High-dimensional counterfactual modelling of lesion-deficit relations in focal brain injury

Lead Research Organisation: University College London
Department Name: Institute of Health Informatics

Abstract

We wish to examine the fidelity of multivariate artificial intelligence (AI) models of observational data in the task of predicting individual response to treatment in focal brain injury such as stroke. Such complex "high-dimensional" models will be compared against the simple "low-dimensional" randomised controlled trials in current use, testing the ability of AI models to resist the bias in treatment allocation typical of observational data. Employing simulated ground truths based on hypothetical treatments selectively effective for different subtypes of stroke presenting with common symptoms (such as visuospatial neglect or anomia), model types including Gaussian processes and support vector machines are compared in their abilities to predict individualised susceptibility. By simulating bias, the resilience of each model is determined, with various parameters including outcome noise, dimensionality, and number of participants. These methods will be translated to empirical observational data involving a ground truth of thrombectomy vs thrombolysis. The circumstances for which it is advantageous to use multivariate modelling for individualised treatment prescription in comparison to traditional randomised controlled trials will be determined.

Publications

10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/S021612/1 01/04/2019 30/09/2027
2252409 Studentship EP/S021612/1 01/10/2019 30/09/2023 Dominic Matthew Giles