Explainable Transformer-Based Neural Network forthe Prediction of Survival Outcomes in Non-SmallCell Lung Cancer (NSCLC)

In this paper, we introduce the "Clinical Transformer" - a recasting of the widely used transformer architecture as a method for precision medicine to model relations between molecular and clinical measurements, and the survival of cancer patients. Although the emergence of immunotherapy offers a new hope for cancer patients with dramatic and durable responses having been reported, only a subset of patients demonstrate benefit. Such treatments do not directly target the tumor but recruit the patients immune system to fight the disease. Therefore, the response to therapy is more complicated to understand as it is affected by the patients physical condition, immune system fitness and the tumor. As in text, where the semantics of a word is dependent on the context of the sentence it belongs to, in immuno-therapy a biomarker may have limited meaning if measured independent of other clinical or molecular features. Hence, we hypothesize that the transformer-inspired model may potentially enable effective modelling of the semantics of different biomarkers with respect to patients survival time. Herein, we demonstrate that this approach can offer an attractive alternative to the survival models utilized in current practices as follows: (1) We formulate an embedding strategy applied to molecular and clinical data obtained from the patients. (2) We propose a customized objective function to predict patient survival. (3) We show the applicability of our proposed method to bioinformatics and precision medicine. Applying the clinical transformer to several immuno-oncology clinical studies, we demonstrate how the clinical transformer outperforms other linear and non-linear methods used in current practice for survival prediction. We also show that when initializing the weights of a domain-specific transformer by the weights of a cross-domain transformer, we further improve the predictions. Lastly, we show how the attention mechanism successfully captures some of the known biology behind these therapies.