🤖 machine learning

A Concept-based approach to Voice Disorder Detection

This paper proposes a concept-based Explainable AI approach using Concept Bottleneck Models and Concept Embedding Models to detect voice disorders with performance comparable to traditional deep learning methods while providing transparent, interpretable decision-making for clinical trust.

Original authors: Davide Ghia, Gabriele Ciravegna, Alkis Koudounas, Marco Fantini, Erika Crosetti, Giovanni Succo, Tania Cerquitelli

Published 2026-06-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Davide Ghia, Gabriele Ciravegna, Alkis Koudounas, Marco Fantini, Erika Crosetti, Giovanni Succo, Tania Cerquitelli

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a very smart, super-fast robot doctor. This robot can listen to a person's voice and tell you instantly if they have a voice disorder (like a hoarse or strained voice) or if their voice is healthy.

The problem with this robot, as described in the paper, is that it's a "black box." It gives you a diagnosis, but it can't explain why. It's like a magic 8-ball that just says "Yes" or "No" without showing you the logic. In healthcare, doctors and patients need to know why a decision was made to trust it.

This paper introduces a new way to build this robot so it acts more like a human expert who can explain their reasoning. Here is how they did it, using simple analogies:

1. The "Translator" Robot (The LLM)

First, the researchers needed to teach the robot what to look for. They had thousands of patient files written by real doctors (called "anamnesis"). These files were messy, like handwritten notes in a diary, not organized data.

To fix this, they used a "Translator Robot" (a Large Language Model, or LLM). Think of this translator as a super-organized secretary. You hand it a messy doctor's note, and it extracts specific, clear facts into a checklist.

The Checklist: Instead of just saying "sick," the translator pulls out specific traits like: Is the voice rough? Is there strain? Is the patient a smoker? Is the voice use professional?
The Result: They turned 14 different medical ideas into a clean list of 9 specific "concepts" the robot could understand (like "Roughness: Yes" or "Strain: No").

2. The Two New Robot Designs

The researchers built two new types of robots that use this checklist. They call these Concept-Based Models.

The "Bottleneck" Robot (CBM):
Imagine a factory assembly line.
1. Input: The robot listens to the voice.
2. The Bottleneck: Before it can make a final decision, it must stop and fill out a form. It has to answer: "Is there roughness? Yes/No. Is there strain? Yes/No."
3. Output: Only after filling out this form does it make the final diagnosis.
- Why this is cool: You can see the form! If the robot says "Pathological," you can look at the form and see, "Ah, it said 'Yes' to roughness and 'Yes' to strain." You know exactly why it made that call.
The "Embedding" Robot (CEM):
This is a slightly more advanced version of the factory. Instead of just checking boxes, it creates a unique "fingerprint" for each concept (like a digital ID card for "Roughness"). It mixes these fingerprints together to make the final decision. It's a bit more complex, but it still keeps the logic transparent.

3. The Big Test: Does it work?

The researchers tested these new robots against the old "Black Box" robot (a standard Deep Neural Network).

The Old Robot: Very accurate, but you couldn't ask it "Why?"
The New Robots: They were almost just as accurate as the old one (getting about 87-88% right compared to the old one's 91%).
The Win: The new robots could explain themselves. If the robot flagged a voice as disordered, it could point to the specific "concepts" it found, like "The voice is breathy and has a glottic gap."

4. The "Ideal" Scenario

The researchers also imagined a "Perfect Robot" (called the Ideal CBM). This robot didn't have to guess the concepts; it was given the perfect checklist by a human. Even with this perfect help, the new robots performed very close to the top. This proved that the concepts they chose (like roughness, strain, breathiness) were actually the right things to look for to diagnose voice problems.

Summary

In short, this paper says: "We built a voice-diagnosing AI that doesn't just guess; it thinks in terms of human-understandable ideas like 'roughness' or 'strain'."

Old Way: "The voice is sick." (No explanation).
New Way: "The voice is sick because it is rough, strained, and breathy." (Clear explanation).

The paper concludes that this method is a huge step forward because it makes AI trustworthy for doctors, allowing them to see the "why" behind the diagnosis without sacrificing too much accuracy. They also suggest that in the future, this system could write a full report for the doctor, but for now, they have just proven it works as a transparent diagnostic tool.

Technical Summary: A Concept-based Approach to Voice Disorder Detection

Problem Statement

Voice disorders affect a significant portion of the population, yet current diagnostic methods often rely on invasive clinical procedures. While recent Deep Neural Networks (DNNs), particularly those based on Transformer architectures, have demonstrated promising performance in classifying pathological versus euphonic voices using non-invasive audio analysis, they suffer from a critical limitation: opacity. These models operate as "black boxes," providing outputs without explaining the decision-making process. In the sensitive context of healthcare, this lack of interpretability raises ethical and legal concerns, hindering the trustworthiness and clinical adoption of AI-driven diagnostics. The paper addresses the need for a diagnostic tool that maintains high predictive performance while offering a transparent, interpretable decision framework.

