A transformer model for cause-specific hazard prediction.

Area of Science:

• Biostatistics
• Machine Learning
• Computational Biology

Background:

• Accurate modeling of discrete-time cause-specific hazards with competing events and non-proportional hazards is crucial but challenging in longitudinal studies.
• Existing models often rely on restrictive proportional hazards assumptions or inadequately handle sequential data, limiting their applicability in complex clinical scenarios.
• The Transformer architecture offers a powerful, assumption-light approach for analyzing complex relationships within sequential data.

Purpose of the Study:

• To propose and evaluate a Transformer-based architecture for predicting cause-specific hazards in discrete-time competing risks scenarios.
• To assess the model's performance against established methods like CoxPH, PyDTS, and DeepHit, particularly in settings with non-proportional hazards.
• To demonstrate the model's capability in handling complex covariate-to-outcome dynamics and its interpretability.

Main Methods:

• Development of a Transformer architecture tailored for discrete-time cause-specific hazard prediction in competing risks.
• Validation using synthetic datasets (2,000-50,000 patients) and the English Longitudinal Study of Ageing (ELSA) cohort.
• Performance comparison using metrics such as the Integrated Brier Score and time-dependent concordance index; model explainability via integrated gradients.

Main Results:

• The Transformer model significantly outperformed CoxPH, PyDTS, and DeepHit in predicting cause-specific hazards, especially when the proportional hazards assumption was not met.
• The model demonstrated superior ability to anticipate hazard evolution at later time steps, even with sparse event data.
• Exceptional performance was observed in predicting dementia and psychiatric conditions within the ELSA cohort, surpassing existing models.

Conclusions:

• The proposed Transformer model achieves state-of-the-art performance in cause-specific hazard prediction without imposing parametric assumptions on hazard rates.
• It is particularly effective in longitudinal cohort studies with complex, non-proportional hazard dynamics, outperforming traditional and deep learning models.
• The model's interpretability through integrated gradients aids in understanding variable importance, making it a valuable tool for time-to-event prediction in clinical research.