Delirium, an acute confusional state, is unfortunately common in Adult Intensive Care Units (AICUs). Its development heralds increased risk of morbidity, mortality, increased length of hospital stay, and worse cognitive outcomes following discharge with cognitive scores not too dissimilar to Alzheimer’s dementia. In this thesis, we explore the use of machine learning techniquesfor the classification of delirium from Electronic Patient Records (EPR) data routinely collected as part of the patient’s clinical care. We hypothesise that delirium can be classified by eye-movements as a measurable biological signal of overt attention as a surrogate marker for covert cognitive attention.

We thus developed, and validated, an eye tracking platform that exceeds empiric requirements of invasiveness, performance, accuracy, and precision. Following appropriate approvals, we then deployed the platform across two hospitals in London. The data gathered facilitated the training of machine learning algorithms that can classify delirium based on eye-movements and scene-information with an Area Under Receiver Operator Curve (AUROC) of 0.69 and 0.78 respectively, and an Area Under the Precision Recall Curve (AUPRC) of 0.8 and 0.83 respectively. These classifiers also provided novel insights into the visual attention of delirium patients.

To explain these classifier’s findings, we further develop a novel, theory of mind inspired, architecture composed of multiple hierarchical competing forward models. Using this architecture, we demonstrate that patients with delirium have globally reduced levels of visual attention throughout the visual processing hierarchy to a statistically significant degree (p = 0.044).