Statistical decision theory and multiscale analyses of human brain data
Introduction
Recent developments in brain recording techniques allow one to record animal brain data with high spatiotemporal resolution (Jun et al., 2017). At the same time, the ability to collect human brain data from large numbers of subjects have revolutionised the study of neurological diseases and disorders (Braund et al., 2018; Williams et al., 2011). Thus, we can now study details at the scale of a local cortical circuit using animal models and describe differences between very large numbers of individuals at the macroscopic scale with non-invasive human electrophysiology. These developments suggest the need to develop multiscale approaches. These will allow us to connect animal and human models. So far, approaches for the analysis of brain data contain information at a single spatiotemporal scale only. We here discuss a multiscale approach for brain imaging data analysis. This is based on combining neural models that describe brain data at different spatial scales.
We focus on a neural mass model that can explain both animal data obtained with thin laminar probes and human MEG data. Neural masses are biophysical models describing neural population responses where ensemble activity is considered as a point process. For a general introduction to these and similar models, see (Deco et al., 2008; Moran et al., 2013). This and other nomenclature used below are defined in Table 1. Following our earlier work (Pinotsis et al., 2017), we provide a mathematical proof that the neural mass model makes similar predictions to a microscopic, compartmental model. This is based on statistical decision theory (Berger, 2013) and shows that both models can be thought of as rules belonging to the same equivalence class. We show that a Bayesian observer could not distinguish between the data predicted separately by each model. Alternatively, if both models are fitted to the same data using Bayesian inference then these fits will have the same error. This suggests the similarity of their predictions.
As an illustration, we considered laminar differences in the excitation to inhibition balance (E–I) in human MEG data reported in (Schwarzkopf et al., 2012). We asked whether cortical function changes at various depths, and focus on differences in the (E–I) balance relevant to both pathophysiology (Chen et al., 2003) and information processing in the brain (Auksztulewicz and Friston, 2015; Friston et al., 2015a; Pinotsis et al., 2014). We find that differences in the E–I balance are expressed in the recurrent connection strengths of inhibitory interneurons in layers 5/6. Although MEG does not provide direct access to laminar data, the use of a neural mass model that makes laminar predictions allowed us to disclose details about cortical function that would otherwise be accessible only by using invasive recordings. In (Pinotsis et al., 2013a,b), we analysed the same dataset using a neural field model. Here, we used a neural mass model instead. Neural masses are a limiting case of neural fields when intrinsic delays on the cortical manifold are neglected, see also Table 1. We did not use a field model because it does not have the same parameters as the compartmental model considered below (see Lemma 1 in the Theory and Calculations section).
In (Pinotsis et al., 2017), we analysed data from laminar electrodes recorded from a single subject. Compared to that earlier work, our current paper includes two new contributions: the analysis of non invasive data and also of data from multiple subjects. Here we analysed MEG data using a hierarchical Bayesian approach (Parametric Empirical Bayes ; PEB) that downweighs neural model parameter estimates from subjects with less reliable data. The neural models used to explain invasive vs non invasive data are different. Although the neural circuitry is the same, the neural model that explains invasive data outputs responses at different depths, while the model that explains non invasive data sums responses across depths. In our earlier work, laminar predictions were fitted to separate electrode tips. These recorded neural activity from superficial and deep cortical layers separately (cf. Figure 7 in Pinotsis et al., 2017). These recordings corresponded to the outputs of neural populations occupying different cortical layers. Here, we used MEG data. Model fitting entailed summing up output responses across all layers. The best fit was achieved via a joint optimization of different weights with which different populations contribute to the MEG signal as well as observation (lead field) parameters of the virtual electrode. Also, the analysis presented here focuses on variability in the structure and function of neural sources between different people that might account for the variability in observed brain responses.
Section snippets
Materials and methods
Data. We used MEG source reconstructed responses reported in (Schwarzkopf et al., 2012). We considered visually induced oscillations between 30−80 Hz from the visual cortex of 16 subjects. In that task, subjects paid attention to the centre of a screen that showed a static, high-contrast, square-wave, vertical grating. MEG data were obtained using a CTF axial gradiometer including 275 sensors, with a sampling rate of 600 Hz. Subject head movement was also recorded and data were preprocessed
Theory and calculations
Here, we present a mathematical proof of the functional equivalence between the compartmental and neural mass model used here and in (Pinotsis et al., 2016a, 2017). We also discuss extensions of our approach that could include computational models predicting data from different modalities (e.g. EEG/fMRI) or tasks (e.g. resting state, task evoked responses). Our proof relies on statistical decision theory. For a primer on this topic, see Appendix A and (Berger, 2013).
