Deep Contrastive Learning For Dynamic Survival Analysis With Competing Risks Study

Survival analysis corresponds to a set of statistical approaches used to investigate whether the event of interest occurred and the time it takes for an event to occur.With the development of informatization technology,the ability to collect a wide variety of data and monitor the observation over long-term periods have been attained in multiple disciplines,which provides a solid foundation for survival analysis studies.Traditional survival analysis models are statisticsbased,which can be divided into parametric,non-parametric and semi-parametric models.Although valuable,statistics models are not suitable for the survival prediction scenarios where the non-linear correlation between event time and patient covariates exists.To alleviate this limitation,machine learning-based models have been developed to survival analysis without assumptions about the form of the underlying distribution of survival time process to estimate the occurrence time of events.However,there are still some problems for machine learning-based survival analysis: 1)Most survival analysis models lack the ability to estimate the distribution of hitting times of more than one event of interests(i.e.,competing risks)2)Most existing studies address the survival analysis problem in a static manner,i.e.,they just utilize a fraction of longitudinal data,ignore the correlations between multiple visits;3)Current models need capture the potential representation of patient better for downstream tasks such as outcome estimation and survival analysis.To address these challenges,this thesis carries out an in-depth study on survival analysis with competing risks.Specifically,this thesis proposes a contrastive-based model for dynamic deep survival analysis,named CD-Surv,to make full use of longitudinal observational data and improve the modeling expression.In CD-Surv,the shared encoder composed of LSTM network is used to transform the trajectories to latent representations.Based on contrastive learning,two novel data augmentation strategies,i.e.,mask-generation and shuffle-generation,are designed to capture the potential representation better and improve the performance of survival analysis.Since most survival analysis methods do not take competing risks into account,this thesis proposes a deep contrastive learning for dynamic survival analysis with competing risks,named Deep-CSA,which combines multi-task learning and CD-Surv,to further improve the performance of survival analysis with competing risks.In this thesis,the proposed models are evaluated by using electronic health record data from a hospital in China and two open-sourced clinical datasets,i.e.,MIMIC-IV and EICU.The experimental results show that: 1)The CD-Surv and Deep-CSA can dynamically predict the survival status of a target patient;2)The contrastive learning module in CD-Surv and Deep-CSA can optimize the latent representation of patients;3)After taking the multi-task learning framework into account,Deep-CSA can predict the occurrence of multiple events better and have good performance on the competing risks problem.The methodology proposed in this thesis outperforms the state-of-the-art survival analysis methods,by means of modeling the covariates of patients in a fine-grain manner to provide dynamic survival analysis with competing risks.