This master's thesis addresses the prediction of time to confirmed disease progression in multiple sclerosis (MS) on the basis of morphometric features computed from magnetic resonance (MR) images of the head. Unlike most existing approaches, which treat progression as a dichotomous outcome within a predetermined time window, the problem is formulated as a survival analysis at the level of the individual visit (per-visit), in which each MR measurement constitutes an independent observation with its own time to event.
The data come from the AI ProMiS study: 166 patients with a total of 612 MR visits from three Slovenian clinical centres, with progression according to the CDP-6 criterion recorded at 9.5 % of visits. From volumetric measurements of brain structures, seven feature sets were assembled (volumes, inter-hemispheric asymmetry, z-index), and four temporal transformations were applied to each (raw values, difference, slope, and principal component analysis). On the 28 datasets thus obtained, random survival forests (RSF), the survival support vector machine, and four deep learning models from the PyCox library were compared, with a conventional random forest and the Cox proportional hazards model serving as references.
The best performance was achieved by the RSF model on the slope-transformed z-index feature set, which reached a C-index of 0.634 ± 0.077 in 5-fold cross-validation and 0.689 on an external hold-out set from a different clinical centre; the time-dependent AUROC remains between 0.66 and 0.69 for horizons from 24 to 60 months. The predictions are explained using Bellatrex for local explanations and SHAP for a global assessment of feature importance. Both methods rank morphometric features — asymmetries of cortical regions, volumes of cortical gyri, and the fourth ventricle — ahead of demographic and clinical variables among the most important predictors. The results show that the time to MS progression can be modelled from MR features even in a relatively small cohort; the principal limitation remains the small number of cases.
|