Brain-fingerprinting

Author: Jason da Silva Castanheira

This tutorial introduces the concept and method of neurophysiological brain-fingerprinting. Using the outputs of the OMEGA tutorial, we will derive spectral brain-fingerprints, as detailed below, for the five example participants.

Introduction

Brain-fingerprinting is a method to assess inter-individual differences in brain activity. It was first proposed by Finn and colleagues in 2015. The present tutorial will demonstrate the method's functionality using Brainstorm outputs.

This demonstration uses sample data from the OMEGA dataset, and can be followed by completing all steps of the OMEGA tutorial.

Preparing the data for brain-fingerprinting

The brain-fingerprinting method requires features of interest to define the so-called brain-fingerprint. These features can be derived from functional connectomes or neural power spectrum, as detailed in da Silva Castanheira et al., 2021. Note that this is not an exhaustive list of all possible neurophysiological features that can be used to define brain-fingerprints.

This tutorial will focus on defining brain-fingerprints using neural power spectra at every parcel of the Destrieux atlas.

To do so, we compute the spectrum (i.e., PSD) using the Welch method at every vertex of the cortical surface, and average the resulting values within each region defined by the Destrieux atlas (see Scout tutorial page). Note, this is done to downsample the spatial features of the brain-fingerprint. If we did not downsample the PSD to a cortical atlas, the resulting feature space would be too large to work with (e.g., 15,000 vertices by 300 frequencies). The choice of atlas and neurophysiological features (e.g., functional connectomes vs power spectra) to define the brain-fingerprint is dependent on the user's specific hypotheses.

See code snippet below:

Brain-fingerprinting requires at least two data segments for every individual. This can be separate recordings used to define the two brain-fingerprints (i.e., between-session fingerprinting), or two segments within the same recording session (i.e., within-session fingerprinting). For this demonstration, we will split the OMEGA recordings in half (i.e., within-session fingerprinting).

See the image below for a summary of the pipeline:

The brain-fingerprinting method

The brain-fingerprinting method is based on the correlational similarity of participants across data segments (see Image above). For each participant in a given cohort, we compute the Pearson correlation coefficient between the brain-fingerprint of the first segment and second segment of all participants in the cohort, including the probe participant. In the case of this demonstration, we are computing the correlation between the topographic distribution of spectral power across data segments. This yields a participant similarity matrix, where off-diagonal elements represent the similarity of a participant i to all other participants and diagonal elements represent the similarity of a participant's brain-fingerprint across data segments.

See the image below for an example participant similarity matrix.

Participant differentiation consists of a lookup procedure along the rows or columns of the participant similarity (correlation) matrix. A participant is said to be correctly differentiated if a participant's brain-fingerprint is more similar to themselves (self-similarity, also refered to as Iself) than all other participants in the cohort (other-similarity, also refered to as Iothers). In other words, if the largest correlation coefficient between the first brain-fingerprint and the second matches the probe participant.

Differentiation accuracy

Differentiation accuracy represents the percent ratio of the number of participants correctly differentiated across the cohort (i.e., differentiation accuracy). Brain-fingerprinting is typically repeated for all possible pairs of data segments. In the case of this tutorial, we only have two data segments—the first and second half of the recording—and therefore only have two differentiation accuracies. These two accuracies are obtained by the lookup procedure along the rows and columns of the participant similarity matrix respectively.

In the case where three data segments are available, we can compute six differentiation accuracies: the accuracy for the columns and rows of data segment 1 vs segment 2, and for each possible pair of recordings (data segment 2 vs 3 and data segment 1 vs 3).

Differentiability

Differentiability is a measure which describes how easyily a given participant is to differentiate from a cohort of individuals. This measure consists of scaling (z-scoring) the self-similarity of brain-fingerprints against the other-similarity (off-diagonal elements). A participant with a high differentiability score will have a higher self-similarity relative to their similarity to others in the cohort (other-similarity) and therefore are easily differentiated (see the last row in the example participant similarity matrix below). In contrast, a participant whose brain-fingerprint is most similar to others in the cohort will be more challenging to differentiate (see the second row of the participant similarity matrix below).

Note other metrics to measure the relative easy of participant differentiation exist: Idiff which is computed as the difference between Iself and Iothers. See Amico & Goñi 2018 and Sareen at el., 2021.

Quantifying the most salient features of the brain-fingerprint

Beyond differentiation accuracy and differentiability, we can quantify the relative contribution of regions and frequencies to participant differentiation.

