Authors: Jason da Silva Castanheira, Alex I. Wiesman, Justine Y. Hansen, Bratislav Misic, Sylvain Baillet, John Breitner, Judes Poirier, Sylvain Baillet, Pierre Bellec, Véronique Bohbot, Mallar Chakravarty, Louis Collins, Pierre Etienne, Alan Evans, Serge Gauthier, Rick Hoge, Yasser Ituria-Medina, Gerhard Multhaup, Lisa-Marie Münter, Natasha Rajah, Pedro Rosa-Neto, Jean-Paul Soucy, Etienne Vachon-Presseau, Sylvia Villeneuve, Philippe Amouyel, Melissa Appleby, Nicholas Ashton, Daniel Auld, Gülebru Ayranci, Christophe Bedetti, Marie-Lise Beland, Kaj Blennow, Ann Brinkmalm Westman, Claudio Cuello, Mahsa Dadar, Leslie-Ann Daoust, Samir Das, Marina Dauar-Tedeschi, Louis De Beaumont, Doris Dea, Maxime Descoteaux, Marianne Dufour, Sarah Farzin, Fabiola Ferdinand, Vladimir Fonov, Julie Gonneaud, Justin Kat, Christina Kazazian, Anne Labonté, Marie-Elyse Lafaille-Magnan, Marc Lalancette, Jean-Charles Lambert, Jeannie-Marie Leoutsakos, Laura Mahar, Axel Mathieu, Melissa McSweeney, Pierre-François Meyer, Justin Miron, Jamie Near, Holly NewboldFox, Nathalie Nilsson, Pierre Orban, Cynthia Picard, Alexa Pichet Binette, Jean-Baptiste Poline, Sheida Rabipour, Alyssa Salaciak, Matthew Settimi, Sivaniya Subramaniapillai, Angela Tam, Christine Tardif, Louise Théroux, Jennifer Tremblay-Mercier, Stephanie Tullo, Irem Ulku, Isabelle Vallée, Henrik Zetterberg, Vasavan Nair, Jens Pruessner, Paul Aisen, Elena Anthal, Alan Barkun, Thomas Beaudry, Fatiha Benbouhoud, Jason Brandt, Leopoldina Carmo, Charles Edouard Carrier, Laksanun Cheewakriengkrai, Blandine Courcot, Doris Couture, Suzanne Craft, Christian Dansereau, Clément Debacker, René Desautels, Sylvie Dubuc, Guerda Duclair, Mark Eisenberg, Rana El-Khoury, Anne-Marie Faubert, David Fontaine, Josée Frappier, Joanne Frenette, Guylaine Gagné, Valérie Gervais, Renuka Giles, Renee Gordon, Clifford Jack, Benoit Jutras, Zaven Khachaturian, David Knopman, Penelope Kostopoulos, Félix Lapalme, Tanya Lee, Claude Lepage, Illana Leppert, Cécile Madjar, David Maillet, Jean-Robert Maltais, Sulantha Mathotaarachchi, Ginette Mayrand, Diane Michaud, Thomas Montine, John Morris, Véronique Pagé, Tharick Pascoal, Sandra Peillieux, Mirela Petkova, Galina Pogossova, Pierre Rioux, Mark Sager, Eunice Farah Saint-Fort, Mélissa Savard, Reisa Sperling, Shirin Tabrizi, Pierre Tariot, Eduard Teigner, Ronald Thomas, Paule-Joanne Toussaint, Miranda Tuwaig, Vinod Venugopalan, Sander Verfaillie, Jacob Vogel, Karen Wan, Seqian Wang, Elsa Yu, Isabelle Beaulieu-Boire, Pierre Blanchet, Sarah Bogard, Manon Bouchard, Sylvain Chouinard, Francesca Cicchetti, Martin Cloutier, Alain Dagher, Samir Das, Clotilde Degroot, Alex Desautels, Marie Hélène Dion, Janelle Drouin-Ouellet, Anne-Marie Dufresne, Nicolas Dupré, Antoine Duquette, Thomas Durcan, Lesley K. Fellows, Edward Fon, Jean-François Gagnon, Ziv Gan-Or, Angela Genge, Nicolas Jodoin, Jason Karamchandani, Anne-Louise Lafontaine, Mélanie Langlois, Etienne Leveille, Martin Lévesque, Calvin Melmed, Oury Monchi, Jacques Montplaisir, Michel Panisset, Martin Parent, Minh-Thy Pham-An, Jean-Baptiste Poline, Ronald Postuma, Emmanuelle Pourcher, Trisha Rao, Jean Rivest, Guy Rouleau, Madeleine Sharp, Valérie Soland, Michael Sidel, Sonia Lai Wing Sun, Alexander Thiel, Paolo Vitali
Categories: Articles, Arrhythmic brain activity, Brain-fingerprinting, Magnetoencephalography, Movement disorders, Neural dynamics, Oscillations, Parkinson’s disease
Source: eBioMedicine
Authors: Jason da Silva Castanheira, Alex I. Wiesman, Justine Y. Hansen, Bratislav Misic, Sylvain Baillet
Research in healthy young adults shows that characteristic patterns of brain activity define individual “brain-fingerprints” that are unique to each person. However, variability in these brain-fingerprints increases in individuals with neurological conditions, challenging the clinical relevance and potential impact of the approach. Our study shows that brain-fingerprints derived from neurophysiological brain activity are associated with pathophysiological and clinical traits of individual patients with Parkinson’s disease (PD).
We created brain-fingerprints from task-free brain activity recorded through magnetoencephalography in 79 PD patients and compared them with those from two independent samples of age-matched healthy controls (N = 424 total). We decomposed brain activity into arrhythmic and rhythmic components, defining distinct brain-fingerprints for each type from recording durations of up to 4 min and as short as 30 s.
The arrhythmic spectral components of cortical activity in patients with Parkinson’s disease are more variable over short periods, challenging the definition of a reliable brain-fingerprint. However, by isolating the rhythmic components of cortical activity, we derived brain-fingerprints that distinguished between patients and healthy controls with about 90% accuracy. The most prominent cortical features of the resulting Parkinson’s brain-fingerprint are mapped to polyrhythmic activity in unimodal sensorimotor regions. Leveraging these features, we also demonstrate that Parkinson’s symptom laterality can be decoded directly from cortical neurophysiological activity. Furthermore, our study reveals that the cortical topography of the Parkinson’s brain-fingerprint aligns with that of neurotransmitter systems affected by the disease’s pathophysiology.
The increased moment-to-moment variability of arrhythmic brain-fingerprints challenges patient differentiation and explains previously published results. We outline patient-specific rhythmic brain signaling features that provide insights into both the neurophysiological signature and symptom laterality of Parkinson’s disease. Thus, the proposed definition of a rhythmic brain-fingerprint of Parkinson’s disease may contribute to novel, refined approaches to patient stratification. Symmetrically, we discuss how rhythmic brain-fingerprints may contribute to the improved identification and testing of therapeutic neurostimulation targets.
Data collection and sharing for this project was provided by the Quebec Parkinson Network (QPN), the Pre-symptomatic Evaluation of Novel or Experimental Treatments for Alzheimer’s Disease (PREVENT-AD; release 6.0) program, the Cambridge Centre for Aging Neuroscience (Cam-CAN), and the Open MEG Archives (OMEGA). The QPN is funded by a grant from 10.13039/501100000156Fonds de Recherche du Québec - Santé (FRQS). PREVENT-AD was launched in 2011 as a $13.5 million, 7-year public-private partnership using funds provided by 10.13039/100008582McGill University, the FRQS, an unrestricted research grant from 10.13039/100013739Pfizer Canada, the Levesque Foundation, the Douglas Hospital Research Centre and Foundation, the 10.13039/501100000023Government of Canada, and the Canada Fund for Innovation. The Brainstorm project is supported by funding to SB from the NIH (R01-EB026299-05). Further funding to SB for this study included a Discovery grant from the 10.13039/501100000038Natural Sciences and Engineering Research Council of Canada of Canada (436355-13), and the CIHR Canada research Chair in Neural Dynamics of Brain Systems (CRC-2017-00311).
