Individual disease course and treatment response prediction using Machine Learning

Multiple Sclerosis (MS) is a chronic and debilitating disease that affects approximately 2.8 million individuals worldwide. MS is complex, unpredictable and currently incurable, only managed through disease-modifying treatments (DMTs) that aim to improve life quality and slow disease progression. Despite the availability of over 25 approved DMTs, there remains a significant gap in understanding the variability in patient responses to these treatments, which impacts the effectiveness of MS management.

We have access to the unique Novartis-Oxford MS (NO.MS) dataset, which is a collection of longitudinal data from over 34,000 individuals from a collaboration between Novartis, the Oxford Big Data Institute (BDI), and MS physicians. Leveraging this dataset, this project aims to develop sophisticated statistical machine learning methods,
focusing on predictive and generative models, to advance our understanding of MS by characterising and predicting individual treatment responses and mapping disease trajectories. Specifically, we implement and extend Bayesian Additive Regression Trees (BART) to causally estimate individual treatment effects, modelling scanner effects separately from biological factors and introducing patient-specific random effects to account for variations in treatment responses over time. This innovative approach allows us to characterise the variability in treatment outcomes between patients and identify biomarkers driving this heterogeneity, ultimately informing the development of new treatments. We then further refine these models to produce more accurate estimates of individual treatment effects (ITE), extending to multivariate outcomes and exploring the use of whole-brain data to extract richer biological features, thereby enhancing predictive accuracy. By leveraging these refined estimates and biomarkers, we aim to cluster patients into treatment efficacy groups. Additionally, we establish innovative dynamic and causal models to predict the future course of the disease, extending our models to survival/time-to-event outcomes and integrating intervention times to analyse their impact on disease progression. Ultimately, we aim to develop novel generative models capable of predicting multiple potential disease trajectories for each patient, dynamically adjusting based on individual biomarkers, treatment history, and potential future interventions, effectively creating "Digital Twins" for each patient.

We develop models capable of handling both cross-sectional and longitudinal data, accommodating univariate and multivariate outcomes, as well as managing data missingness, and we incorporate uncertainty quantification to ensure reliable predictions. The rich and extensive longitudinal NO.MS dataset provides a unique opportunity to develop and refine our methods, allowing us to advance the state-of-the-art in MS research.

The potential impact of this work is profound, as it aims to reduce the unpredictability of MS, ultimately leading to better-informed treatment decisions and improved patient outcomes. Moreover, we anticipate our innovative methods could be applied to the study of different diseases.

This project falls within the EPSRC Artificial Intelligence technologies and Clinical Technologies (excluding imaging) research areas, contributing to the healthcare technology theme.