Head motion does not drive success of dynamical classification approaches for fMRI

1484 

Abstract Submission 

Grace Huckins¹, David Glahn², Russell Poldrack³

¹Stanford University, Stanford, CA, ²Department of Psychiatry, Harvard Medical School, Boston, MA, ³Stanford University, Palo Alto, CA

Grace Huckins  
Stanford University
Stanford, CA

David Glahn  
Department of Psychiatry, Harvard Medical School
Boston, MA

Russell Poldrack  
Stanford University
Palo Alto, CA

Historically, analysis of resting-state fMRI has largely focused on static measures like functional connectivity. In contrast, dynamical approaches remain underexplored, but they do present challenges. In particular, head motion in the scanner introduces artifacts that can exert a strong effect on the results of analyses [3], but throwing out high-motion volumes would disrupt the dynamical structure of the data.

In work presented at OHBM 2023, we demonstrated that a dynamical, hidden Markov model-based approach can successfully classify resting-state fMRI data and achieve higher accuracy than static approaches. Those previous analyses, however, did not account for the fact that head motion differences may have been driving classification performance. In the current study, we establish that the success of dynamical classification methods is not driven by head motion artifacts. Additionally, we apply our dynamical classification approach to a problem where head motion is a particular concern: classifying resting-state fMRI data according to psychiatric diagnosis.

All of the experiments in this study used the same HMM-based dynamical classification approach, which is detailed in Figure 1. For the first portion of this study, we used data from the MyConnectome dataset, which includes dozens of resting-state scans from a single individual. Data were preprocessed as in [1]. We previously demonstrated that dynamical classification can distinguish runs where the subject had and had not ingested caffeine; for that study, high-motion volumes were censored and censored volumes were reconstructed using linear interpolation. To test the impact of interpolation on model performance, we generated two versions of the MyConnectome data-one containing only the continuous sequences of low head-motion volumes, and the other containing interpolated sequences. Dynamical classification performance was evaluated using the SSM Python package.

The second portion of this study used a dataset including individuals with and without a diagnosis of psychosis [2]. Data were preprocessed using fMRIPrep and XCP_D, and high-motion volumes were removed and reconstructed using linear interpolation. Dynamical classification of the psychosis data was performed using the Dynamax Python package, which implements the same models as the SSM packing using Jax and is therefore significantly faster (but cannot handle ragged sequences).

As shown in Fig 2a, linear interpolation over censored volumes appears to have had no effect on classification performance. Across a wide range of hidden states, MyConnectome classification performance was virtually identical, whether or not head motion had been fully censored or subject to interpolation. In both cases, performance was significantly higher than the chance level of 50%, despite the fact that the amount of data had been significantly limited by censoring high-motion volumes.

In contrast, Fig 2b demonstrates that neither the dynamical classification approach nor a baseline approach, which applied linear SVM to functional connectivity matrices, was able to achieve above-chance performance in distinguishing individuals with psychosis from healthy controls.

This study shows that the success of dynamical classification on MyConnectome rsfMRI data is not driven by differences in head motion between classes. Completely eliminating high head motion volumes from the dataset had no effect on classification performance. This HMM-based classification strategy is thus a robust and effective method for classifying rsfMRI data, and its high performance indicates that examining whole-brain dynamics in resting-state data may be a fruitful path forward for the field.Additionally, the fact that the dynamical approach was unable to distinguish two groups who differed in their head motion frequency (people with psychosis and healthy controls) provides further evidence that the success of dynamical classification is not driven by head motion.

Psychiatric (eg. Depression, Anxiety, Schizophrenia)

Classification and Predictive Modeling ¹

fMRI Connectivity and Network Modeling

Task-Independent and Resting-State Analysis ²

Computational Neuroscience

Data analysis

DISORDERS

FUNCTIONAL MRI

Machine Learning

Modeling

Psychiatric

Psychiatric Disorders

Schizophrenia