Research in contextEvidence before this studyRecent research with healthy young adults has shown that certain brain activity features are remarkably stable over extended periods, enabling the definition of so-called individual brain-fingerprints. However, this stability is reduced in neurological disorders, raising questions about the underlying neurophysiological causes and the clinical utility of brain-fingerprinting.Added value of this studyOur study introduces a brain-fingerprinting method using spontaneous brain electrophysiological activity, recorded during rest through magnetoencephalography, from Parkinson’s disease patients compared to age-matched healthy controls. This method revealed the origins of variable brain activity patterns in Parkinson’s patients and identified stable rhythmic features that distinguish individuals by their unique brain-fingerprints. We show that our approach accurately differentiates Parkinson’s patients from healthy individuals with an accuracy of about 90%, pinpointing rhythmic activity in sensory and motor cortical areas as key indicators of Parkinson’s disease and symptoms. Furthermore, we found that the brain-fingerprint cortical topography in Parkinson’s patients corresponds with the distribution of specific neurochemical systems related to the disease.Implications of all the available evidenceOur findings provide a detailed account of how cortical brain activity is altered in Parkinson’s disease. They show that individualized brain-fingerprints can be established for each patient, offering the potential for monitoring disease progression and tailoring treatment strategies. This study underscores the value of brain-fingerprinting in understanding neurological disorders and enhancing patient care.
The neurophysiological underpinnings of Parkinson’s disease (PD) are characterized by a spectrum of motor and non-motor symptoms that vary widely among patients, and their fundamental nature continues to be a subject of extensive research.1, 2, 3 This variation in symptoms parallels PD’s diverse structural alterations,4, 5, 6, 7 along with changes in hemodynamic and electrophysiological brain activity compared to healthy individuals.^1^^,^8, 9, 10, 11, 12 Notably, electrophysiological changes in PD involve both the rhythmic (oscillatory) and arrhythmic (background 1/f) components of neurophysiological signals.^8^^,^^9^^,^12, 13, 14, 15 Brain-network characteristics, as highlighted in previous studies using functional connectome analysis with functional magnetic resonance imaging (fMRI) and other brain mapping techniques, also deviate from those in healthy individuals and correlate with PD’s hallmark motor and cognitive impairments.16, 17, 18, 19
Recent methodological advances have employed fMRI connectomes to derive brain-fingerprints, providing biometric differentiation based on individual neuroimaging phenotypes.20, 21, 22, 23 This concept posits that an individual's neuroimaging phenotype remains relatively stable over time, forming the basis for distinctive brain-fingerprints.^20^^,^^21^^,^^24^ Such brain-fingerprinting has enabled the exploration of the neurophysiological bases of complex traits and behaviors in healthy subjects.20, 21, 22, 23, 24, 25, 26
However, subsequent studies have reported greater variability of brain-fingerprints in clinical populations,^25^^,^^27^^,^^28^ which challenges inter-individual differentiation based on brain-fingerprints in patients affected by movement disorders29, 30, 31 and mental health disorders.^27^^,^^28^ A recent study with magnetoencephalography (MEG), examined brain-fingerprints of patients with PD derived from beta-band (13–30 Hz) connectomes and found that differentiation accuracy declined with the severity of the motor symptoms.^29^ The authors attributed this effect to the reduced self-similarity of the connectome brain-fingerprints of patients. We are not aware of any previously published studies of brain-fingerprints derived from hemodynamic signals in Parkinson’s disease. A comprehensive understanding of these effects in terms of their rhythmic and arrhythmic components remains unexplored.
Neurophysiological brain activity comprises rhythmic (oscillatory) and arrhythmic (1/f) components, each with distinct associations to behaviour,32, 33, 34, 35, 36, 37 and disease symptoms.^15^^,^^38^^,^^39^ These two components may also reflect interdependent physiological mechanisms.^40^ For instance, computational modeling and empirical evidence suggest that arrhythmic activity relates to the balance of excitatory vs. inhibitory currents in local circuits.^34^^,^40, 41, 42
The greater longitudinal variability of brain-fingerprints in Parkinson’s disease may result from more pronounced instability of patients' brain activity over short periods of time. For instance, previous work has shown that hemodynamic signals from functional near-infrared spectroscopy (fNIRS) are more variable in patients with severe PD symptoms.^43^ This suggests that electrophysiological activity in PD is also likely to exhibit greater temporal variability, especially in brain regions with strong coupling between electrophysiological and hemodynamic signals.^44^
This short-term variability within patients challenges the definition of a stable brain-fingerprint profile that accurately characterizes an individual’s disease stage. Further research is needed to determine whether this increased variability affects the entire frequency spectrum of electrophysiological activity or is confined to specific rhythmic or arrhythmic components.^32^^,^^40^
In a recent study with healthy young adult participants, we demonstrated that frequency-specific measures of electrophysiological activity across the cortex, derived from brief, task-free MEG data, define spectral brain-fingerprints that are unique to each individual over remarkably prolonged periods.
In the present study, we extend this approach and confirm that the electrophysiological brain-fingerprint of patients with PD exhibits greater variability over time compared to that of healthy controls. This variability is predominantly driven by the arrhythmic component of the neurophysiological power spectrum. In contrast, rhythmic features of the PD brain-fingerprint remain remarkably stable, enabling effective differentiation between PD patients and healthy controls, as well as among patients themselves. We discuss the unique longitudinal variability in arrhythmic and rhythmic neurophysiological components of the PD brain-fingerprint, suggesting they are influenced by distinct physiological mechanisms. Additionally, we underline the clinical relevance of the brain-fingerprint’s stable rhythmic features, particularly their correlation with symptom laterality. We also show that the brain-fingerprint’s distinctive features align with the cortical topography of specific neurotransmitter systems, highlighting aspects of PD neuropathophysiology.^45^
Participants for this study were selected from a diverse age group (40–82 years) and included healthy controls as well as patients with mild to moderate idiopathic Parkinson’s Disease (PD). We aggregated data from multiple sources. We aggregated data from 79 PD patients who were part of the Quebec Parkinson Network (QPN; https://rpq-qpn.ca^46^). These patients had undergone extensive clinical, neurophysiological, and biological profiling. All enrolled patients in the QPN study were on a stable dose of antiparkinsonian medication and demonstrated satisfactory clinical responses. They were instructed to continue their medication regimen as prescribed before any data collection. We included data from QPN participants who had complete and usable MEG, (275 channels whole-head CTF; Port Coquitlam, British Columbia, Canada) clinical, and demographic data.
Our main control group comprised demographically matched participants from the PREVENT-AD (N = 50)^47^ and OMEGA (N = 4)^48^ studies, ensuring a comparison group that mirrors the age and demographic characteristics of the PD group. Participants self-reported their biological sex. We replicated our observations using a second sample of healthy controls from the Cambridge Center for Aging Neuroscience (Cam-CAN) dataset (N = 370 healthy adults, 40–78 years old, mean age = 58.67, SD = 11.04; 185 Females) recorded on a different MEG instrument. See “Cam**-**CAN sample of healthy controls” below for more details.
All participants underwent resting-state eyes-open MEG recordings. With the exception of the Cam-CAN dataset, these recordings were conducted using a 275-channel whole-head CTF system (Port Coquitlam, British Columbia, Canada) at a sampling rate of 2400 Hz, with a 600-Hz antialiasing filter. We also applied the system's built-in third-order gradient filters to the recordings. Consistency in data collection was maintained by conducting all recordings at the same site, each lasting a minimum of 10 minutes.
Note that the MEG data can be assumed to be missing at random. The main contributors to participant exclusion (i.e., metal implants, dental work, non-compliance, and claustrophobia) are expected to be equally common in both our patient and healthy control groups.
Our sample sizes were primarily dictated by the amount of data available from the QPN,^46^ PREVENT-AD^47^ and OMEGA^48^ datasets that met our inclusion criteria. At the time of data aggregation and preprocessing for the present study, 103 individuals had provided MEG data as part of the QPN initiative. After exclusions, 79 patients with PD and 10 healthy control participants were included from this cohort.
Of the 124 participants in the PREVENT-AD cohort who underwent an MEG recording, 21 were excluded from our study due to poor data quality. From the remaining sample (N = 103), we selected participants to match the demographics of the QPN sample (see Sample Matching Procedure).
All 228 participants from the OMEGA sample underwent an MEG scan. We excluded those whose recordings did not last a minimum of 10 minutes. After this exclusion step, we selected participants to match the demographic makeup of the QPN PD participant group, rsulting in a final sample of N = 4.
First, we identified all potential participants with Parkinson’s disease without comorbid pathology from the QPN database who had provided MEG, clinical, and demographic data. This resulted in a sample of 79 patients.
Next, we retrieved from the PREVENT-AD and OMEGA repositories participants who (1) matched the demographics of the patient sample and (2) met the inclusion criteria. Preference for inclusion was then given to individuals (1) of the same biological sex toward which the PD group was biased (i.e., male) and (2) closest to the mean level of education of the PD group. See Supplemental Table S1 for a summary of the sample’s demographics.
