Early stage lung cancer detection from speech sounds in natural environments

Abstract

In the diagnosis of early-stage lung cancer, conventional methods often rely on periodic imaging techniques using medical devices. However, recent studies suggest that speech sounds could offer valuable insights into the diagnosis of disease. This study investigates the different characteristics of speech sounds recorded in natural environments between individuals diagnosed with lung cancer and those who are healthy. Using signal processing techniques and a self-supervised contrastive learning approach, we investigate the classification of these speech sounds for the diagnosis of early-stage lung cancer. Our results show that it is possible to utilize naturally recorded speech sounds. Using the Graph Attention Transformer Fine-Tuning Contrastive Learning (GAT-ftCL) model, which leverages graph neural networks to capture complex relationships in data and fine-tunes the learning process through contrastive learning, we achieve an accuracy of 90.90% in distinguishing individuals diagnosed with lung cancer from healthy individuals. Furthermore, the model achieves a remarkable accuracy of 92.85% in specifically identifying individuals with early stage (stage 1) lung cancer within the stage 1 group and healthy individuals. These results underline the diagnostic potential of natural speech sounds, especially in the detection of early-stage lung cancer.