LAUSR

employed multivariate Cox proportional hazard models on selected clinical variables. Training was performed on defined sets of cases identified molecularly similar for each individual case, and on all cases in the database; Cox proportional hazard models including clinical variables and molecular categories was also learned, as an additional comparison. Various models were assessed by the average time dependent Brier score. Results: Gene set enrichment analysis showed that activated B‐cell like signatures and cell cycle signatures are highly enriched amongst prognostically unfavourable genes, while germinal centre B‐cell like signatures and signatures associated with immune response are highly enriched amongst favourable genes. Figure 1 shows that personalised survival models learned from various subsets of molecularly similar patients provide superior prognostic predictions. Conclusion: These preliminary results support the use of patients with molecularly similar disease from large databases in personalized prognostic and predictive modelling, an approach that should increase in power as patient databases grow in size and molecular content.