For the Cam-CAN cohort, we aimed to include a larger and more demographically diverse sample than the PREVENT-AD and OMEGA controls. Therefore, we included all individuals from this cohort aged 40 to 78, reflecting the age range of the initial healthy control sample.
The procedures for the collection, curation and analysis of all data reported in this study were reviewed and approved in accordance with the institutional ethics policies of McGill University's and the Montreal Neurological Institute’s Research Ethics Boards (ref no. 2021–7536). In compliance with the Declaration of Helsinki, written informed consent was obtained from each participant following a detailed description of the study.
The funders of this project played no role in study design, data collection, analysis, decision to publish, or manuscript preparation.
We preprocessed the MEG data using Brainstorm^49^ (March-2021 distribution) on MATLAB 2019b (Mathworks, Inc., Massachusetts, USA). We adhered to established good practice guidelines^50^ and replicated the following preprocessing steps as applied in previously published studies on similar data.^15^^,^^51^
We filtered the MEG sensor signals between 1 and 200 Hz to minimize slow-wave drifts and high-frequency noise. We then removed line noise artifacts at 60 Hz and harmonic frequencies using a notch filter bank. Cardiac and ocular artifacts were corrected using Signal-Space Projectors (SSPs) derived from electrocardiogram and electrooculogram recordings, through an automated procedure in Brainstorm.^49^ We segmented the MEG recordings into non-overlapping 6-second epochs and downsampled them to 600 Hz. Finally, we screened and excluded data segments with peak-to-peak signal amplitude or maximum signal gradient exceeding 3 absolute deviations from the median across all epochs.
We derived brain source models from each participant’s individual T1-weighted MRI data. We segmented and labeled the MRI volumes using FreeSurfer.^52^ We coregistered the MEG data to these segmented MRIs using approximately 100 head points digitized on the day of the MEG sessions. For 14 PD patients and 3 controls who lacked usable MRI data, we used BrainStorm procedures to warp the default FreeSurfer anatomy to match their available head digitization points and anatomical landmarks.
We created biophysical head models for each participant using the Brainstorm overlapping-spheres model with default parameters. The MEG cortical maps comprised 15,000 elementary dipole sources, constrained to the cortical surface, with free orientation. We computed source maps for each participant and each 6-second epoch using dynamic statistical parametric mapping (dSPM) with Brainstorm’s default parameters. To model environmental noise statistically, we applied the same processing approach to 2-minute empty-room recordings collected around the time of each participant’s visit.
Each vertex of the cortical surface (i.e., cortical location) consisted of three elementary time courses. We performed a principal component analysis (PCA) and retained the first local principal component to obtain a single source time series per cortical location. We then grouped the resulting 15,000 time series—one for each vertex—into the 68 regions of the Desikan-Killiany cortical parcellation.^53^ To do this, we extracted the first principal component of all elementary source time series within each parcel, yielding one representative time series per cortical region.
We derived brain-fingerprints from the power spectrum of the ROI source time series. We calculated the Power Spectral Density (PSD) for each parcel using Welch’s method, with a 3-second time window and 50% overlap. This approach yielded PSDs in the frequency range of 0–150 Hz, with a frequency resolution of 1/3 Hz.
Each individual’s spectral brain-fingerprint was composed of the PSDs of all 68 cortical parcels, averaged across all 6-second epochs. As detailed in Results, we derived two sets of spectral brain-fingerprints based on epochs from either the first or second half of the entire MEG session recordings. Additionally, we generated spectral brain-fingerprints from shorter data segments comprising 30-second non-overlapping segments.
Each spectral brain-fingerprint comprised a total of 68 × 451 features. We performed subsequent analyses using in-house developed code in Python (version 3.7.6) and R (version 4.2.1).
We replicated a previously published fingerprinting approach based on the correlational differentiability of participants between data segments (as illustrated in Fig. 1a–b).^24^ For each participant, we calculated all Pearson correlation coefficients between their first spectral brain-fingerprint and the second brain-fingerprint of every individual in the same cohort, including the participant being analyzed. The fingerprinting process involved a simple lookup along the rows or columns of the symmetrical inter-individual correlation matrix, where the highest correlation coefficient in this matrix indicated the matching participant.Fig. 1Brain-fingerprinting pipeline and study design. (a) For each participant, the power spectral density of MEG source time series for each cortical parcel of the Desikan-Killiany atlas is estimated from two data segments (1 and 2), each approximately 4 minutes long. The resulting power spectra across all cortical parcels generate one brain-fingerprint for each data segment (b-fp1 and b-fp2). The similarity between two brain-fingerprints, measured with cross-correlation statistics, produces an inter-individual confusion matrix. The diagonal elements of this matrix represent the self-similarity (Iself) between each participant’ two consecutive brain-fingerprints, while the off-diagonal elements represent the other-similarity (Iother) of a participant’s brain-fingerprint with those from the other participants in the cohort. (b) We tested the brain-fingerprinting approach in three inter-individual differentiation i) between healthy controls, ii) between patients with Parkinson’s Disease (PD), and iii) between patients and healthy controls. (c) We derived an individual differentiability score for each participant based on the self-similarity of their two consecutive brain-fingerprints in relation to their other-similarity with the rest of the participants. The score of individual differentiability is calculated by z-scoring the self-similarity score against the other-similarity scores.
We repeated this approach for all participants in the cohort, generating an inter-individual confusion matrix based on the two instances of their respective brain-fingerprints. We determined the overall differentiation accuracy of the brain-fingerprinting procedure by calculating the percentage of correctly differentiated individuals.
We defined individual differentiability as the ability to distinguish a participant from others in the cohort based on their brain-fingerprint. This measure is derived as the self-similarity (Iself^25^) of an individual’s two consecutive brain-fingerprints, z-scored against the mean and standard deviation of the other-similarity (Iother^25^) of the participant’s brain-fingerprint with those from the other participants in the cohort.
To establish confidence intervals for the average differentiation accuracy scores obtained from the fingerprinting procedure, we employed a bootstrapping method across the tested cohorts. This involved randomly selecting a subset of participants constituting 90% of the cohort and performing brain-fingerprinting on their data to obtain a differentiation accuracy score for that subset.
We repeated this process 1000 times, each time with a different random subset of participants from the cohort. From the empirical distribution of these differentiation accuracies, we derived a 95% confidence interval using the 2.5th and 97.5th percentiles.
We examined the potential impact of environmental noise and biophysical recording artifacts on the differentiation of individual participants. To do this, we correlated individual differentiability scores with the root-sum-squares (RSS) values of ocular, cardiac, and head movement signals that were recorded simultaneously with MEG. These signals included data from electrocardiogram (ECG), horizontal electrooculogram (HEOG), vertical electrooculogram (VEOG), and head-coil triplet channels.
We analyzed the correlations between these three measures and the individual differentiability of each participant from the entire cohort. Additionally, we included the head motion RSS measure as a nuisance covariate in our regression model, which explored the relationship between the self-similarity (Iself) of brain-fingerprints and increasing gap durations.
To assess whether environmental and instrument noise, which can vary daily, could have biased individual differentiation, we utilized the empty-room recordings collected alongside each MEG session. From these recordings, we derived “mock brain-fingerprints” for each participant by projecting the empty-room time series data onto the cortical source maps. We then calculated the differentiation accuracies obtained by utilizing these mock brain-fingerprints to differentiate the true observed brain-fingerprints, following the same procedure as described above.
To evaluate the respective contributions of arrhythmic and rhythmic spectral components to individual differentiation, we parameterized (decomposed) each individual’s brain-fingerprint spectral features using specparam (Brainstorm version) over the 2–40 Hz frequency range. The parameters for specparam peak width limits between 0.5 and 12 Hz, a maximum of 3 peaks, a minimum peak amplitude of 3 arbitrary units (a.u.), a peak threshold of 2 standard deviations, a proximity threshold of 2 standard deviations, and a fixed aperiodic mode.
The specparam outcomes comprise a set of Gaussian curves representing each spectral peak as a narrowband rhythmic component and a broadband arrhythmic component modeled as a Lorentzian function b -log(F^α^), where b is the broadband offset and α is the slope of the 1/f arrhythmic spectral component.^32^
The arrhythmic brain-fingerprint comprised the set of 1/f components across all cortical parcels.
