Stacking models of brain dynamics improves prediction of subject traits in fMRI

1429 

Abstract Submission 

Ben Griffin¹, Christine Ahrends², Fidel Alfaro Almagro³, Mark Woolrich⁴, Stephen Smith⁴, Diego Vidaurre⁵

¹Oxford University, Oxford, Oxford, ²Aarhus University, Aarhus, Aarhus, ³WiN FMRIB - University of Oxford, Oxford, Oxfordshire, ⁴University of Oxford, Oxford, Oxfordshire, ⁵Aarhus University, Aarhus, Aarhus C

Ben Griffin  
Oxford University
Oxford, Oxford

Christine Ahrends  
Aarhus University
Aarhus, Aarhus

Fidel Alfaro Almagro  
WiN FMRIB - University of Oxford
Oxford, Oxfordshire

Mark Woolrich  
University of Oxford
Oxford, Oxfordshire

Stephen Smith  
University of Oxford
Oxford, Oxfordshire

Diego Vidaurre  
Aarhus University
Aarhus, Aarhus C

Beyond structural and static brain measures, brain dynamics can add important information when investigating individual cognitive traits [1]. One way to look at dynamics is through the Hidden Markov Model (HMM), a probabilistic model of temporal dynamics of functional connectivity, which can be combined with machine learning models to generate subject-specific trait predictions [1]. However, there are two potential sources of variability in these predictions. First, the run-to-run variability (e.g., due to different initialisations) [2]; and second, the choice of hyperparameters (e.g., number of HMM states). We propose an approach that leverages this variability to improve prediction accuracy by combining information from different models and runs. We implement a stacked generalisation scheme [3, 4] that combines predictions from multiple models of brain dynamics to accomplish two goals: (i) produce robust predictions across multiple cognitive traits; (ii) improve prediction accuracy by combining predictions created by HMMs with varying hyperparameters.

We aimed to predict cognitive traits for subjects in the Human Connectome Project (HCP) [5] and UK Biobank [6] datasets using models based on their resting-state fMRI functional connectivity dynamics. To achieve this, we used a group-level HMM on surface-based node timeseries concatenated across all subjects, which were generated using 25-dimensional group-ICA surface maps [1, 7]. Subsequently, subject-specific HMM parameters were "dual-estimated" using the parameters obtained from the group-level HMM, which involved generating subject-specific state-time courses, from which subject-specific HMM parameters were inferred [1]. We then used the Fisher kernel method, a mathematically principled way of combining generative models (here, the HMM) with discriminative methods (here, kernel ridge regression) [8]. This method operates directly on the Riemannian manifold of the HMM parameters, capturing the natural geometry of the HMM. We are able to construct a The method constructs a subject-by-subject matrix that quantifies the similarities between subjects by examining how the HMM likelihood function varies locally with respect to the subjects' HMM parameters. This entire process was repeated 100 times; applying the HMM 50 times with fixed hyperparameters but different random initialisations (to investigate the HMM run-to-run variability) and 50 times with varying hyperparameters (to investigate hyperparameter selection variability). Separately for each group-level HMM, Fisher kernels were computed and used with kernel ridge regression to produce subject-specific predictions, which were subsequently combined using stacking [3].

By combining predictions generated using the Fisher kernel method from group-level HMMs, we achieve our first goal (i) of producing a robust prediction across traits. Furthermore, by varying the hyperparameters of the HMM, our second goal (ii) of improving prediction accuracy is achieved. The diversity created by varying the HMM hyperparameters enabled distinct yet complementary predictions to be generated, and stacking resulted in boosting the prediction accuracy. Stacking performs particularly well when, for a given trait, certain predictions are much better than the remaining predictions (e.g., see Figure 2b 'Pic_Vocab_Unadj').

For predictions to be useful (e.g., in a clinical setting), we require that they are both robust and accurate. We used the Fisher kernel method, leveraging the natural geometry of the HMM, to generated accurate predictions. To enhance robustness and accuracy further, we employed stacking, combining predictions from different HMM configurations that captured distinct patterns in the data. Looking forward, stacking predictions opens up avenues for integrating a wider variety of data or models, for example combining predictions from different brain imaging modalities, or static functional connectivity and structural information.

Classification and Predictive Modeling ¹

fMRI Connectivity and Network Modeling

Methods Development ²

FUNCTIONAL MRI

Machine Learning

Modeling