Machine learning om de prognose van multiple sclerose patiënten te modelleren

Multiple sclerosis is a neurological autoimmune disorder with onset typically in young adulthood, affecting millions of individuals worldwide. Despite the existence of several disease-modifying therapies, multiple sclerosis can still result in a severe reduction in quality of life. This could be mitigated by administering the right treatment with appropriate efficacy to a patient as soon as possible, as early treatment can delay or even prevent long-term disability accrual. However, finding this patient-treatment match is challenging, as the disease course is very heterogeneous and unpredictable, and the optimal treatment might not even exist yet. Although several data-driven prognostic models have been proposed, the field often focuses only on a single patient outcome represented in a binary classification problem. Additionally, explainability and interpretation remains underexplored, which are crucial in the medical domain for trust in the system and accountability of decisions. In this thesis, we aim to develop novel approaches for modeling prognosis in multiple sclerosis on the patient, group and population level. On the patient level, we advance personalized prognosis for patients, by investigating alternatives to the currently dominant binary classification paradigm predicting whether patients will worsen in a fixed number of years. On the group level, we investigate outcome-informed patient stratification based on multiple time-to-event patient outcomes. On the population level, we improve model and disease understanding, by using explainability techniques and interpreting their results. Three new learning frameworks were developed. First, we propose a multi-target binning approach for modeling sparse sporadically observed time series with different lengths of follow-up, as typically encountered in retrospective observational chronic care datasets. Second, we conduct an extensive comparison of the binary classification paradigm to a survival analysis alternative for time-to-worsening, for multiple sclerosis and beyond. Third, we extend a survival analysis outcome-informed clustering algorithm to the multi-event setting and apply it to composite scores in multiple and systemic sclerosis, achieving methodological and external validation. Explainability and domain knowledge interpretations are interwoven throughout the development of these three methods. However, we also specifically develop a dashboard to explain individual predictions of random forest-based models for the expert end user. Our findings indicate that survival analysis is a valuable replacement for the binary classification paradigm. When thresholded at a specific point in time, the single threshold-agnostic survival analysis model performs equally well as a binary classification model trained for that specific threshold. This observation holds true for a variety of benchmark datasets from different domains: without compromises, a single survival model on the full prediction horizon matches the performance of the specialized binary classification model. For multiple sclerosis, post-hoc explainability techniques applied to the time- to-worsening model showcase that the model uncovered interesting variable interaction patterns. Likewise, the inherent explainability of the multi-event survival-informed clustering algorithm allowed us to identify clusters with higher risks of worsening for some patient outcomes but lower risks of worsening for other outcomes. These explanations are partly corroborated with the domain literature, but also partly remain potential research hypotheses. In conclusion, we have extensively advocated and motivated the competitive performance and increased expressiveness of survival analysis over binary classification for worsening in multiple sclerosis. When integrated together with the multi-target binning approach and local explainability dashboard into a clinical decision support system, this paves the way for personalized medicine. These alternative paradigms allow clinicians to get a more comprehensive view on the future progression of the individual patient, easing the treatment selection process. Additionally, since many patient outcomes are naturally represented as a time-to-event, the proposed multi-event survival-informed clustering algorithm is also broadly applicable for patient stratification purposes. When developing novel treatments, this algorithm can be used to split the candidate cohort for a clinical trial into more homogeneous subgroups. More homogeneous means equal statistical power for lower sample sizes, and thus overall more cost-effective drug trials. Finally, we show how explainability techniques can be applied and interpreted to increase model and disease understanding. These explanations could generate scientific hypotheses to be tested, as we showcased for the prognostic relevance of a biomarker called CHIT1.