Incremental feature fusion based time series forecasting with cumulative risk constraint for longitudinal overall survival prediction.

Abstract

Background: Overall survival (OS) prediction methods usually adopt pre-operative data which lack important prognosis-related information, such as post-operative lesion status and evolution during treatment, leading to unsatisfactory performance. Incorporating longitudinal data into OS prediction, however, introduces two main challenges: (1) variable time span; and (2) implicit spatiotemporal information.

Purpose: This study aims to break the limitation of pre-operative data based OS prediction by addressing the aforementioned two main challenges and leveraging longitudinal data to achieve accurate OS prediction.

Methods: We propose a novel longitudinal data based OS prediction method. Specifically, a new incremental feature fusion (IFF) based time series forecasting module is presented to derive accumulated features up to each time point and fill missing time points in longitudinal data. It addresses the challenge of variable time span with high computational efficiency compared to the widely applied decoder-only transformer with causal-attention (DoT-CA). Based on the accumulated features in the IFF module, corresponding survival risks up to each time point are predicted under a cumulative survival risk (CSR) constraint, where the survival risks are encouraged to be monotonically increased over time, effectively exploring the spatiotemporal information.

Results: In the experiment, both in-house and public multimodal MR datasets (BraTS2020) containing 1678 patients of diffuse glioma are used to evaluate our method, and the experimental results show that our method outperforms all state-of-the-art (SOTA) methods with statistical significance. Further ablation study shows that both proposed IFF module and CSR constraint are effective in longitudinal OS prediction. Moreover, the proposed IFF module is more efficient than DoT-CA, enabling scalable offline longitudinal analysis on large patient cohorts.

Conclusion: Longitudinal data contain important spatiotemporal information related to prognosis, based on which more accurate OS prediction can be achieved comparing with existing pre-operative data based methods. For longitudinal data with variable time span, the evolution pattern of lesions can be effectively learned and used to fill up the missing time points. Codes of our method are available at https://github.com/BH-MICom/OSTimes.