Each Gaussian component was defined by three its center frequency (in Hz), spectral width (standard deviation around the peak, in Hz) and the peak amplitude (arbitrary units). However, we did not use these rhythmic components to define the rhythmic brain-fingerprint because the number of Gaussian peaks varies between participants. This results in a sparse spectral distribution with many frequency bins containing unassigned, noninformative spectral power values. Instead, we defined the rhythmic brain-fingerprint using the broadband residual obtained by subtracting the arrhythmic component from the original power spectrum at each cortical parcel. This approach yields a rhythmic brain-fingerprint that is defined over the entire frequency spectrum and can be readily compared across individuals.
We then conducted the same brain-fingerprinting analyses as previously described, applying them separately to both arrhythmic and rhythmic brain-fingerprints.
We quantified the contribution of each cortical region to individual differentiation using intraclass correlations coefficients (ICC). ICC assess the agreement between two measures; in this context, they indicate how consistent a particular brain-fingerprint feature is across the two brain-fingerprints of each individual compared to others in the cohort. A higher ICC for a given brain-fingerprint feature implies greater consistency across an individual’s brain-fingerprints relative to the cohort.
To illustrate the saliency of these features, we created ΔICC maps, as shown in Fig. 3a and Supplemental Fig. S3. We first averaged the ΔICC values within each of the canonical frequency bands and then averaged these across all bands. This process involved averaging ΔICC within the following frequency delta (1–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), beta (13–30 Hz), and gamma (30–50 Hz). This resulted in six ΔICC maps, one for each frequency band, which were then averaged to obtain a broadband ΔICC map.
The rationale behind this method was to give equal weight to each frequency band in deriving the broadband ΔICC, regardless of their respective bandwidths. For example, while the delta band spans 4 Hz, the high-gamma band covers 100 Hz. This approach ensures a balanced representation of all frequency bands in assessing the contribution of cortical regions to individual differentiation based on brain-fingerprint features.
We tested whether the spatial divergence between the rhythmic brain-fingerprints of patients and controls (see the ΔICC cortical maps in Fig. 3a) would replicate with using another cohort of age-matched healthy controls. For this, we used healthy-control data from the Cambridge Center for Aging Neuroscience (Cam-CAN) repository, consisting of resting-state, eye-closed MEG recordings collected using a 306-channel VectorView MEG system (Elekta Neuromag, Helsinki). We selected a Cam-CAN data subsample of 370 healthy adults aged 40 to 78 years, which we had preprocessed for another study using a similar pipeline that was used for the PREVENT-AD sample. Differences included a notch filter adjusted for the power line frequency in the UK (50 Hz vs. 60 Hz in Canada for PREVENT-AD), a distinct source mapping method (linearly constrained beamformer in Brainstorm with default parameters), and a 2-second time window with 50% overlap for power spectrum density estimation (compared to a 3-second time window with 50% overlap for the PREVENT-AD control sample).
We ensured that neuroanatomical features, including those altered by Parkinson’s disease, did not influence the differentiation of participants based on their spectral brain-fingerprints. To achieve this, we measured z-scored deviations in cortical thickness for each cortical parcel in PD patients. These deviations were calculated using FreeSurfer’s recon-all, based on the mean and standard deviation of cortical thickness observed in the age-matched healthy controls.
We employed linear regression models to investigate two i) whether patients who were most differentiable based on their brain-fingerprints also exhibited greater deviations in cortical thickness, and ii) the relationship between deviations in regional cortical thickness and regional ΔICC (as depicted in Fig. 3a, top panel).
We used individual Hoehn & Yahr scores as markers for disease staging in PD patients.^54^^,^^55^ We binarized these scores around a value of 2, creating two distinct one for patients with unilateral symptoms and one for those with bilateral symptoms.
We trained a linear support vector machine (SVM) classifier in R with default parameters to identify each patient’s symptom laterality category (i.e., unilateral vs. bilateral) based on their respective spectral brain-fingerprint features.
We conducted SVM classification for each cortical parcel independently. To train the SVM classifier, we used data from a random sample of 80% of the patients, while the remaining 20% served as the test set. We recorded the percentage of these held-out patients for whom the classifier accurately identified their Hoehn & Yahr category. We repeated this classification process 1000 times for each cortical parcel, generating an empirical distribution of symptom laterality classification accuracy across the cortex.
Next, we examined the spatial correlation between the cortical topographies of binarized Hoehn & Yahr category decoding accuracy and the regional ΔICC values from the brain-fingerprints (as discussed above in Saliency of Brain-fingerprint Features). Specifically, we correlated the differences in ICC values—calculated by subtracting those from the control-cohort experiment from those in the PD-cohort fingerprinting experiment—across cortical ROIs with the decoding accuracies obtained from Hoehn & Yahr score decoding.
In addition, we assessed whether a binarized form of Hoehn & Yahr scores (see Decoding Hoehn & Yahr category from brain-fingerprints for details) could be decoded from regional brain-fingerprint features using a linear support vector machine (SVM) classifier with leave-one-out cross-validation (LOOCV).
We randomly selected two subsets of participants, one with bilateral symptoms and the other with unilateral symptoms, to train an SVM classifier on the brain-fingerprint features identified at each cortical parcel. This classifier was then tested to predict whether a randomly selected held-out participant from the remainder of the cohort had bilateral or unilateral symptoms. This procedure was repeated for every participant in a leave-one-out fashion, where, for each iteration, two new random subsets of participants with either bilateral or unilateral symptoms were generated.
We examined whether the brain-fingerprints of individuals with Parkinson’s Disease (PD) align with the cortical topography of functional hierarchies in the cortex.^56^ To investigate this, we focused on the spatial relationship between the ΔICC brain-fingerprint topography and the first gradient of the cortical functional hierarchy.
We calculated Pearson’s spatial correlation between the ΔICC topography of brain-fingerprints and the atlas map of the first gradient of the cortical functional hierarchy. The gradient map used, available from neuromaps^45^, was parcellated into the 68 regions of the Desikan-Killiany atlas.
To statistically evaluate the significance of these correlations, we computed Bayes Factors using the correlationBF function in R. Additionally, we estimated p-values using permutation tests that accounted for the spatial autocorrelation inherent in the data (see Correlation with cortical neurotransmitter systems).^57^^,^^58^
Using a similar approach, we assessed the spatial correlation between the ΔICC values of brain-fingerprints and the normative atlas maps of various neurotransmitter systems. These systems were represented by maps for 19 receptors and transporters across 9 neurotransmitter systems, obtained from neuromaps. The neurotransmitter systems and their corresponding receptors and transporters dopamine (D1, D2, DAT); serotonin (5-HT1a, 5-HT1b, 5-HT2a, 5-HT4, 5-HT6, 5-HTT); acetylcholine (α4β2, M1, VAChT); GABA (GABAa); glutamate (NMDA, mGluR5); norepinephrine (NET); histamine (H3); cannabinoid (CB1) and opioid (MOR).
Each neurotransmitter system map was parcellated into the 68 regions of the Desikan-Killiany atlas. We then calculated Pearson’s spatial correlations between these neurochemical maps and the regional ΔICC values of brain-fingerprints.
To determine statistical significance, we corrected for multiple comparisons using the False Discovery Rate (FDR) method implemented in R’s p.adjust function.^59^ We also computed Bayes factors using the correlationBF function in R to quantify the evidence in favor of the alternative hypothesis that a spatial correlation exists.
For each significant spatial correspondence observed, we estimated p-values based on spatially constrained permutation tests.^57^^,^^58^ We conducted 1000 permutations of the neurochemical atlases using the Hungarian method.
To investigate the temporal variability of the PD brain-fingerprint, we utilized brain-fingerprints derived from 30-second recordings of data. This approach builds on our previous work, which demonstrated the robustness of spectral brain-fingerprints derived from brief recordings.^60^
To predict the Fisher z-transformed self-similarity (i.e., Iself) of successive brain-fingerprints, we employed second-order polynomial hierarchical regression models constructed using the lme4 package in R.
In our modeling, we nested the slope of gap duration within each subject, allowing for second-order polynomial fits for the gap duration between brain fingerprints.
artanh(self-similarity) ∼ poly(gap duration,2) ∗ group (PD vs. CTL) + head motion + random effect (1 + poly(gap duration, 2) |SubjId).
We collected at least two task-free MEG recordings, each lasting 5 minutes with participants’ eyes open, from 79 patients with PD and 54 age-matched healthy controls (PREVENT-AD sample; demographic details in Supplemental Table S1). We then applied source-imaging to the MEG sensor data, using individual cortical surfaces derived from T1-weighted structural MRI scans.^61^ For each participant, we estimated the power spectral density (PSD) of their cortical MEG time series in the 0–150 Hz frequency range, across the cortical regions defined by the Desikan-Killiany atlas.^53^ This process generated one spectral brain-fingerprint for each participant’s MEG recordings (see Methods).
