Poster E81, Monday, March 26, 2:30-4:30 pm, Exhibit Hall C
The limits of behavioural outcome prediction following focal brain injury
Tianbo Xu1, Ashwani Jha1,2, Hans Rolf Jager1,2, Michel Thiebaut de Schotten6,7, Geraint Rees1,3,4,5, Parashkev Nachev1,2; 1Institute of Neurology, UCL, London, WC1N 3BG, UK., 2National Hospital for Neurology and Neurosurgery, Queen Square, UK., 3Institute of Cognitive Neuroscience, UCL, London WC1N 3AR, UK., 4Faculty of Life Sciences, UCL, London, WC1E 6BT, UK., 5Wellcome Trust Centre for Neuroimaging, UCL, London WC1N 3BG, UK., 6Brain Connectivity Behaviour group, Paris, France, 7Sorbonne Universités, UPMC Univ Paris 06, Inserm, CNRS, Institut du cerveau et la moelle (ICM) - Hôpital Pitié-Salpêtrière, Boulevard de l’hôpital, F-75013, Paris, France
Our capacity to predict the behavioural outcome of focal brain injury is commonly conceived to be limited by the complexity and variability of the underlying functional anatomy. But a more proximal limit is set by the spatial characteristics of the focal lesions through which lesion-outcome relationships are established in the first place. Here we sought to quantify this hard upper limit on outcome prediction by evaluating a comprehensive set of hypothetical lesion-deficit classification models with the largest published collection of anatomically registered maps of ischaemic focal brain injury (N=1172). Theoretically maximal predictive performance was quantified for lesion-deficit models assuming either single-region or two-region functional dependence, across the space of all possible two-region combinations, defined by a Brodmann area parcellation of the brain. Predictions derived from mass-univariate analyses yielded an average theoretical maximum receiver operating characteristic (ROC) area under the curve (AUC) of 0.923 (95%CI=0.914-0.933) for two area models, and 0.981 (95%CI=0.978-0.983) for one area models, whereas predictions derived from high-dimensional multivariate analyses yielded an average theoretical maximum AUC of 0.963 (95%CI=0.958-0.968) and 0.988 (95%CI=0.986-0.990), respectively. These results show that the spatial structure of natural focal lesions limits not only the fidelity of functional-anatomical inferences drawn from them but also the fidelity of predictive models of behavioural outcome that are agnostic of the underlying anatomy. High-dimensional multivariate methods lift this limit higher than the mass-univariate methods in widespread current use, and ought to be preferred.
Topic Area: METHODS: Neuroimaging