Machine learning for precision medicine and health economics: a novel model for survival analysis and subgroup identification
Background: The emergence of precision medicine, which is described as an approach that tailors health interventions to a group of patients based on their characteristics, brings challenges to Health Technology Assessment (HTA). HTA seeks to provide policymakers with the necessary information to better understand the benefits of health technologies and make better funding decisions, and ultimately improve resource allocation. Precision medicine differs from traditional medicine in that it is based on patient subsets, whereas traditional medicine is based on the entire patient population. Therefore, identifying appropriate patient subgroups is an important part of the evaluation of precision medicines.
Objective: This research aims to develop Machine Learning tools to address the subgroup identification problem in precision medicine.
Methodology: A novel Machine Learning model is proposed in this research using Multi-Task Learning and Support Vector Machine. The survival analysis problem is decomposed into a series of classification problems, and Support Vector Machine is applied to improve the classification accuracy. Moreover, a ℓ₂,₁ norm is used for feature selection. This model has two functions: (1) predict event times based on censored data; (2) select important covariates.
Results: The prediction accuracy of the proposed model is compared with two benchmark statistical methods, Cox-LASSO and Cox-EN, and three state-of-the-art Machine Learning methods. The proposed model outperforms the other methods. The proposed model significantly outperforms the Cox models in terms of feature selection. In the simulation study, the proposed model outperforms the other methods in selecting patient subgroups with enhanced treatment effects
Conclusion: The model proposed in this research outperforms several other state-of-the-art methods in time-to-event prediction and feature selection. Therefore, the proposed model can be used for precision medicine to select high-risk patients or patient subgroups with enhanced treatment effects and has potential applications in other fields involving survival analysis.