상세 보기
음성 기반 ALS 중증도 예측을 위한 하이브리드 특징 기반 딥러닝 기법
- 송채린;
- 류정찬;
- 이채린;
- 김민수;
- 곽일엽
WEB OF SCIENCE
0초록
Early and accurate prediction of disease progression in neurodegenerative disorders such as Amyotrophic Lateral Sclerosis (ALS) is essential for timely intervention and effective treatment planning. In this study, we propose an AI-based framework to predict the ALS Functional Rating Scale-Revised (ALSFRS-R) score at the final follow-up using patients’ initial speech recordings and clinical metadata. Experiments were conducted on the dataset provided by the Speech Analysis for Neurodegenerative Diseases (SAND) Challenge at ICASSP 2026. To capture subtle articulatory impairments caused by ALS, we adopt a hybrid feature extraction strategy that combines deep semantic embeddings from a pre-trained Whisper model with handcrafted acoustic features reflecting physiological characteristics of speech. This design enables complementary modeling of high-level linguistic information and low-level speech distortions. To mitigate data imbalance and sparsity, we employ Sliding Window-based strategy and Synthetic Minority Over-sampling Technique (SMOTE), improving training stability. As a result, the proposed hybrid pipeline achieved performance improvement over baseline and secured 4th place in the competition, demonstrating its ability to predict disease progression stages in an imbalanced data environment. By validating the efficacy of combining deep learning and acoustic features with sampling strategies, this study contributes to the development of non-invasive, objective, biomarker-based patient monitoring systems.
키워드
- 제목
- 음성 기반 ALS 중증도 예측을 위한 하이브리드 특징 기반 딥러닝 기법
- 제목 (타언어)
- A deep learning approach using hybrid features for ALS severity prediction from speech
- 저자
- 송채린; 류정찬; 이채린; 김민수; 곽일엽
- 발행일
- 2026-08
- 유형
- Article
- 저널명
- 응용통계연구
- 권
- 39
- 호
- 4
- 페이지
- 401 ~ 417
- 언어
- KOR
- 출판사
- 한국통계학회
- 발행국가
- 대한민국
- 분량
- 17 페이지
- ISSN
- E 2383-5818
P 1225-066X