A comparison of approaches for combining predictive markers for personalised treatment recommendations.

摘要: BACKGROUND In the presence of heterogeneous treatment effects, it is desirable to divide patients into subgroups based on their expected response treatment. This formalised via a personalised recommendation: an algorithm that uses biomarker measurements select treatments. It could be multiple, rather than single, biomarkers better predict these subgroups. However, finding optimal combination multiple can difficult prediction problem. METHODS We described three parametric methods for in recommendation, using randomised trial data: regression approach models outcome by interactions; proposed Kraemer forms combined measure from individual weights, calculated all treated and control pairs; novel modification Kraemer's utilises prognostic score sample matched subjects. Using Monte Carlo simulations under data-generating models, we compare approaches draw conclusions improvement recommendation compared standard The are applied data home-delivered pragmatic rehabilitation versus as usual with chronic fatigue syndrome (the FINE trial). Prior analysis this indicated some effect heterogeneity correlated biomarkers. RESULTS outperformed across scenarios. leads improved recommendations, except case where there was strong unobserved biomarker. example, method weak its algorithm. CONCLUSIONS does not perform combining All sensitive misspecification models.