF in several regions—including lateralized subcortical areas, occipital lobe, limbic areas, frontoparietal network (FPN), and the DMN—compared to matched healthy controls^26,32–35^. Decreased ALFF is also observed in somatosensory cortex^25,36,32^, limbic areas, and cerebellum^37,38^ for major depressive disorder (MDD) compared to controls. In bipolar disorder, increased ALFF in the right caudate and putamen^39^, bilateral insula, medial prefrontal cortex, and decreased ALFF in the left cerebellum^40^ have been observed. In contrast, lfSE has not been studied in mood and psychosis-spectrum disorders using fMRI. Moreover, to our knowledge no previous work has comprehensively contrasted ALFF with lfSE in mood and psychosis-spectrum disorders in one study.
Mega-analyses (e.g., a method that pools together the raw, individual level data from multiple studies), are a promising approach for finding shared and distinct biological indices across disorders. By pooling data from several smaller studies, mega-analyses maximize population variance, increase generalizability, and improve statistical power^41,42^. ALFF and lfSE are ideal for these analyses because they are computationally tractable in large samples, easy to pool across diverse datasets due to their conceptual simplicity, and neurobiologically interpretable at different spatial scales (i.e., regional, network, or whole brain levels). By better accounting for psychiatric heterogeneity, mega-analyses analyses can support the development of treatments that target shared and distinct clinical mechanisms^43–45^.
Here, we perform a mega-analysis in over 1000 individuals to compare ALFF and lfSE in individuals with mood disorders, individuals with psychosis-spectrum disorders and healthy controls. A secondary objective was to assess the utility of lfSE compared to ALFF. Our study expands on previous empirical findings by adding a complexity measure (i.e., lfSE)^46^, using a large-scale sample size (n=1081), and comparing measures across a wider range of symptoms. We also assess their association with relevant symptom measures. Based on findings from previous works, we hypothesized that ALFF and lfSE case-control group differences will be in the FPNl^47,48^, SAL^47^, and DMN^47,49^.
We used four publicly available resting-state fMRI datasets (see Table S1 for participant demographics), the Center for Biomedical Research Excellence^50^ (COBRE, n=99), the Human Connectome Project Early Psychosis^51^ (HCP-EP, n=169), the Strategic Research Program for Brain Sciences^52^ (SRPBS, n=609) Multi-disorder Connectivity Dataset, and the University of California Los Angeles (UCLA) Consortium for Neuropsychiatric Phenomics^53^ (CNP, n=204). The sample included 535 healthy controls (HCs), 228 mood disorder (major depressive disorder, bipolar disorder, and dysthymia), and 318 psychosis-spectrum disorder (early psychosis and schizophrenia spectrum disorder) participants. All patients were diagnosed according to the DSM-5^54^. Demographics, symptom scores, medication information, and diagnosis breakdowns are summarized in the Supplement methods and tables S1, S5, and S6.
Psychosis symptoms were assessed with the Positive and Negative Syndrome Scale^55^ (PANSS), which measures the presence and severity of positive, negative, and general psychopathology for an individual within the past week. The scale is widely used in clinical psychosis studies^56^ and has demonstrated reliability in assessing psychopathology of schizophrenia populations^57^. 230 participants had PANSS scores (HCP-EP=107, SRPBS=123). Mood symptoms in individuals (n=62) with a mood disorder (SRPBS=39) and health controls (SRPBS=23) were assessed with the Beck’s Depression Inventory^58^ (BDI-II). The BDI-II measures hallmark symptoms of depression and has high internal consistency^58,59^.
The acquisition and imaging parameters for the datasets are detailed elsewhere^50,51,60,52^. However, an abridged version is available in the Supplement (see Table S2). Images were motion and slice time corrected using SPM12^61^. BioImage Suite was used to perform mean white matter regression, cerebral spinal fluid regression, and gray matter time, removing linear trends, and low-pass filtering^62^. The Shen-268 atlas^61^ was warped from Montreal Neurological Institute (MNI) space into single-subject space. Next, the mean-time course of each node was calculated by averaging its constituent voxels’ time-series data.