Our goal with brain-fingerprinting was to quantify the distinctiveness of individual features in the brain-fingerprints of patients and healthy controls. We compared the cortical spectral features from each participant’s MEG recordings with those of all other participants in our sample. Specifically, brain-fingerprinting compares the similarity of a participant’s two consecutive brain-fingerprints (self-similarity; Iself^25^) to the similarity of a participant’s brain-fingerprint with those from the other participants in the cohort (other-similarity; Iother^25^). If a participant’s brain-fingerprint from the first data segment is more similar to their own brain-fingerprint from a second segment than to those of others, the participant is considered to be correctly differentiated from others. We calculated self- and other-similarity scores within each participant’s respective group (i.e., healthy controls or patients with PD), as detailed in the Methods section.
We conducted three differentiation experiments based on brain-fingerprinting: i) among healthy controls, ii) among patients with Parkinson’s disease and iii) between patients and healthy controls, where each patient with Parkinson’s disease was differentiated from all age-matched healhty controls iteratively (Fig. 1). Differentiating healthy controls aimed to replicate our earlier study with younger adults^24^ in an older participant group, providing a benchmark differentiation accuracy for the age group of patient participants’ age group.
We found that healthy controls could be differentiated from each other with an accuracy of 89.8% (CI [88.0, 94.0]; Fig. 2a), patients with PD from each other with 77.2% accuracy (CI [74.7, 81.7]), and patients from healthy controls with 81.1% accuracy (CI [81.0, 83.5]; Fig. 2a) using full spectral features.Fig. 2Differentiating patients with Parkinson’s disease from healthy controls using spectral brain-fingerprints. (a) Accuracy in distinguishing participants from their brain-fingerprints derived from full, arrhythmic, and rhythmic brain-fingerprints, estimated from 4-minute (bar plots) and 30-second (scatter plots) data segments. The scatter plots indicate the differentiation accuracy for all brain-fingerprint pairs derived from all possible contiguous 30-second segments from the original 4-minute data segments. Grey segments at the base of the bar plots represent control differentiation performances based on empty-room MEG recordings collected during each participant’s visit (see Supplemental Fig. S1). Error bars represent bootstrapped 95% confidence intervals. (b) Self-similarity statistics within participants for full spectral, rhythmic, and arrhythmic brain-fingerprints. The plots show the empirical density of self-similarity statistics between two consecutive brain-fingerprints in control and PD cohorts. Thehe PD group displays a wider distribution, suggesting more variability in patients for full spectral and arrhythmic features (1/f). (c) Self-similarity of brain-fingerprints from brief (30-second) data segments across full spectral, rhythmic, and arrhythmic features across all frequency bands. Patients with PD show lower self-similarity with increased gap durations between data segments used to derive brain-fingerprints (y-intercept shift downwards). The self-similarity of patient full spectrum brain-fingerprints decreases more rapidly with the gap duration between recordings. In contrast, the self-similarity of patient brain-fingerprints from rhythmic components was more self-similar than controls at short gap durations, and became comparable at longer durations. Shaded regions indicate the standard error on the mean. See Supplemental Fig. S5 for a narrow-band description of the observed rhythmic effects.Fig. 3Comparative analysis of brain-fingerprint differentiation in Parkinson’s disease and control groups. (a) Cortical maps comparing ICC scores for differentiating between patients and controls. Orange areas indicate regions where differentiation of individual patients is more effective than in controls. We replicated this finding in two independent samples of healthy the PREVENT-AD dataset (top panel) and the Cam-CAN dataset (bottom panel). (b) Differentiation accuracy from brain-fingerprints defined by top features for differentiating patients (left, cortical areas shown in orange) and top features for differentiating controls (right, cortical areas shown in purple).
We then sought to determine whether the present participants could be similarly differentiated based on brief, 30-second segments, replicating our previous observations in healthy younger participants with older healthy adults and patients.^24^ We observed a similar pattern of differentiation 84.9% for differentiation among healthy controls (computed 95% CI [83.1, 86.7]), 77.2% for differentiation among patients (95% CI [74.4, 79.9]), and 81.2% for differentiation between patients and healthy controls (95% CI [78.7, 83.7]) using full spectral features. These results demonstrate the robustness of the spectral brain-fingerprinting approach with respect to data length (scatter plots in Fig. 2a).
We aimed to understand why the accuracy of differentiating patients from healthy controls varied so significantly, both with full spectral brain-fingerprints (77.2% vs. 89.8% accuracy) and arrhythmic ones (66.5% vs. 74.1% accuracy). To address this, we compared the self-similarity of the brain-fingerprints of patients with PD and healthy controls (Iself^25^). Similarly, we compared the other-similarity (Iother^25^) of brain-fingerprints between patients and between healthy controls. Our analysis revealed no significant difference in other-similarity between healthy controls and patients with PD when comparing full spectral brain-fingerprints (see Supplemental Fig. S2). However, we noted a significant reduction in the self-similarity of the patients’ full spectral brain-fingerprints (t = 2.24, p = 0.02; permutation t-tests; Fig. 2b).
To better understand this effect, we analyzed the impact of arrhythmic vs. rhythmic neurophysiological spectral components on participant differentiation.
We assessed the respective contributions of both constituent components of brain activity to inter-individual differentiation by parameterizing the regional power spectra of the cortical time series as detailed in the Methods section. We defined arrhythmic brain-fingerprints as the modeled arrhythmic spectrum (i.e., 1/f) and rhythmic brain-fingerprints as the difference between the regional power spectra and the modeled arrhythmic component (i.e., 1/f).
The accuracy of inter-individual differentiation based on arrhythmic brain-fingerprints decreased to 74.1% among healthy controls (CI [72.0, 78.0]), 66.5% among patients (CI [62.9, 71.4]), and 71.5% accuracy individual patients and healthy controls (CI [69.6, 75.9]; Supplemental Information). In contrast, the accuracy of inter-individual differentiation based on rhythmic brain-fingerprints increased to 92.6% (CI [90.0, 96.0]) among healthy controls, 86.7% (CI [82.9, 91.4]) among patients, and 90.5% (CI [89.9, 92.4]) between individual patients and healthy controls (Fig. 2a bar plots). Both brain-fingerprints of the arrhythmic and rhythmic components derived from brief 30-second segments exhibited similar patterns, with arrhythmic brain-fingerprints differentiating between patients with lower accuracy than among healthy controls (Fig. 2a points).
Next, we aimed to understand whether the lower differentiation accuracy observed in the PD cohort for arrhythmic brain-fingerprints was due to reduced self-similarity. We found that arrhythmic brain-fingerprints in patients demonstrated reduced self-similarity (t = 4.86, p < 0.01; permutation t-tests), unlike for rhythmic brain-fingerprints (t = 1.77, p = 0.09; permutation t-tests; Fig. 2b).
Further, we investigated whether this discrepancy could be linked to the increased moment-to-moment variability in the brain activity of patients with PD within the recording session (as detailed in Methods under ‘Temporal variability of the PD brain-fingerprint’; see Supplemental Tables S2–S4 and Fig. 2c). Using the shorter 30-second data segments, we discovered that for full spectrum brain-fingerprints, the self-similarity (Iself) decreases more rapidly in patients than in healthy controls as the gap duration between data segments increases (β = −3.77, SE = 1.73, 95% CI [−7.16, −0.38], p = 0.029; detailed in Supplemental Table S2). This trend was not significant for arrhythmic brain-fingerprints (β = −3.67, SE = 2.33, 95% CI [−8.25, 0.92], p = 0.117; Supplemental Table S3). Conversely, rhythmic brain-fingerprints revealed a different for shorter time gaps between data segments, the neurophysiological activity in patients with PD was more self-similar (Iself) than that of controls, becoming comparable over longer durations (β = −3.33, SE = 1.69, 95% CI [−6.65, −0.02], p = 0.049; Supplemental Table S4). Breaking down these rhythmic self-similarity effects into the constituent narrow bands, we observed that, similar to the broadband effect (Fig. 2c right), the theta and alpha bands showed increased self-similarity (see Supplemental Fig. S5). Together, these results suggest that the decreased differentiation accuracy observed in PD is related to increased moment-to-moment variability of arrhythmic (1/f) brain-fingerprints in PD.
