Exploring Deep Learning Techniques in the Prediction of Cancer Relapse Using an Open Brazilian Tabular Database

The early prediction of the risk of cancer relapse can bring various benefits for healthcare, and open databases can help to address this challenge. Although the literature presents several studies to predict cancer relapse based on images and genomics data, these are not always available. To deal with this issue, we investigated the use of deep learning techniques with tabular data of an open Brazilian database. This database aggregates hospital tabular records from over 70 hospitals located in Brazil. We analyzed models based on Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN) and FT-Transformers. CNN results for the 202 types of cancer show an average Area Under the ROC Curve (AUC) of 0.62, while the best AUC of 0.85 is for cancer type ICD-O C718 (Malignant neoplasm of overlapping sites of brain). The best F1 score and average accuracy results were obtained using FT-Transformer-based model, reaching 0.92 and 0.86, respectively, while the best individual F1 and accuracy were 0.99 and 0.98 for the cancer type ICD-O C258. In addition to the model performance, we also bring explainability to the models by identifying the most relevant features through the calculation of SHAP values. The corresponding feature analysis indicates that the most relevant explanatory variables are TNM classification codes and the time interval to begin treatment after the diagnosis. This article also discusses the limitations, challenges, and future works.