Participants were excluded if they had an average frame-to-frame displacement exceeding 0.2 mm (n=28), insufficient quality of linear or nonlinear registration (n=16), had >100 missing nodes (n=10) or had less than 75% coverage of relevant nodes for a given network or whole-brain analysis (n=15, see Supplement for more details on data preservation process). This left a final sample of 1081 subjects for further 823 for whole-brain analyses (averaging across all nodes) and varying sample sizes for network analyses (Table S3–4).
First, the power spectral density of each node’s time series was calculated using periodogram function from the SciPy ^63^ signal processing API in Python 3.9. This calculation entailed converting each participant’s node time-series into the frequency domain via the fast Fourier transform (FFT). Then, the power spectral density was integrated over the low-frequency range (0.01 Hz ≤ frequency ≤ 0.08 Hz) using SciPy’s trapezoid integration method to calculate total power. Next, we took the square root of that operation to quantify ALFF.
Note: f=frequency and df=delta frequency
ALFF in this study is considered a proxy for the intensity of neural-associated brain activity or the energy content of the low-frequency band. A high ALFF signifies greater oscillation intensity while a low ALFF signifies lower oscillation intensity in the low-frequency range. ALFF was assessed group-wise at the whole-brain by averaging all nodes and network-levels using 10 canonical networks from the Shen 268 atlas.
lfSE was also calculated at the node-level. Individual node time-series underwent a band-pass filter in the low-frequency range (0.01 Hz ≤ frequency ≤ 0.08 Hz) using SciPy’s butter and filtfilt functions (order = 2)^21^. The signal was transformed into the frequency domain using FFT, followed by power spectral density calculation and normalization to a probability density function. Then, lfSE was calculated as the Shannon entropy of the power spectral density was calculated using AntroPy^64^ package in Python 3.
Note: f=frequency, Fs=sampling frequency, numf= number of frequencies derived by the FFT
As a complexity measure, high lfSE (flatter power spectral density spectrum) represents low complexity while low lfSE (power spectral density distribution with more peaks) represents high complexity. lfSE was assessed group-wise at the whole-brain by averaging all nodes and network-levels using 10 canonical networks from the Shen 268 atlas.
Statistical analyses were constrained to the whole-brain and network levels to capture macroscale brain dynamics. Broader scale such as these have greater power than granular levels of inference^65^. Group comparisons of each low-frequency brain measure were evaluated at p<0.05 using the Wilcoxon two-sided rank sum test while correcting for multiple comparisons with false discovery rate (FDR). We controlled for self-reported age, motion, and sex as variables of non-interest. First, these factors were regressed^66^ using a generalized linear model on ALFF and lfSE values at the node level. Second, we applied the neuroCombat^67^ R package to harmonize data across sites. For empirical ranges, please refer to table S5). Where relevant, we report corrected (q) and uncorrected (p) results for transparency but only interpret significant corrected results (q<0.05).
We correlated symptoms scores (PANSS and BDI-II) and covariate-regressed ALFF and lfSE values using Spearman’s correlation at the whole-brain and network levels. Participants for the PANSS analyses came from multiple sites and underwent site harmonization using neuroCombat^67^. Participants for the BDI-II analyses came from one site, so harmonization was not applied. Network-level associations were considered statistically significant at q<0.05 after FDR correction. Additionally, we evaluated similarity between the ALFF and lfSE by correlating these values across participants (p<0.05) using Spearman’s rho at the whole-brain and network levels. We also calculated these similarities independently for each group. These correlations were Fisher transformed and compared across groups using a two-sample z-test.
We used different frequency cutoffs (0.01-0.045 Hz and 0.035-0.09 Hz) to assess if finer spectral divisions were driving the identified trends and associations^68^. We also examined the effect of medication exposure by including current medication status (binary: 1=on, 0=off) as an additional covariate.