Methodology

The authors propose a concept-based approach using Explainable AI (XAI), specifically focusing on Concept Bottleneck Models (CBM) and Concept Embedding Models (CEM). The methodology involves a two-stage process:

1. Concept Annotation via Large Language Models (LLMs)

Since the target dataset (Italian Pathological Voice - IPV) contains unstructured textual anamnesis (medical history) rather than pre-labeled concepts, the authors developed a pipeline to extract 14 candidate concepts from these texts.

Extraction: They utilized an LLM (specifically Gemini-pro, selected after comparative testing with Mistral 7B) to parse the anamnesis and output a structured JSON object containing concept values.
Concept Definition: The initial set included 14 concepts (e.g., smoking, professional voice use, dysphonia severity, irregular mucous wave, roughness, breathiness).
Refinement: Through consultation with medical experts, the set was refined. Concepts requiring invasive physical exams (e.g., specific mucous colors) were excluded. Others were merged (e.g., "dysphonia light-moderate" merged with "moderate"). The final set comprises 9 predictable concepts (derived from the audio/anamnesis via the model) and 5 patient-provided concepts (e.g., gender, smoking status). The predictable concepts largely align with the GRBAS scale (Grade, Roughness, Asthenia, Strain).
Encoding: Multi-class concepts (e.g., dysphonia severity) were one-hot encoded, resulting in a total of 20 concept dimensions before final selection.

2. Model Architecture

The proposed framework utilizes HuBERT, a pre-trained audio Transformer, as the feature extractor to generate audio embeddings. Two architectures are implemented:

Concept Bottleneck Model (CBM): The audio embeddings are passed through a classification head to predict the binary values of the 9 concepts (the "bottleneck"). These predicted concepts are concatenated with the 5 patient-provided concepts. A second classifier then maps this combined concept vector to the final binary diagnosis (pathological vs. euphonic). Crucially, the final classifier has no access to the raw audio signal, relying solely on the concepts.
Concept Embedding Model (CEM): Similar to CBM, but instead of predicting binary logits, the bottleneck layer learns two distinct embeddings for each concept (active and inactive states). The final prediction is a weighted combination of these embeddings based on the predicted probability of the concept being active.

3. Training Strategy

The models are trained using a joint loss function that minimizes both the concept loss (accuracy of predicting the intermediate concepts) and the task loss (accuracy of the final diagnosis). A hyperparameter $\lambda$ balances these losses. The authors employed a warm-up strategy, freezing the task loss for the first two epochs to prioritize learning robust concept representations before optimizing for the final classification.

Key Contributions

Application of Concept-based XAI to Voice Disorders: This is the first work to apply CBM and CEM architectures specifically to the detection of voice pathologies, moving beyond standard end-to-end black-box approaches.
Automated Concept Extraction Pipeline: The paper demonstrates a novel workflow using LLMs to extract high-level medical concepts from unstructured clinical text (anamnesis) to serve as ground truth for training concept-based models.
Interpretability without Performance Loss: The study validates that concept-based models can achieve performance comparable to state-of-the-art end-to-end Transformers while providing explicit justifications for their predictions in terms of clinically relevant features (e.g., "dysphonia," "roughness").

Results

Experiments were conducted on the Italian Pathological Voice (IPV) dataset, comprising 513 patients (170 euphonic, 340 dysphonic), with 385 cases containing written anamnesis. The study utilized 10-fold cross-validation.

Concept Annotation: The Gemini-pro model achieved a concept accuracy of 98.65% and a macro F1 score of 98.10% on a manually annotated validation set, outperforming Mistral 7B.
Voice Disorder Detection:
- Baseline: A standard end-to-end HuBERT model with a 2-layer classification head achieved a Task Accuracy of 91.33% and an F1 score of 0.9085.
- Concept-Based Models:
  - CBM: Achieved a Task Accuracy of 87.76% and F1 of 0.8566, with a concept accuracy of 84.43%.
  - CEM: Achieved a Task Accuracy of 87.30% and F1 of 0.8599, with a concept accuracy of 84.50%.
- Ideal CBM: When provided with ground-truth concept labels (simulating perfect concept extraction), the CBM achieved 90.52% accuracy, nearly matching the end-to-end baseline.

The results indicate that the drop in performance for the actual concept-based models is primarily due to the imperfect nature of the LLM-generated concept annotations (approx. 84.5% accuracy) rather than a fundamental limitation of the architecture. The models successfully learned to use concepts like "dysphonia," "breathiness," and "roughness" to justify the final diagnosis.

Significance and Claims

The paper claims that concept-based models represent a significant step forward in making AI in healthcare transparent and trustworthy. By explicitly linking the final diagnosis to intermediate, human-understandable concepts (such as the GRBAS scale parameters), these models allow clinicians to verify the reasoning behind a prediction. The authors assert that while the end-to-end approach remains slightly more accurate, the concept-based approach offers a "competitive" performance level with the distinct advantage of interpretability, enabling human experts to intervene or derive counterfactual predictions. The authors remain modest, acknowledging that the current performance gap is largely attributable to the quality of the automated concept annotations and that future work could further enhance the system by integrating sustained vowels and generative reporting tools.