Thesis Reducción de dimensión y clustering no supervisado para el análisis de señales de voz ambulatorias captadas por acelerómetro
Loading...
Date
2025
Authors
Journal Title
Journal ISSN
Volume Title
Program
Ingeniería Civil Electrónica
Departament
Campus
Campus Casa Central Valparaíso
Abstract
Los trastornos de la voz asociados a hiperfunción vocal representan un desafío diagnóstico y terapéutico debido a la ausencia de marcadores objetivos estables y a la variabilidad del comportamiento vocal diario. En este contexto, el monitoreo ambulatorio con acelerometría laríngea permite registrar señales de fonación durante la vida cotidiana, generando grandes volúmenes de datos que requieren nuevas estrategias de análisis. El presente trabajo propone un pipeline no supervisado para caracterizar el comportamiento vocal ambulatorio en mujeres con patología fonotraumática y sus controles, combinando reducción de dimensión mediante UMAP y clusterización con HDBSCAN. Se utilizaron registros de 13 pares de sujetos (pacientes y controles) obtenidos en el Massachusetts General Hospital, segmentados en ventanas de 50 ms y subsecuencias de 1 s. La calidad del embedding y de los agrupamientos se evaluó con métricas internas (DBCV, Silhouette, Davies-Bouldin) y de preservación espacial-temporal. Los resultados revelan una estructura latente estable que distingue tres patrones principales: (i) un clúster basal asociado a fonación modal cotidiana, (ii) un clúster vinculado a voz cantada y (iii) un clúster minoritario con flujos glóticos anómalamente altos, posiblemente atribuibles a artefactos de modelado. La comparación entre condiciones mostró redistribuciones moderadas tras la terapia vocal, así como la emergencia de patrones adicionales al aumentar la escala temporal de análisis. Asimismo, los estudios de caso individuales evidenciaron la capacidad del método para detectar eventos vocales transitorios de interés clínico. Estos hallazgos demuestran que la combinación UMAP + HDBSCAN constituye una herramienta robusta y escalable para explorar millones de ventanas de datos de voz ambulatoria, aportando bases objetivas para la caracterización de patrones vocales, el monitoreo terapéutico y la detección temprana de comportamientos hiperfuncionales en la vida diaria.
Voice disorders associated with vocal hyperfunction represent a diagnostic and therapeutic challenge due to the lack of stable objective markers and the high variability of daily vocal behavior. In this context, ambulatory monitoring using a necksurface accelerometer enables long-term phonation recording in everyday life, generating large volumes of data that demand advanced analysis strategies. This work proposes an unsupervised pipeline to characterize ambulatory vocal behavior in women with phonotraumatic pathology and matched controls, combining dimensionality reduction through UMAP with clustering via HDBSCAN. Data from 13 subject pairs (patients and controls) collected at Massachusetts General Hospital were analyzed, segmented into 50-ms windows and 1-s subsequences. The quality of embeddings and clustering was evaluated using internal metrics (DBCV, Silhouette, Davies-Bouldin) as well as spatial-temporal preservation indices. Results revealed a stable latent structure distinguishing three main patterns: (i) a basal cluster associated with everyday modal phonation, (ii) a cluster strongly linked to singing voice, and (iii) a minority cluster with abnormally high glottal airflow values, likely attributable to modeling artifacts. Comparisons across clinical conditions showed moderate redistributions following voice therapy, while longer temporal windows allowed the emergence of additional phonatory patterns. Case studies further demonstrated the method’s ability to detect transient vocal events of clinical interest. These findings demonstrate that the combination of UMAP and HDBSCAN provides a robust and scalable framework for exploring millions of windows of ambulatory voice data, offering objective foundations for the characterization of vocal patterns, therapy monitoring, and early detection of hyperfunctional behaviors in daily life.
Voice disorders associated with vocal hyperfunction represent a diagnostic and therapeutic challenge due to the lack of stable objective markers and the high variability of daily vocal behavior. In this context, ambulatory monitoring using a necksurface accelerometer enables long-term phonation recording in everyday life, generating large volumes of data that demand advanced analysis strategies. This work proposes an unsupervised pipeline to characterize ambulatory vocal behavior in women with phonotraumatic pathology and matched controls, combining dimensionality reduction through UMAP with clustering via HDBSCAN. Data from 13 subject pairs (patients and controls) collected at Massachusetts General Hospital were analyzed, segmented into 50-ms windows and 1-s subsequences. The quality of embeddings and clustering was evaluated using internal metrics (DBCV, Silhouette, Davies-Bouldin) as well as spatial-temporal preservation indices. Results revealed a stable latent structure distinguishing three main patterns: (i) a basal cluster associated with everyday modal phonation, (ii) a cluster strongly linked to singing voice, and (iii) a minority cluster with abnormally high glottal airflow values, likely attributable to modeling artifacts. Comparisons across clinical conditions showed moderate redistributions following voice therapy, while longer temporal windows allowed the emergence of additional phonatory patterns. Case studies further demonstrated the method’s ability to detect transient vocal events of clinical interest. These findings demonstrate that the combination of UMAP and HDBSCAN provides a robust and scalable framework for exploring millions of windows of ambulatory voice data, offering objective foundations for the characterization of vocal patterns, therapy monitoring, and early detection of hyperfunctional behaviors in daily life.
Description
Keywords
Trastornos de la voz, Hiperfunción vocal, Reducción de dimensión, Clustering, UMAP, HDBSCAN, Acelerometría