To do so, we calculated intraclass correlations (ICC). ICC quantifies the ratio of within-participant variance and between-participant variance, with participants as their own raters across data segments (see Amico et al., 2018 & da Silva Castanheira et al., 2021 for details). Features that contribute the most to participant differentiation should be highly consistent within individuals, but show great inter-individual variance. Larger values of ICC indicate that a neurophysiological feature (e.g., region of interest or frequency band) contributes more to participant differentiation than smaller values.

The following code snippet computes ICC for every frequency band and region of the brain-fingerprint:

Applications of the brain-fingerprint

Beyond biometric differentiation, brain-fingerprints advance the neurobiological origins of individual traits and inform clinical. Below we summarize some of the most recent applications of brain-fingerprints.

• Decoding individual traits: Previous work demonstrates that beyond biometric identification, brain-fingerprints can predict individual differences in intelligence test performance (e.g., Finn et al., 2015; Amico et al., 2018) and demographics (da Silva Castanheira et al., 2021).

• Clinical application: Recent work has explored the clinical utility of bran-fingerprints. This work has shown across multiple neurological diseases that the brain-fingerprints of patients show increased variability (lower self-similairty/ Iself ) in brain-fingerprints of clinical populations (e.g., Kaufmann et al., 2017, 2018; Sorrentino et al., 2021; Troisi Lopez et al., 2023; da Silva Castanheira et al., 2024). A potential explanation for the de-individualization of brain activity in observed in Parkinson’s disease may be an inherent instability in the arrhythmic brain activity of patients over short periods. In contrast, rhythmic brain activity appears particularly characteristic of patients with PD (da Silva Castanheira et al., 2024). Future work is exploring the application of brain-fingerprints to the design of personalized medicine.

Additional documentation

Please cite previous work on brain-fingerprinting:

• da Silva Castanheira, J., Orozco Perez, H.D., Misic, B. et al. Brief segments of neurophysiological activity enable individual differentiation. Nat Commun 12, 5713 (2021). https://doi.org/10.1038/s41467-021-25895-8

• Finn, E., Shen, X., Scheinost, D. et al. Functional connectome fingerprinting: identifying individuals using patterns of brain connectivity. Nat Neurosci 18, 1664–1671 (2015). https://doi.org/10.1038/nn.4135

Further reading:

• Amico, E., Goñi, J. The quest for identifiability in human functional connectomes. Sci Rep (2018). https://doi.org/10.1038/s41598-018-25089-1

• Sareen, E., Zahar, S., Ville, D. V. D., Gupta, A., Griffa, A., & Amico, E. Exploring MEG brain fingerprints: Evaluation, pitfalls, and interpretations. NeuroImage (2021). https://doi.org/10.1016/j.neuroimage.2021.118331

• Sorrentino, P., Rucco, R., Lardone, A., Liparoti, M., Troisi Lopez, E., Cavaliere, C., Soricelli, A., Jirsa, V., Sorrentino, G., & Amico, E. Clinical connectome fingerprints of cognitive decline. NeuroImage, (2021). https://doi.org/10.1016/j.neuroimage.2021.118253

• Troisi Lopez, E., Minino, R., Liparoti, M., Polverino, A., Romano, A., De Micco, R., Lucidi, F., Tessitore, A., Amico, E., Sorrentino, G., Jirsa, V., & Sorrentino, P. Fading of brain network fingerprint in Parkinson’s disease predicts motor clinical impairment. Human Brain Mapping, (2023). https://doi.org/10.1002/hbm.26156

• Kaufmann, T., Alnæs, D., Brandt, C. L., Bettella, F., Djurovic, S., Andreassen, O. A., & Westlye, L. T. Stability of the Brain Functional Connectome Fingerprint in Individuals With Schizophrenia. JAMA Psychiatry, (2018). https://doi.org/10.1001/jamapsychiatry.2018.0844

• Kaufmann, T., Alnæs, D., Doan, N. T., Brandt, C. L., Andreassen, O. A., & Westlye, L. T. Delayed stabilization and individualization in connectome development are related to psychiatric disorders. Nat Neurosci, (2017). https://doi.org/10.1038/nn.4511

• da Silva Castanheira, J., Wiesman, A. I., Hansen, J. Y., Misic, B., Baillet, S., Network Quebec Parkinson, & PREVENT-AD Research Group. (2023). Neurophysiological brain-fingerprints of motor and cognitive decline in Parkinson’s disease. medRxiv.