Kurz, Jason A.Sleigh, James W.Mayo, Christopher2026-08-242026-08-242026https://hdl.handle.net/10289/18576Nociception is defined as the neural encodings of noxious stimuli, and is a key dimension of general anaesthesia that clinicians seek to control. The use of machine learning (ML) to predict intraoperative nociception from single-channel electroencephalography (EEG) data is poorly established, with existing indices relying on opaque proprietary algorithms or limited validation. A key challenge is the lack of reliable ground truth: many prior studies evaluate models solely on their ability to predict postoperative pain scores, which may not accurately reflect intraoperative nociceptive processes. This thesis addresses this problem through two complementary contributions. Firstly, a hypothesis-driven unified validation framework was created to evaluate nociception indices against multiple physiologically motivated criteria, including responsiveness to surgical incision and analgesic administration, rather than relying solely on postoperative pain scores. Secondly, a range of ML approaches were developed to predict nociception from EEG. A dataset comprising 1,541 total hours of intraoperative EEG across 567 patients from three previous studies was collated, preprocessed, and aligned with drug dosage and surgical event data. This was used to develop and evaluate putative nociception indices using a range of ML approaches, including unsupervised approaches, representation learning methods to reduce confounding, and a novel EEG-informed pharmacodynamic non-linear mixed effects model (EEG-PD-NLME) that exploits fentanyl pharmacology to provide a supervisory signal. The best-performing candidate indices were a principal component-based index (PCA-6) and an EEG-PD-NLME model with additional adversarial training (NLME+adv.), satisfying most of the criteria in the proposed framework. However, despite this, both models performed poorly when predicting postoperative pain (AUROC 0.511 and 0.388, respectively). This discrepancy suggests a mismatch between intraoperative nociception and postoperative pain, challenging the use of postoperative pain scores as a validation endpoint. A well-validated and well-described index of nociception could be utilised clinically, allowing for improved titration of analgesic medications during surgery. This may lead to improved postoperative outcomes such as reduced postoperative pain. More broadly, this thesis demonstrates the importance of comprehensive evaluation when learning latent physiological variables in the absence of ground truth data.enAll items in Research Commons are provided for private study and research purposes and are protected by copyright with all rights reserved unless otherwise indicated.artificial intelligenceEEGanaesthesiaThesis