Survival Analysis Based On Deep Learning

Survival analysis for medical focuses on the hazard of patient experiencing a target endpoint event(e.g.death)during the study time period and the potential relationship between patient features and the target endpoint event.It has a widespread application in the prognostic analysis of clinical diseases.Depending on the number of target endpoint events,survival analysis models can be divided into multi-target competing risks and single events.Existing survival analysis methods have the following problems:(1)Conventional feed-forward neural networks have limited learning capacity,which makes it difficult to learn the interactions between features and deep representations of data.(2)Most competing risk models assume that all target endpoint events are independent of each other and do not account for the fact that model performance can be affected by an uneven distribution of events.(3)The Cox model is the most commonly used model in single events.It assumes a linear relationship between features and hazard,which cannot deal with non-linear relationships and time-dependent effects.(4)How to design a more unified network architecture,applicable to both single events and competing risks,is also an open question to be addressed.To address these problems,we proposed to use Transformer to extend traditional survival analysis methods to enhance the model accuracy and interpretability.Specifically,to address the problem of limited learning capability of conventional neural networks,we used the multi-headed attention mechanism of Transformer to establish interactions between features and use stacked self-encoders to extract the depth representation of the input data layer by layer in order to improve prediction accuracy.Different from the assumption of event independence of target endpoints in conventional competitive risk models,Transformer-based Competing Risks(TransCR)model was proposed in this paper.TransCR model does not assume any underlying survival distribution and uses inverse propensity score technology to reduce result bias due to unbalanced event distributions.In the single events model,to automatically learn the non-linear relationship between features and the time-dependent effect of features,we proposed a TransCox model and a time-dependent TransCoxTime model based on the combination of Transformer and Cox models.Finally,to deal with competing risks and single events simultaneously,we fused TransCR,TransCox and TransCoxTime models to construct a multi-task learning model,TransMTL.We also added two auxiliary tasks,mortality prediction and duration prediction,to make the full use of limited data for survival analysis.This study was conducted on PyTorch and R language,and four proposed models,include TransCR,TransCox,TransCoxTime and TransMTL,were evaluated on three open available medical datasets.The results show that,compared to conventional neural network models,the four Transformer-based survival analysis models proposed in this paper perform well in terms of discrimination,calibration and interpretab ility.TransCR model superior to current competing risk models that ignore selection bias and achieves good results on small probability events.TransCox and TransCoxTime are extremely competitive in single event models,increasing the AUC values of METABRIC dataset by 1.9%and 2.3%respectively compared to the best baseline model.TransMTL achieved optimal performance in both single-event and competing risk analyses,with the two auxiliary tasks improving prediction accuracy in all main tasks.The mortality prediction task contributes more to model performance improvement.The visualization of Transformer attention score demonstrates that the hormone therapy,radiotherapy and age are very significant prognostic factors,with cretain correlations with all other features.