Explainable transformer-based deep survival analysis in childhood acute lymphoblastic leukemia

Area of Science:

• Computational biology
• Machine learning in healthcare
• Oncology research

Background:

• Acute lymphoblastic leukemia (ALL) is a life-threatening childhood cancer with increasing incidence.
• Predictive survival models are crucial for timely and effective treatment of ALL.
• Existing models struggle with complex data; Transformers offer advanced feature dependency analysis.

Purpose of the Study:

• To develop an explainable Transformer-based deep survival model for predicting patient-specific survival probabilities in ALL.
• To enhance interpretability of survival predictions using Shapley Additive Explanations (SHAP).

Main Methods:

• Proposed an explainable Transformer-based deep survival model integrating feedforward networks.
• Trained the model to minimize the difference between predicted and actual survival outcomes.
• Utilized SHAP for global and local interpretation of clinical attribute contributions.

Main Results:

• The model achieved a high concordance index (C-index) of 0.945, indicating strong predictive accuracy.
• SHAP analysis identified prognosis status, diagnosis year, and histology as key survival factors.
• Outperformed state-of-the-art deep survival models when incorporating these key variables.

Conclusions:

• The explainable Transformer model accurately predicts patient-specific survival in ALL.
• SHAP-derived insights enhance model interpretability for clinical decision support.
• Potential to improve prognosis and treatment strategies for ALL patients.