 |
International Society of Science and Applied Technologies |
| A Two-Stage Attention-Based Deep Learning Model for Predicting Mortality in Hemodialysis Patients | ||||
| Author |
Yen-Chun Huang
|
|||
| Co-Author(s) |
Ming-Hsien Tsai; Mingchih Chen
|
|||
| Abstract | In End-stage renal disease (ESRD) patients depend on hemodialysis (HD) to sustain life, and accurately predicting their survival duration is crucial for personalized medical decision-making. However, traditional machine learning and deep learning models face two major challenges when handling real-world survival data: right censoring bias and scale explosion in predictive fusion. These issues often lead to a pronounced regression-to-the-mean effect, thereby suppressing the ability to predict long-term survival potential. This study proposes a novel two-stage regression framework. In the first stage, heterogeneous machine learning expert models are constructed, incorporating a time anchoring design and an unbiased recursive feature elimination with cross-validation (RFECV) technique. Through stratified sampling based on implicit targets, an unbiased distribution of the feature space is ensured. In the second stage, a Decoupled Cross-Attention Fusion Network is developed, which disentangles the raw predictions of expert models from the attention weight learning pathway. This network is paired with a newly derived Censored Mean Squared Error (Censored MSE) loss function, effectively addressing penalty misestimation in right-censored data. Furthermore, the study transforms the genetic algorithm (GA) from a static weight searcher into a global hyperparameter optimizer, enabling multi-generational optimization of the deep fusion network. Finally, by analyzing feature importance, the study elucidates the underlying pathological mechanisms driving prediction decisions, providing highly interpretable and practical AI-assisted indicators for clinical survival risk stratification.
|
|||
| Keywords | Hemodialysis (HD); Mixture of Experts (MoE); Mortality; Genetic algorithm (GA) | |||
| Article #: RQD2026-41 | ||||
Proceedings of 31st ISSAT International Conference on Reliability & Quality in Design |