Workshop: Machine learning for Clinical Outcome Prediction

• Radboud University/Radboud UMC: Ton Coolen (Department of Biophysics, Donders Institute), Kit Roes and Marianne Jonker (Biostatistics, Radboud UMC)
• External Partners: SMART B.V., King’s College London, University of Manchester, Saddle Point Science Ltd.

The analysis of data from clinical trials or medical databases usually aims to quantify patterns that allow us to predict patients’ clinical outcomes from clinical, genomic, molecular, or environmental variables and to understand underlying biological processes. In many clinical settings, outcome data take the form of event times, e.g. time to death, cancer relapse, or stroke. Predictive regression from time-to-event data is called survival analysis. Many traditional survival analysis methods are now recognized to be outdated and unable to handle the complexities of modern medical data, such as high dimensionality, disease heterogeneity, informative censoring by comorbidities, or confounding factors. Modern statistical and computational survival analysis methods often have a significantly higher level of complexity and are therefore often computationally challenging. Hence, progress in survival analysis may benefit from combining transparent statistical approaches with clever model parametrizations and powerful machine learning (ML) implementations.

In our workshop, we focus on one specific problem that is especially prominent in cancer research: survival analysis for longitudinal medical data. Patients do not only provide baseline data but are monitored and treated for extensive periods, giving us additional rich longitudinal information. Unfortunately, there are presently few survival analysis methods that are simultaneously sufficiently powerful and computationally feasible for the effective predictive analysis of longitudinal data. Most do not model the interplay between longitudinal observations and clinical interventions and their crucial combined impact on survival.

We will organize a small one-week technical workshop with around ten participants on novel statistical and interpretable machine learning methods for the predictive analysis of longitudinal medical data. Discussion sessions to generate new ideas (morning sessions) will alternate with break-out sessions (afternoons), where these new ideas undergo preliminary scrutiny and feasibility testing. This workshop will bring together several stakeholders in this domain, from Radboud’s Science Faculty (Donders Institute, Data Science Group), Radboud UMC, King’s College London, the University of Manchester, and from the private sector (two SMEs involved in medical data analytics: SMART BV from Nijmegen and Saddle Point Science Ltd from London).