Significant group differences were found for self-reported sex (χ^2^=18.081, df = 2, p<0.001), age (W=75.48, df=2, p<0.001) and average frame displacement (W= 8.599, df=2, p=0.014). Participants with psychosis-spectrum (29.8 ± 13.21 years) were significantly younger than mood disorders (38.31 ± 12.56 years) and controls (39.89 ± 15.3 years). Participants with mood disorders also had significantly higher motion frame displacement (0.087 ± 0.045 mm) compared to matched HCs (0.083 ± 0.046 mm) and those with psychosis-spectrum disorder (0.075 ± 0.037 mm).
At the whole-brain level, mood disorders exhibited the greatest ALFF, followed by controls and then psychosis-spectrum (Figure 2A, Table 1). Psychosis-spectrum ALFF values were significantly lower than both mood disorders and controls (p’s<0.001). No differences in whole-brain ALFF were detected between mood disorders and controls. Similar differences were observed at the network level, where all networks showed significant (q’s<0.05, FDR-corrected) differences when comparing psychosis-spectrum to mood disorders or health controls (Table 1). Results were similar when potential outliers were removed (Table S8) and when removing psychosis-spectrum patients with known affective disorders (Table S9).
At the whole-brain level, psychosis-spectrum had the highest lfSE, followed by mood disorders and then HCs (Figure 2B). Controls were significantly lower than both psychosis-spectrum and mood disorders at the whole-brain (p’s<0.001). No differences in whole-brain lfSE were detected between psychosis-spectrum and mood disorders. At the network level, all networks exhibited significant differences between health controls and psychosis-spectrum (q’s<0.05, FDR-corrected). The medial frontal, FPN, motor, subcortical, and cerebellar networks exhibited significant differences between HCs and mood disorders (q’s<0.05, FDR-corrected). While no differences between psychosis-spectrum and mood disorders existed at the whole-brain level, every network (other than the cerebellar network) exhibited significant between-group differences (q’s<0.05, FDR-corrected), likely due to the increases sample size used in the network analyses (Table 2). Results were similar when removing psychosis-spectrum patients with affective disorders (Table S10).
The psychosis-spectrum lfSE distribution was bimodal. Since the second (higher) peak in the distribution may influence the results, we repeated the analyses after removing these high-leverage points (see Supplemental Methods, Figure S4). Results were similar (Table S11); however, mood disorders now showed the highest lfSE, and only visual networks remained significantly different between psychosis-spectrum and mood disorders (q’s<0.05, FDR-corrected).
Whole-brain ALFF is positively associated with BDI-II scores (rho=0.36, p= 0.0046; Figure 2C, Table S12) but had no association with PANSS scores (Table S13). Additionally, whole-brain lfSE is negatively associated with positive PANSS (rho=−0.13, p=0.0418, Table S14) and BDI-II scores (rho=−0.38, p=0.002; see Figure 2D–E, Table S15). No other significant associations with ALFF or lfSE at the whole-brain or network-level were observed (Tables S12–15).
At all scales of analysis, ALFF and lfSE were negatively correlated, such that individuals with higher ALFF values had lower lfSE values (rho’s = −0.39 to −0.22, Table S16–17). This relationship was also observed in our group comparison, where ALFF was consistently greater in mood disorder compared to psychosis-spectrum across scales and lfSE was consistently greater in psychosis-spectrum compared to mood disorder across scales.
Results at the finer-scale frequency ranges were broadly similar to the main results, with wide-spread differences observed between groups (Tables S18–25). However, results involving the mood disorder group exhibited the most changes in finer-scale frequency ranges. For example, the lfSE of the DMN was significantly different between controls and mood disorders at 0.035-0.090 Hz, but not 0.01-0.045 Hz or the full frequency range. Additionally, whole-brain ALFF was positively associated with BDI-II scores at 0.01-0.045 Hz but negatively associated at 0.035-0.09 Hz (Table S22).
Results after controlling for medication were broadly similar for lfSE, but not ALFF (Table 3, Table S26). Only comparisons involving the mood disorder group were significant for ALFF. The ALFF comparisons between controls and psychosis spectrum were not significant.
In this study, we conducted a mega-analysis using two spectral measures—oscillation intensity (ALFF) and complexity (lfSE) —to compare mood and psychosis-spectrum disorders. Our results highlight extensive differences between these two clinical conditions that span multiple dimensions of brain dynamics.
