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

Abstract

Background

Acute lymphoblastic leukemia (ALL) is the most common type of leukemia among children and adolescents and can be life-threatening. The incidence of new cases has been increasing in recent years. Developing a predictive model to forecast the risk of death can help improve survival rates by enabling clinicians to provide timely and effective treatments. Traditional statistical survival models are limited by predefined assumptions, while current deep survival models, despite their flexibility, struggle with capturing complex and dynamic feature dependencies. Transformers provide a promising solution by using self-attention and multi-head attention mechanisms to overcome these challenges. Moreover, building on recent work in interpretable medical AI, the combination of Transformers and explainable methods can quantify the contributions of each feature to the survival probability prediction.

Methods

This paper proposes an explainable Transformer-based deep survival model to predict patient-specific survival probabilities for ALL. The model combines feedforward networks with Transformer architecture and is trained to minimize a loss function that measures the difference between predicted and actual survival outcomes. In addition, we use Shapley Additive Explanations (SHAP) to interpret the contributions of clinical attributes to the predictions, providing insights from both global and local perspectives.

Results

The proposed model demonstrates robustness by consistently providing higher average survival probabilities for censored patients compared to deceased patients. It achieves an average concordance index (C-index) of 0.945, demonstrating strong predictive accuracy. Through SHAP analysis, we identify three key factors affecting survival outcomes, namely prognosis status, diagnosis year, and histology. Experiment results reveal that our model outperforms state-of-the-art deep survival models in terms of C-index when these important variables are included.

Conclusion

The proposed explainable Transformer-based deep survival model shows strong potential for providing accurate patient-specific survival predictions for ALL. Moreover, the insights gained from SHAP improve the model's interpretability for clinicians, helping them make better-informed decisions regarding prognosis and treatment.