Given the noted temporal stability of rhythmic neurophysiological features in patients with Parkinson’s Disease (PD), we calculated the intraclass correlation (ICC) scores for each cortical region to identify the most consistent neurophysiological features in the rhythmic brain-fingerprints across individuals.^21^^,^^62^ We found distinctive patterns of rhythmic neurophysiology in varying brain regions between healthy controls and patients with PD. The highest ICC values were in frontal and medial cortical regions for healthy controls (Supplemental Fig. S3a), and in the right pre- and post-central regions for patients (Fig. 3a and Supplemental Fig. S3b).
We replicated these findings with an external sample of age-matched healthy controls from the Cambridge Center for Aging Neuroscience (Cam-CAN) dataset^63^ (Fig. 3a, see Methods). We computed a cortical map of ICC values for the independent sample of older healthy controls and contrasted this map with the topography of patients with PD (i.e., ΔICC map). The cortical maps of the distinctive patterns of the PD brain-fingerprint obtained using the two separate control samples were strongly correlated across samples (r = 0.75, p < 0.001, pspin < 0.001).
To gauge how the spatial divergence between the rhythmic brain-fingerprints relates to individual differentiation, we created brain-fingerprints for both patients with PD and healthy controls using the top 10% of ICC features specific to each group (Fig. 3a). Utilizing the most distinctive features of the brain-fingerprints from patients, we achieved a differentiation accuracy of 78.7% (CI [76.6, 80.8]; Fig. 2a) among healthy controls, and 88.7% (CI [85.5, 91.8]; Fig. 3b) among patients. In contrast, using the features most salient in healthy control brain-fingerprints, we differentiated healthy controls with 92.7% accuracy (CI [90.6, 94.9]; Fig. 3b), and patients with only 66.9% accuracy (CI [62.4, 71.5]; Fig. 3b).
We then explored whether the rhythmic brain-fingerprint of patients indicates their respective clinical stage of the disease. We implemented classifiers to decode whether a patient experiences unilateral or bilateral symptoms from the regional features of their rhythmic brain-fingerprint (see Methods, Decoding disease staging from brain-fingerprints). The laterality of symptoms was derived from each patient’s binarized score on the Hoehn & Yahr clinical scale (HY < 2 for unilateral, HY ≥ 2 for bilateral).^54^^,^^55^
The cortical map of regional decoding accuracies revealed that it is possible to distinguish symptom laterality through electrophysiological brain activity, with accuracies exceeding chance levels. The most notable brain regions enabling this decoding were the right post–central gyrus and the left caudal middle frontal gyrus, showing decoding accuracies of 69.6% and 68.8%, respectively (Fig. 4a). This data-driven approach uncovered that, in these specific regions, individuals with bilateral symptoms exhibit a suppression of faster brain activity above 15 Hz and an increase in slower activity (6–9 Hz) (Fig. 4a, right panel).Fig. 4Decoding Parkinson’s disease symptom laterality from brain-fingerprints. (a) Cortical topography of decoding accuracies for Parkinson’s symptom laterality (based on binarized Hoehn & Yahr scores). On the right, power spectra of resting-state neurophysiological activity in the right postcentral gyrus, the cortical region with the highest accuracy for decoding symptom laterality. The plots represent the average power spectrum for each healthy controls, individuals with unilateral symptoms, and individuals with bilateral symptoms, with shaded areas indicating standard errors across groups. (b) Scatter plot showing how the decoding accuracy of binarized Hoehn & Yahr scores from brain-fingerprint features of each cortical parcel correlates with the saliency of each parcel, as determined by its ΔICC score.
Moreover, we found that the cortical map for binarized Hoehn & Yahr category decoding aligns with the map of ICC difference scores (Fig. 4b). We replicated this alignment using the Cam-CAN sample of healthy controls, with correlations of r = 0.36 (p < 0.01, pspin < 0.001) and r = 0.51 (p > 0.001, pspin < 0.001), respectively. This consistency in findings was robust regardless of the cross-validation method employed for training the binarized Hoehn & Yahr classifiers (Supplemental Figs. S7 and S8).
We found that the regional disparities in prominent features of the rhythmic brain-fingerprint between patients with PD and controls (indicated by ΔICC; see Fig. 3a) aligned with the unimodal-to-transmodal functional gradient of the cortical hierarchy^56^ (r = −0.49, p < 0.001, pspin < 0.001; Fig. 5a, with details in Methods). The most notable rhythmic brain-fingerprint features in healthy adults were associated with transmodal cortical regions. Conversely, the distinct features of the Parkinson's rhythmic brain-fingerprint were more closely related to unimodal (i.e., primary sensorimotor) areas within the functional hierarchy of the cortex. We replicated this effect using the Cam-CAN sample of healthy controls (r = −0.53, p < 0.001, pspin < 0.00**1; Supplemental Fig. S9).Fig. 5Correlation of spectral brain-fingerprints with cortical functional hierarchy and neurotransmitter systems. (a) Top: Cortical map illustrating the first unimodal-to-transmodal functional gradient, sourced from neuromaps.^45^ Bottom: Linear association between the weights of cortical regions in this functional gradient (sourced from neuromaps) and their prominence in the PD brain-fingerprint (Fig. 3a, top). (b) Top: Bayes factor analysis of the topographical alignment between PD brain-fingerprint features (from Fig. 3a) and atlases of various cortical neurochemical systems, highlighting strong correlations particularly with serotonin, cannabinoid, mu-opioid, and norepinephrine systems. Each row represents data from different control samples. Bottom: Selected neurochemical cortical atlases, as obtained from neuromaps.
We further investigated whether the most prominent features of the Parkinson’s brain-fingerprint were related to the cortical distribution of major neurotransmitter systems. Using neuromaps,^45^ we obtained 19 normative cortical maps representing 9 neurotransmitter systems (Fig. 5b, bottom) and assessed their spatial correlation with the cortical map of ICC difference scores (Fig. 3a, PREVENT-AD sample; see Methods). Our analysis revealed significant correlations with several neurotransmitter systems, including serotonin-2a (r = −0.39, pFDR = 0.006, pspin < 0.001), serotonin-4 (r = −0.37, pFDR = 0.008, pspin = 0.008), cannabinoid-1 (r = −0.41, pFDR = 0.00045, pspin < 0.001), mu-opioid (r = −0.34, pFDR = 0.018, pspin = 0.007) receptors, and the norepinephrine transporter (r = 0.43, pFDR = 0.0040, pspin < 0.001). Notably, the cannabinoid, opioid, and serotonin systems, concentrated in temporal and frontal cortical regions, corresponded with the most salient rhythmic brain-fingerprint features in healthy controls (Fig. 5b & Supplemental Fig. S3a). Conversely, the pronounced presence of norepinephrine transporters in the somatomotor cortices mirrored the significance of rhythmic neurophysiology in these areas in patients with PD (Figs. 5b and 3a).
This effect was replicated using the CamCAN sample of healthy controls (Fig. 3a). We found alignments with the cortical distributions of serotonin-2a (r = −0.31, pFDR = 0.02, pspin = 0.005), serotonin-4 (r = −0.43, pFDR = 0.002, pspin = 0.002), cannabinoid-1 (r = −0.35, pFDR = 0.01, pspin = 0.003), mu-opioid (r = −0.37, pFDR = 0.007, pspin = 0.002) receptors, and the norepinephrine transporter (r = 0.52, pFDR < 0.001, pspin < 0.001). Additionally, we observed correspondence with the dopamine-1 (r = −0.31, pFDR = 0.02, pspin = 0.005), dopamine-2 (r = −0.28, pFDR = 0.04, pspin = 0.035), and serotonin transporter maps (r = −0.43, pFDR = 0.001, pspin = 0.003; Fig. 5b).
To ensure the reliability of spectral brain-fingerprints, we tested their robustness against environmental and physiological artifacts. We first evaluated environmental factors, particularly those related to recording conditions on different days. To this end, we used empty-room MEG recordings conducted around each participant’s visit. By processing these recordings identically to the participant data and mapping them onto the participant’s cortex using the same imaging procedure, we computed an empty-room “mock brain-fingerprint” for each participant. We then used these mock brain-fingerprints to test that environmental factors influenced inter-individual differentiation. As expected, we found that inter-individual differentiation accuracies from these mock brain-fingerprints were substantially lower than those using the actual brain-fingerprints (<5%; see Fig. 2a & Supplemental Fig. S1). We conclude that environmental factors did not significantly contribute to inter-individual differentiation based on actual brain-fingerprints.
