Acoustic biomarkers and AI: Transforming NSCLC detection and personalized care.

Abstract

1568 Background: Implementing a mass screening program for lung cancer using low-dose chest CT presents significant challenges, including financial constraints and concerns about radiation exposure. Nonetheless, recent evidence reveals that lung cancer is not limited to smokers, as it also affects non-smoker populations who are currently excluded from existing screening programs. As part of the I3LUNG study (NCT05537922), we investigated the use of AI-based forced cough analysis as a non-invasive approach to distinguish NSCLC patients undergoing immunotherapy (IO) from healthy individuals. Additionally, we examined whether cough features could differentiate patients based on their baseline clinical features. Methods: Machine Learning-based preprocessing isolated meaningful cough events and extracted 39 acoustic features from the time and frequency domains. To reduce redundancy and improve model performance, highly correlated features ( > 85%) were eliminated. Support Vector Machines (SVM) and Deep Learning (DL) models were then employed to distinguish NSCLC patients from healthy controls. Additional statistical analyses of acoustic features were conducted on cough recordings from patients to evaluate differences based on smoking status (current, former, or never smokers) using the Kruskal-Wallis test with Benjamini-Hochberg post-hoc correction. Similarly, differences based on the presence or absence of lung metastases were assessed using the Mann-Whitney test. Results: A total of 200 individuals were enrolled in the study, including 91 stage IIIB-IV NSCLC patients undergoing IO and 109 healthy controls. Cough recordings were analyzed, with the SVM model achieving an accuracy of 82% and a specificity of 92% on the test set. The DL model demonstrated superior performance, with an accuracy of 95% and a specificity of 100%. Significant differences were observed in the peak-to-root-mean-square value ratio and cough duration among smokers (current, former, or never), with P-values of 0.026 and 0.042, respectively. Furthermore, spectral features - including centroid, rolloff, spread, kurtosis, bandwidth, and flatness - differed significantly between patients with and without lung metastases (P < 0.01). Conclusions: These findings highlight the potential of cough as a valuable digital biomarker for NSCLC diagnosis. The tool's high sensitivity facilitates the effective identification of individuals at risk for lung cancer, while its exceptional specificity makes it a promising initial screening method, efficiently triaging positive cases for follow-up chest CT scans. Future studies should validate these results on larger cohorts. Moreover, the correlation of specific cough features with smoking status and the presence of lung metastases suggests that this tool could extend beyond screening to monitoring disease progression over time.