Assume two Bayesian decision
Results
As an illustration, we applied our approach to human MEG data from (Schwarzkopf et al., 2012). We fitted power spectra of visually induced oscillations between 30–80 Hz from different people, similar to our previous work (Pinotsis et al., 2013a,b). The corresponding gamma-band frequency correlated with the retinotopically determined surface area of central V1 (Schwarzkopf et al., 2012). Data were recorded from the visual cortex of 16 subjects while they viewed a static, high-contrast vertical
Discussion
The ability to process and store big datasets, known as Big Data, has revolutionised cognitive and computational neuroscience among other fields. We now have the ability to record electrophysiological data at an unprecedented spatial and temporal resolution, e.g. the first generation of Neuropixel probes could record more than 700 neurons from five rat brain structures (Jun et al., 2017). We can also collect and process data from thousands of healthy and patient subjects in large scale clinical
Author contributions
DAP: conceptualization, data curation, analysis and writing the ms.
Acknowledgments
This work was supported by UKRI ES/T01279X/1. I thank Professor Earl Miller and Dr Daniel Gibson for important suggestions. I also thank and Dr Sam Schwarzkopf for informative discussions and providing data.
References (46)
- et al.
O43. Cognitive and emotional biomarkers of anxious major depressive disorder: an iSPOT-D report
Biol. Psychiatry
(2018) - et al.
A disturbance of nonlinear interdependence in scalp EEG of subjects with first episode schizophrenia
Neuroimage
(2003) - et al.
Reduced compartmental models of neocortical pyramidal cells
J. Neurosci. Methods
(1993) - et al.
Processing of global, but not local, motion direction is deficient in schizophrenia
Schizophr. Res.
(2003) - et al.
Dynamic causal modelling of electrographic seizure activity using Bayesian belief updating
Neuroimage
(2016) - et al.
Cortical surface-based analysis: II: inflation, flattening, and a surface-based coordinate system
Neuroimage
(1999) - et al.
Post hoc Bayesian model selection
Neuroimage
(2011) - et al.
LFP and oscillations—what do they tell us? Curr
Opin. Neurobiol.
(2015) When brain rhythms aren’t ‘rhythmic’: implication for their mechanisms and meaning
Curr. Opin. Neurobiol.
(2016)- et al.
Dynamic causal models of steady-state responses
Neuroimage
(2009)
Dynamic causal modeling with neural fields
Neuroimage
Dynamic causal modelling of lateral interactions in the visual cortex
NeuroImage
Contrast gain control and horizontal interactions in V1: a DCM study
NeuroImage
Linking canonical microcircuits and neuronal activity: dynamic causal modelling of laminar recordings
Neuroimage
Predicting violent behavior: what can neuroscience add?
Trends Cogn. Sci. (Regul. Ed.)
Machine learning and big data in psychiatry: toward clinical applications
Curr. Opin. Neurobiol.
Decision making, movement planning and statistical decision theory
Trends Cogn. Sci. (Regul. Ed.)
Bayesian fusion and multimodal DCM for EEG and fMRI
NeuroImage
Synchronization in directionally coupled systems: some rigorous results
DISCRETE Contin. Dyn. Syst. Ser. B
Attentional enhancement of auditory mismatch responses: a DCM/MEG study
Cereb. Cortex
On the nature and use of models in network neuroscience
Nat. Rev. Neurosci.
Statistical Decision Theory: Foundations, Concepts, and Methods
A general framework for updating belief distributions
J. R. Stat. Soc. Ser. B Stat. Methodol.
Cited by (5)
Multi-modal bioelectrical signal fusion analysis based on different acquisition devices and scene settings: Overview, challenges, and novel orientation
2022, Information FusionCitation Excerpt :The purpose of fusion is to select useful features to enhance features (redundancy), or select signals with complementary effects to supplement the deficiencies of other signals. Common fusion methods include weighted average method [180,181], Kalman filter [182,183], Bayesian network [184], statistical decision theory [185], D-S evidence theory [186,187], clustering analysis [188], deep learning fusion [189–191], fuzzy set theory [186,192]. Among them, spatio-temporal feature fusion, parallel (serial) feature fusion, feature reconstruction fusion, weight coefficient fusion are common fusion components.
Multiscale modeling and analytical methods in neuroscience: Molecules, neural circuits, cognition and brain disorders
2021, Journal of Neuroscience MethodsCortical origin of theta error signals
2023, Cerebral CortexDifferences in visually induced MEG oscillations reflect differences in deep cortical layer activity
2020, Communications Biology