Further, we evaluated the influence of common physiological artifacts in MEG recordings—such as head motion, heart-rate variability, and eye blinks—on brain-fingerprinting. Our findings indicated that inter-individual differentiability was not significantly affected by cardiac or ocular artifacts (r = −0.04, p = 0.71 and r = −0.08, p = 0.46, respectively). However, there was a modest association with head movements in the PD cohort (r = 0.24, p = 0.04; Bayesian post-hoc analysis BF = 2.04; see Supplemental Fig. S6). Consequently, we included head motion as a nuisance covariate in all subsequent regression analyses (detailed in Methods). Notably, there were no significant differences in physiological artifact profiles between healthy controls and patients with PD (head t(64.34) = 0.41, p = 0.68; EOG: t(123.88) = −0.91, p = 0.36; ECG: t(64.41) = −1.24, p = 0.22).
Lastly, considering previous reports of cortical thickness abnormalities in PD,4, 5, 6, 7 we investigated whether these structural changes could partly explain the differentiability of PD patients from healthy controls. We derived cortical thickness measures from the structural MRI data of both groups, when available (n = 134; Supplemental Fig. S4a). We standardized the patients’ cortical thickness maps using z-score transforms based on healthy controls. Our analysis revealed no significant linear relationship between individual differentiability and the average standardized cortical thickness in PD patients (b = −0.03, SE = 0.07, 95% CI [−0.16, 0.11], p = 0.69; Supplemental Fig. S4b). Additionally, the cortical topography of the most salient Parkinson’s brain-fingerprint features did not align with the cortical thickness changes observed in patients (Pearson’s r = 0.04, t(66) = 0.34, p = 0.73, pspin = 0.36). Thus, we conclude that the individual differentiability observed in PD patients based on their brain-fingerprints is not significantly influenced by cortical thickness alterations associated with the disease.
Our study demonstrates the application and relevance of brain-fingerprinting to Parkinson’s disease (PD) research. We derived brain-fingerprints from task-free MEG recordings and first replicated the prior observation that the brain-fingerprints of patients with PD exhibit increased variability over short periods of time compared to healthy controls.^29^^,^^43^ We identified that this effect is largely due to the enhanced temporal variability of the arrhythmic component of the neurophysiological brain activity of patients with PD, making them less distinguishable from one another. However, we observed that patients can be accurately differentiated from each other and from healthy controls based on brain-fingerprints derived from the rhythmic components of their ongoing electrophysiological brain activity. Futhermore, we show that the distinct features of these rhythmic fingerprints correlate with symptom laterality and align with neurochemical systems impacted in PD, highlighting the potential for targeting neuromodulation therapies based on rhythmic cortical neurophysiology in PD.
Previous studies have highlighted frequency-specific signaling abnormalities in PD, particularly in motor and subcortical structures.^64^^,^^65^ Our findings align with this literature,^10^^,^^66^^,^^67^ showing that the most distinctive brain-fingerprint features in PD patients localize to the primary sensorimotor cortex (Fig. 3b, left panel & Supplemental Fig. S3c). This localization correlates with our ability to decode symptom laterality from rhythmic brain-fingerprints (Fig. 3b). Specifically, we found evidence linking between atypical beta and theta band activities in the postcentral gyrus and symptom laterality. This observation supports previous findings that associate beta-bursting in the motor network and sensorimotor cortex with symptom severity and treatment response to medication^8^^,^^10^ and deep brain stimulation of the subthalamic nucleus.^68^
In the present study, we restricted disease staging to whether patients experienced unilateral or. bilateral symptoms. Our data show that this disease stage is best decoded from the regional rhythmic features of their brain-fingerprint, particularly in the pre- and post–central cortices bilaterally. This finding highlights the clinical relevance of the neurophysiological brain-fingerprint and may encourage further research to associate more nuanced individual symptom presentations with specific brain-fingerprint features.
We observed that the most salient brain-fingerprint features of healthy controls align with regions of the default-mode network (DMN; Fig. 5a & Supplemental Fig. S3a). Prior studies have noted functional decoupling of the DMN in PD during rest and task-based activities,69, 70, 71, 72 often linked to the dopaminergic system.69, 70, 71, 72 However, our data from patients on stable antiparkinsonian medication regimens may have moderated the saliency of DMN regions in the patients’ brain-fingerprints (Fig. 3). Thus, our observation that the DMN and other transmodal brain regions of the functional hierarchy do not contribute substantially to the Parkinson brain-fingerprint (Fig. 5a) may reflect a normalization effect of medications.^70^^,^73, 74, 75 These findings prompt further investigation into how responsiveness to medications relates to brain-fingerprints in transmodal brain regions.
Our data suggest that monoamine neurotransmitters are closely associated with the brain-fingerprint of PD (Fig. 5b). Specifically, we found that the cortical topography of serotonin 2a and 4 receptor densities is inversely related to the PD brain-fingerprint, whereas there is a direct association with the norepinephrine transporter.
This finding aligns with prior reports of monoamine degradation in PD.^76^ We observed a negative relationship between the PD brain-fingerprint and dopamine systems (Fig. 5b), and interpret the weak and inconsistent effect as possibly stemming from limitations in the present data. One consideration is that PD-induced alterations in dopaminergic signaling might predominantly impact subcortical structures,^77^ which were not the focus of our analyses. Future research is warranted to explore whether a similar spatial correlation exists between dopaminergic receptor concentrations and the distinct neurophysiology of PD in deep brain nuclei.
We also observed a negative alignment of the Parkinson's brain-fingerprint with the cannabinoid receptor-1 (CB1) system (Fig. 5b), supporting prior research highlighting CB1 as a potential therapeutic target in PD.^78^ Our present results also highlight the potential role of the cannabinoid system in the neuropathophysiology of PD and encourage further research in this area.
While the present findings do not specifically address whether neurochemical effects underpin the distinction between motor and non-motor symptoms in Parkinson’s Disease, emerging research has begun to explore the relationship between neurochemical systems and changes in neurophysiology in PD.^15^^,^^79^ Notably, increased neurophysiological slowing in somatomotor regions correlates with deteriorated attention performance in PD.^15^ Furthermore, the association between neurophysiological slowing and clinical impairment, encompassing both motor and non-motor symptoms, demonstrates a ventromedial-dorsolateral spatial gradient. This gradient corresponds to the cortical distribution of the monoamine system, including receptors and transporters such as D1, D2, DAT, 5HT1-a, 5HT2-a, 5HT4, and 5HT-T. Thus, the neurophysiological changes observed in PD, which coincide with the monoamine system’s spatial distribution, may reflect the spectrum of motor and non-motor symptoms.
We note that the processing of the two control samples differed, as detailed in Cam**-**CAN sample of healthy controls. Despite differences in the respective pre-processing of the PREVENT-AD and Cam-CAN MEG healthy controls’ data, we observed qualitatively similar cortical topographies for the PD brain-fingerprint (i.e., ΔICC maps; Fig. 3).
Our study revealed that the brain-fingerprints of patients with Parkinson’s disease fluctuate more over short time spans compared to age-matched healthy individuals. This finding aligns with the decreased accuracy initially observed in differentiating individuals within the patient group (Fig. 2b &c).
We anticipated greater variability in PD brain activity based on previous fNIRS research, which suggested a correlation between symptom severity and hemodynamic signal variability.^43^ Additionally, studies using fMRI connectome brain-fingerprinting indicated reduced self-similarity (Iself) in individuals at risk of or with mental health disorders^27^^,^^28^ and in PD patients.^29^ Our data extend these findings to electrophysiology, highlighting increased within-participant variability of arrhythmic brain activity in PD as a possible source of such variability. We noted that differentiation accuracy using full spectral and arrhythmic brain-fingerprints in PD patients was lower compared to rhythmic brain-fingerprints, which achieved similar differentiation to that seen in healthy controls (Fig. 2a).
Recent research has linked alterations in arrhythmic brain activity in patients with PD to symptom severity.^15^^,^^39^^,^^80^ Preliminary studies further suggest that baseline arrhythmic activity in the subthalamic nucleus may predict responses to neuro-stimulation protocols.^13^^,^^14^ While these studies focused on group-level mean differences, our findings emphasize the significance of within-patient variability in arrhythmic brain activity for understanding individual disease manifestations.
