Voice-Based Screening of Lung Cancer and COPD Using Combined Spectral, Cepstral, and Learned-Embedding Features
12th International Conference on Electrical and Electronics Engineering, ICEEE 2025, İstanbul, Türkiye, 24 - 26 Eylül 2025, ss.521-525, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/iceee67194.2025.11261927
- Basıldığı Şehir: İstanbul
- Basıldığı Ülke: Türkiye
- Sayfa Sayıları: ss.521-525
- Anahtar Kelimeler: Acoustic Biomarkers, Mel-Frequency Cepstral Coefficients (MFCC), Non-Invasive Respiratory Disease Screening, SHAP-Based Feature Importance
- Ankara Üniversitesi Adresli: Evet
Özet
Early, accessible detection of lung cancer and chronic obstructive pulmonary disease (COPD) remains challenging due to invasive and costly standard methods. Here, we explore a non-invasive, voice-based approach to distinguish healthy controls from lung cancer (Stages I-IV) and COPD (Stages I-III) patients. We collected speech recordings from 128 Turkish-speaking individuals (42 healthy, 50 lung cancer, 36 COPD) reading standardized paragraphs in a quiet setting. Recordings were down-mixed to mono, high-pass filtered, and segmented into 100 ms frames. From each frame, we extracted time-domain statistics (RMS, skewness, zero-crossing rate), frequency-domain descriptors (spectral centroid, bandwidth, flux), MFCCs 1-13, spectrogram and chroma features, and learned embeddings via a 64-dimensional autoencoder and PCA-reduced components. Dimensionality reduction and K-means clustering revealed that learned embeddings and select spectral features achieved clear grouping of healthy, COPD, and lung cancer samples. An XGBoost classifier leveraging all features, coupled with SHAP analysis, identified higher-order MFCCs (MFCC5-MFCC7), spectral bandwidth, skewness, and zero-crossing rate as most discriminative. Physiological interpretation suggests that COPD-induced airflow obstruction leads to breathiness and turbulence - reflected in elevated zero-crossing rates and amplitude irregularities - whereas lung cancer yields formant distortions and heavier spectral tails, captured by higher-order MFCCs and kurtosis. These findings underscore the promise of acoustic biomarkers for scalable, cost-effective respiratory disease screening.