Predicting Treatment Outcome Through Patient Subgroup Evolution-A Multi-Layer Snapshot Network Approach

Precision medicine involves the stratification of patients into more homogeneous subgroups in order to tailor treatment decisions to the characteristics of each patient. Traditional approaches assume static subgroups. However, patient data (e.g. laboratory values, symptoms, or self-reported health) can change substantially. These changes may be indicative of the success of treatment, i.e. predict treatment outcome. We propose a method that models patient subgroups and their evolution during treatment on a set of multi-layer snapshot networks (MLSNs) and exploits subgroup transitions to augment the prediction of treatment outcomes. Next to inter-feature similarity and intra-feature similarity, we formalize patient migration across subgroups. We further introduce a mechanism that assigns new patients to the subgroups without reconstruction of the network. We evaluate our method on self-report questionnaire data of patients with chronic tinnitus from a multi-center randomized clinical trial. We demonstrate that regularized regression models predicting the treatment outcome perform better when subgroup information is added to the feature space. In our experiments comparing with conventional clustering algorithms, the predictive performance of the models using subgroups found with our method was competitive, despite using a smaller feature subset. We further demonstrate that the proposed strategy, which involves grouping new patients into pre-existing subgroups and using this information to predict treatment outcomes, outperforms a scenario in which subgroup information is not used.