Introducing machine learning for full MS patient trajectories improves predictions for disability score progression.

Abstract

A. Background and Goals: It is now well known that patient medical history (or trajectories) is of crucial importance for both prognosis and optimal treatment selection for Multiple Sclerosis (MS) patients. This longitudinal data is nowadays being collected more systematically and its availability in patient registries opens the way towards precision medicine in MS. However, this information is very challenging to process computationally. By their observational nature, longitudinal patient data are scarcely and irregularly observed (i.e. only few observations are scattered over the whole patient history). Several works have recently attempted to address this issue by substituting full patient history with summary statistics in the modelling (i.e. using metrics such as the maximum and minimum disability score over the patient history as a proxy for the full medical trajectory). This strategy potentially discards relevant prognostic information. In this work, we present a method capable of handling the full information about disease history and its value for individual prognostic modelling. B. Method(s): We propose a novel statistical method relying on state-of-the-art machine learning models (Bayesian Probabilistic Tensor Factorisation, BPTF). Due to its generative nature, this method is able to process the full MS patient trajectories as input to deliver predictions. We infer the tensor decomposition (rank 70) with Gibbs sampling. To illustrate the performances of our approach, we consider the challenging task of predicting patient worsening within 2 years from current visit using 3 years clinical history. Patient worsening was defined with the 2 strata definition of disability progression as in Kalincik et al. (Brain, 2015). Available longitudinal data include Expanded Disability Status Scale (EDSS), magnetic resonance imaging and relapses. Patient static covariates include, among others, age at onset, gender, disease course and first observed EDSS. C. Result(s): We show that our method results in an AUC of 0.79, which outperforms baseline models using only static covariates by over 10 points of AUC and provide an improvement of 4 points compared to methods using summary statistics for the trajectories (maximal difference in EDSS, last observed EDSS and highest observed EDSS over the 3 years observation window). These results confirm that use of longitudinal information sets with BPTF improves the accuracy of prognostic models in MS.