Previous studies have also documented increased intra-individual variability in cognitive task performance in PD,81, 82, 83 correlating with cognitive symptom severity.^81^^,^^82^^,^^84^ The biological basis of this increased behavioural variability remains poorly understood.^85^ fMRI research has linked moment-to-moment brain activity variability with cognitive performance,86, 87, 88 and recent studies have related BOLD signal variability to the arrhythmic components of electrophysiology.^89^ Consequently, we hypothesize that the heightened variability in PD behavioural markers may be associated with the observed increased temporal variability in arrhythmic brain activity.
The arrhythmic and rhythmic components of the neurophysiological spectrum are related to distinct underlying mechanisms.^32^^,^^40^^,^^90^ The slope of the arrhythmic spectrum is believed to mirror the balance between excitation and inhibition within neuronal circuits.^32^^,^^40^ Thus, our findings imply that PD pathophysiology may alter cortical excitability.
Recent investigations have reported that patients with PD exhibit more stereotyped network dynamics when off their medications.^91^^,^^92^ Greater stereotyped network dynamics could challenge inter-individual differentiation between patients due to increased other-similarity. However, Our findings do not support this interpretation. Future research should reconcile these observations by examining arrhythmic and rhythmic network dynamics in Parkinson’s disease.
We note that the windowing approach used to derive brain-fingerprints from brief recordings imposes limits on the temporal resolution of our analysis and potentially increases the variability of the arrhythmic component. It is important to emphasize, however, that any such bias in increased variability should equally affect both patient and healthy participants, and therefore not drive the observed group effects.
We also note that, although we took every precaution to model the neurophysiological power spectra using specparam, this approach has limitations such as the need for manual tuning of hyperparameters. Future efforts should focus on developing automatic model selection approaches should to address these issues and improve the robustness of the analysis.
Our findings demonstrate that brief brain recordings can distinguish individuals,^24^ including those with Parkinson’s disease. We highlight the consistent within-participant stability of rhythmic brain-fingerprints in both patients and healthy controls (see Fig. 2c), offering unique insight into individual-specific brain activity. This consistency aligns with prior research showing the stabilization of spectral content in resting-state brain activity within 30–120 seconds of MEG recording.^60^ This rapid stabilization is especially beneficial for clinical applications, particularly for patients with cognitive or motor impairments who may find longer recording sessions challenging.
We observed that the rhythmic brain-fingerprints of patients with Parkinson’s disease are more self-similar at short gap durations (Fig. 2c). This finding aligns with the slowing hypothesis in PD, which posits that patients with PD show increases in low-frequency brain activity (i.e., delta [2–4 Hz] and theta [4–8 Hz]) and decreases in high–frequency activity (e.g., alpha and beta [above 8 Hz]).^1^^,^^12^^,^^15^^,^^93^^,^^94^ Recent research suggests that the neurophysioloigical slowing observed in PD is not related to changes in arrhythmic brain activity and reflects both adverse and compensatory effects depending on their cortical topography.^15^ At this stage, we can only speculate that greater low-frequency fluctuations in brain activity observed in PD may induce greater self-similarity of brain activity over short periods. We noted that the increased self-similarity of rhythmic brain-fingerprints at short gap durations in patients was particularly expressed in the slower frequency range (theta 4–8 Hz; Supplemental Fig. S5). Future work is needed to establish whether increases in low–frequency activity drive the observed increases in rhythmic brain-fingerprint self-similarity.
Our results dovetail with previous work suggesting that brain networks may be differentiable at short time scales, particularly from visual-somatomotor regions in fMRI.^95^ The authors of that study conjectured that this ability to differentiate individuals at short time scales might be driven by oscillatory neuronal activity.^95^ Our findings expand upon these results by providing further evidence that rhythmic brain activity measured over brief segments can differentiate between individuals (Fig. 2).
Although our observations provide insights into the pathophysiology of Parkinson’s disease, brain-fingerprinting is not intended to replace standardized clinical assessments. The data and the approach presented here enhance the fundamental understanding of the neurophysiological pathways of the disease. We anticipate that future developments in neurophysiological brain-fingerprinting may also refine patient stratification by highlighting, in a data-driven manner, personalized expressions of PD pathophysiology.
Recent studies have shown that deep brain stimulation (DBS) protocols at high frequencies (130 Hz) lead to an increase in the slope of the arrhythmic spectrum within the subthalamic nucleus target.^13^^,^^42^ Similarly, baseline arrhythmic activity within the subthalamic nucleus has been found to predict the clinical effectiveness of these neurostimulation protocols.^14^ Our data suggest that systematically quantifying moment-to-moment fluctuations in arrhythmic brain activity could enhance the accuracy of predicting the effectiveness of neurostimulation protocols in individual patients. Additionally, this approach could facilitate a more structured monitoring of their clinical trajectory.
We also observed that, over short periods, rhythmic theta and alpha activities distinguish between patients more effectively than between controls, as shown in Supplemental Fig. S5. This observation aligns with previous research indicating that rhythmic brain activities are associated with cognitive and motor symptoms in PD.^12^^,^^15^^,^^65^^,^^96^^,^^97^ Specifically, theta-, alpha-, and beta-frequency rhythms originating in the fronto-motor cortices have been highlighted as potential targets for non-invasive neurostimulation protocols aimed at normalizing disease-related neurophysiology in PD.^68^^,^98, 99, 100, 101 Although the target frequencies in most stimulation protocols are standardized across patients, our findings suggest that the features of individual rhythmic brain-fingerprints could provide a basis for determining personalized frequency targets for neurostimulation.
Our present findings rely on non-invasive cortical neurophysiology techniques, making them more readily translatable to tracking the effectiveness of non-invasive neurostimulation protocols, such as transcranial magnetic stimulation (rTMS).^99^ We believe, however, that brain-fingerprinting can also be applied to monitor the cortical effects of deep brain stimulation interventions in PD and the resulting longitudinal trajectories of patients.
We also highlight the need to account for increased intra-individual variability of brain activity in disease states when developing statistical and machine learning models for disease classification. This consideration is crucial for ensuring the scalability and generalizability of patient stratification methods.
Our approach to brain-fingerprinting can, in principle, be applied to both electroencephalography (EEG) and the evolving MEG technology based on optically pumped magnetometers (OPMs), which are poised to become more affordable and accessible in clinical settings.
Our results should be interpreted with some important caveats. First, studying inter-individual differences in a disease as heterogeneous as PD requires large sample sizes. The results presented here should be replicated in a larger cohort with greater symptom diversity, allowing for more individualized mapping between neurophysiological fingerprints and symptom profiles. Additionally, demographic variability is a crucial factor when studying inter-individual differences. Detailed demographic information (e.g., ethnicity, gender expression, socioeconomic status, sexual orientation) was not available for most of the open datasets we used in our analyses. Consequently, our samples may not be entirely representative in this regard. Future research should assess the generalizability of our findings in a more demographically diverse sample.
Second, the patients included in the sample were all on a stable dose of anti-parkinsonian medications, which are known to normalize aberrant neurophysiological activity.^97^ Future research should explore the effects of dopaminergic medications on neurophysiological brain-fingerprints.
In conclusion, our study underscores the clinical significance of brain-fingerprinting based on rapid neurophysiological activity dynamics. It illuminates the clinical aspects of Parkinson’s disease, identifying specific brain regions and rhythms where the disease impacts neurophysiological stability. We anticipate these insights will catalyze further research in population neuroscience and the development of personalized neuromodulation therapies for Parkinson’s disease and other neurodegenerative conditions.
JDSC contributed to the conceptualization of the project, curated data, formal analysis, methodology, software, visualization, and writing of the original draft. AW contributed to the conceptualization, curation of data, methodology, and reviewing & editing of the manuscript. JYH contributed to the software, methodology, and reviewing & editing of the manuscript. BM contributed to the methodology, supervision, and reviewing & editing of the manuscript. SB contributed to the conceptualization, methodology, supervision, funding acquisition, and reviewing & editing of the manuscript. JDSC and AW accessed and verified the underlying data. The PREVENT-AD Research Group and the Quebec Parkinson Network collected and curated the data. All authors read and approved the final version of the manuscript.
The data are available through the Clinical Biospecimen Imaging and Genetic (C-BIG) repository (https://www.mcgill.ca/neuro/open-science/c-big-repository), the PREVENT-AD repository (https://openpreventad.loris.ca/ and https://registeredpreventad.loris.ca),^47^ and the OMEGA repository (https://www.mcgill.ca/bic/resources/omega).^48^ Normative neurotransmitter density data are available from neuromaps (https://github.com/netneurolab/neuromaps).^45^
All authors declare no competing conflicts of interest. The listed funding sources in the Acknowledgements did not play any role in the writing of the manuscript or the decision to submit this manuscript for publication.