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

Beyond structural and time-averaged functional connectivity brain measures, the way brain activity dynamically unfolds can add important information when investigating individual cognitive traits. One approach to leveraging this information is to extract features from models of brain network dynamics to predict individual traits. However, there are two potential sources of variation in the models' estimation which will in turn affect the predictions: first, in certain cases, the estimation variability due to different initialisations or choice of inference method; and second, the variability induced by the choice of the model hyperparameters that determine the complexity of the model. Rather than merely being statistical noise, this variability may be useful in providing complementary information that can be leveraged to improve prediction accuracy. We propose stacking, a prediction-driven approach for model selection, to leverage this variability. Specifically, we combine predictions from multiple models of brain dynamics to generate predictions that are accurate and robust across multiple cognitive traits. We demonstrate the approach using the Hidden Markov Model, a probabilistic generative model of brain network dynamics. We show that stacking can significantly improve the prediction of subject-specific phenotypes, which is crucial for the clinical translation of findings.