Notably, ALFF and lfSE uniquely differentiated clinical groups, providing distinct insights into the biological underpinnings of mood and psychosis-spectrum disorders. Psychosis-spectrum participants exhibited larger, widespread differences in lfSE compared to mood disorder participants and controls, while mood disorder participants exhibited more focal differences in lfSE compared to controls. Also, in line with several studies ^25,26,32–38^, we observed significant differences between mood disorder participants and controls in ALFF after controlling for medication. In some cases, ALFF and lfSE results even converged. For example, despite our initial hypotheses, there is widespread overlap in the functional networks that had significant group differences between mood and psychosis-spectrum disorders.
Our sensitivity analyses revealed that ALFF results in mood disorders vary by frequency band. In the narrower range of 0.035 to 0.090 Hz, larger group differences and significant associations with mood symptoms were discovered. Moreover, there was a reversal in the sign of significant mood symptom associations in the two finer-scale frequency bands (0.01-0.045 Hz and 0.035-0.090 Hz). Aligned with these findings, a previous study showed that associations between depressive symptoms and spectral measures in the subgenual gyrus differ between frequency bands^69^. These diverging associations are evidence of the functional specificity of frequency bands in the blood-oxygen level dependent (BOLD) signal^70–72^, a phenomena that has largely been attributed to EEG studies of mental health^73^. As such, the notion that different frequency bands reflect physiology occurring at a range of intrinsic time scales^73–76^ is likely a universal principle of brain function. Under this consideration, the finer bands represented in our work comprise multiple physiological processes that contribute to mood disorder pathology. If too broad a frequency range is adopted, interactive effects^69^ can dominate and obscure significant results in a competitive or mutually constructive way^77^. Continued investigations of the functional relevance of specific frequency bands in mood disorders and other psychiatric conditions are needed in fMRI research.
Out of the two measures, lfSE performed the best in detecting significant group differences and symptom associations. It was also robust to common sources of variance, which is crucial for mega-analyses, as they often encompass real-world conditions (e.g., use of different scanners, participants with varying medication histories). lfSE may more effectively capture changes in brain dynamics because as a measure of complexity and irregularity^78^, it considers the entire energy distribution across a frequency band. Previous work employing similar methods in other imaging modalities have identified signal irregularity and high signal variance as key features of the psychosis-spectrum^79–82^. Our results, alongside broader literature findings, position lfSE—a standard measure in the EEG literature^24^ and related modalities^78,83–85^—as a valuable index for increased application in fMRI. Future research should apply the lfSE metric under different conditions and test its robustness against different covariates to fully assess its mechanistic insights and resilience to non-interest variance sources.
In contrast, we observed a change in the significance of two group difference pairs in ALFF analyses with and without medication (Table 3). This inconsistency may indicate that ALFF is more susceptible to medication effects, or more specifically, medication effects in mood and psychosis-spectrum disorders. Future studies should further probe ALFF’s sensitivity to medication in other psychiatric populations to confirm.
This study has several strengths. We combined distinct diagnoses into broader groupings (e.g., mood and psychosis-spectrum disorders), which may better reflect the dimensional nature of disorders and allow datasets with different diagnostic criteria to be combined in a mega-analysis. We also used two spectral measures, including lfSE (which is uncommon in fMRI analysis). It also has several limitations. First, participants were assumed to exclusively have mood or a psychosis-spectrum disorder. However, comorbidities are common, and individuals could belong in multiple groupings ^86^. Second, we focus on measures that compress brain dynamics into a single number. More complex summaries of dynamic measures^87^ may prove to be more sensitive and is a natural next step. Third, although we identify associations with symptoms scores, causality cannot be established.
Through a mega-analysis, we investigated group differences using spectral measures of brain dynamics in mood and psychosis-spectrum disorders. Our approach revealed distinct brain signal properties in addition to shared characteristics between the two patient groups. Future studies should continue to elucidate common denominators as well as diagnosis-, symptom-, and individual-specific brain deviations that build on top of disorder commonalities. Better characterization of shared brain features, and disorder specific deviations could pave the way towards plausible prevention and tailored treatment efforts.