Computational limits to the legibility of the imaged human brain

Abstract

Our knowledge of the organisation of the human brain at the population-level is yet to translate into power to predict functional differences at the individual-level, limiting clinical applications and casting doubt on the generalisability of inferred mechanisms. It remains unknown whether the difficulty arises from the absence of individuating biological patterns within the brain, or from limited power to access them with the models and compute at our disposal. Here we comprehensively investigate the resolvability of such patterns with data and compute at unprecedented scale. Across 23 810 unique participants from UK Biobank, we systematically evaluate the predictability of 25 individual biological characteristics, from all available combinations of structural and functional neuroimaging data. Over 4526 GPU*hours of computation, we train, optimize, and evaluate out-of-sample 700 individual predictive models, including fully-connected feed-forward neural networks of demographic, psychological, serological, chronic disease, and functional connectivity characteristics, and both uni- and multi-modal 3D convolutional neural network models of macro- and micro-structural brain imaging. We find a marked discrepancy between the high predictability of sex (balanced accuracy 99.7%), age (mean absolute error 2.048 years, R² 0.859), and weight (mean absolute error 2.609Kg, R² 0.625), for which we set new state-of-the-art performance, and the surprisingly low predictability of other characteristics. Neither structural nor functional imaging predicted an individual's psychology better than the coincidence of common chronic disease (p < 0.05). Serology predicted chronic disease (p < 0.05) and was best predicted by it (p < 0.001), followed by structural neuroimaging (p < 0.05). Our findings suggest either more informative imaging or more powerful models will be needed to decipher individual level characteristics from the human brain. We make our models and code openly available.

Keywords: Artificial intelligence; Brain function; Brain structure; Deep learning; Functional connectivity; High performance computing; Machine learning; Magnetic resonance imaging; Neuroimaging.

Fig. 1

Sum of all lesion images in MNI space. Indices refer to z coordinates in MNI space.

Fig. 2

NeuroQuery meta-analytic map of activations associated with the keyphrase “Stroop”. The connected component located on the medial wall, encompassing the ACC, has been outlined in white.

Fig. 3

Graph lesion-deficit mapping of phonemic fluency showing the behaviour-associated mean edge weights for SBM-defined regions significantly associated with the behaviour compared with lesion co-occurrence. Note inferred dependence on left prefrontal regions, including within the ACC ROI (outlined in white).

Fig. 4

Graph lesion-deficit mapping of Stroop showing the behaviour-associated mean edge weights for SBM-defined regions significantly associated with the behaviour compared with lesion co-occurrence. Note inferred left frontotemporal dependence, excluding the ACC ROI (outlined in white).

Fig. 5

Schematic illustrating the impossibility of “internal” error feedback at a posited controller stage where the behaviour requires learning from success or failure of the achieved goal. A controller, Cr, directs two controlled substrates, Cd, producing two different actions, A, in pursuit of a goal, G. Since it is definitional of a controller that is has knowledge of what it controls, learnt operation of this circuit must receive an error signal at the controlled substrate stage. But this is impossible, for the definition of success or failure in the taskd−achieving the goal−is available only from the final action and its sensorially registered consequences, spanning both controller and controlled.