Artificial Intelligence for Non-Invasive Diagnosis of Voice Disorders
Duration: 01/2024 - 12/2025
Scientific Coordinator: Gabriele Ciravegna
Project Type: Non-commercial collaborative agreement
PoliTO Role: Project Coordinator
Project Objectives
The project aims to develop an innovative non-invasive diagnostic system based on artificial intelligence (AI), designed to improve the accuracy and accessibility of diagnostic processes. The main objective is the early detection and classification of voice disorders and other medical conditions using data collected from everyday devices (e.g., smartphones, smartwatches, and computers). This system aims to provide personalized and non-invasive clinical support, enhancing patients’ quality of life and reducing the need for traditionally invasive diagnostic procedures such as laryngoscopy.
Methodologies Used
The project employs advanced machine learning algorithms, with a particular focus on end-to-end trained transformer models for the analysis of raw voice data. To overcome limitations related to data scarcity and imbalance, techniques such as data augmentation and synthetic data generation based on real voices of both healthy and pathological patients are used. Furthermore, the approach utilizes a multi-modal analysis pipeline that combines predictions derived from different types of vocal samples (e.g., sentence reading, sustained vowels) through an ensemble model based on a mixture of experts (MoE) and other fusion techniques, leveraging the intrinsic capability of attention mechanisms to handle diverse data types. The integration of multiple data sources enables a holistic understanding of patients’ health status, facilitating accurate and personalized diagnoses.
Expected Results
The proposed system promises to significantly enhance healthcare services in three key areas:
- Clinical: Early and personalized detection will allow timely interventions, improving treatment outcomes and reducing the need for invasive procedures.
- Economic: The efficiency of AI-based diagnostic methods could lead to overall healthcare cost reductions while enabling large-scale screening.
- Technological: The developed models can be reused or fine-tuned in other diagnostic contexts characterized by limited data availability, thus broadening the project’s impact.
In summary, the project represents a step forward toward a more accessible, effective, and patient-centred healthcare system with tangible benefits on multiple levels.
Involved Structure
- DAUIN, Department of Control and Computer Engineering
Partners
- Head Neck Cancer Unit
- San Giovanni Bosco Hospital
- San Feliciano Hospital
Location
- Politecnico di Torino main campus, Corso Duca degli Abruzzi 24
Project Team
- Dr. Gabriele Ciravegna (DAUIN), Post-doctoral Researcher and Scientific Project Supervisor
- Student Alkis Koudounas (DAUIN), PhD Candidate and Model Implementation Lead
- Tania Cerquitelli (DAUIN), Project Coordinator
- Marco Fantini (DAUIN), MD, Data Collection and Results Analysis Lead
- Giovanni Succo (DAUIN), MD, Coordinator
Sustainable Development Goals (SDGs)
3. Good health and well-being
10. Reduced inequalities