Shahin, Ahmed H.; (2025) Multimodal Machine Learning for Prognostic Modelling in Idiopathic Pulmonary Fibrosis. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text Main.pdf - Submitted Version Download (13MB) | Preview

Abstract

Idiopathic Pulmonary Fibrosis (IPF) is a severe lung disease characterized by rapid progression and high mortality, with a highly variable prognosis between patients. This thesis leverages machine learning to enhance prognosis prediction in IPF by analysing clinical data and volumetric imaging. We first address the challenge of missing data in patient records by applying latent variable models to accurately impute missing attributes based on the available information in each record. Next, we use the Cox proportional hazards model to predict mortality risk from patient data. As a ranking objective, the Cox model requires many samples per training iteration, which is computationally expensive and often infeasible for volumetric data. We introduce a scalable memory bank-based training approach for efficient model training with volumetric data. Recognizing the inherent constraints of the Cox model, we also propose a new method, CenTime, which better utilizes censored data and directly predicts the time-to-mortality. CenTime relaxes the assumptions of the Cox model, provides a more precise estimation of patient outcomes, and leverages right-censored data more effectively. Our methods are validated on a comprehensive dataset of IPF patients, demonstrating significant improvements in prediction accuracy over existing approaches. This work can advance personalized prognosis in IPF, aiding clinicians in developing tailored treatment strategies.

Download activity - last month

Export as CSVXMLJSON

Download activity - last 12 months

Export as JSONCSVXML

Downloads by country - last 12 months

Export as CSVXMLJSON

Archive Staff Only

View Item