Machine learning-powered audio-omics processing method as an auxiliary diagnostic approach for advanced
Ting You1, Fangna Huan2, Zhenzhen Luo3
1The First Affiliated Hospital, Department of Emergency, Hengyang Medical School, University of South China, Hengyang, 421001, China.
Purpose:
Nasopharyngeal carcinoma (NPC) is located in the nasopharyngeal mucosa and is a malignant tumour of the head and neck, and approximately 70% of patients have intermediate to advanced disease at the time of initial diagnosis. Epstein-Barr virus (EBV) is closely correlated with etiology and pathogenesis of NPC, serological detection of EBV antibodies is a common screening method for NPC. However, only approximately 60% of NPC cases are associated with EBV infection. Herein, this work aimed to develop and internally evaluate a machine learning-based acoustic signal processing model as a preliminary non-invasive auxiliary diagnostic approach for advanced NPC.
Materials And Methods:
First, we collected the audio files from 359 advanced NPC patients and 304 healthy controls in our hospital from 2022 to 2025. The machine learning-powered Nasopharyngeal Carcinoma Screening (ML-NPCS) system for screening NPC. And the ML-NPCS system is composed of three steps: speech acquisition, acoustic features extraction, and classification decision-making.
Results:
In the independent test set, ML-NPCS achieved an accuracy of 84.2% (95% CI, 77.1%-89.4%), a sensitivity of 88.9% (95% CI, 79.6%-94.3%), and a specificity of 78.7% (95% CI, 66.9%-87.1%); the independent test set comprised 133 participants (72 patients with advanced NPC and 61 healthy controls).
Conclusion:
The ML-NPCS model demonstrated preliminary potential for distinguishing advanced NPC patients from healthy controls using voice-derived acoustic features. Further prospective evaluation in early-stage disease, symptomatic controls, and external cohorts is required before population-level screening use can be considered.
