TL;DR
This paper introduces a novel method using Sparse Autoencoders (SAEs) to interpret "black box" deep learning models for Parkinson's Disease (PD) detection from speech. By uncovering interpretable internal representations, they demonstrate that PD speech biomarkers are linked to specific acoustic features like spectral flux, which in turn correlate with brain structure changes, paving the way for more trustworthy clinical AI.
What Problem Does This Paper Address?
Parkinson's Disease (PD) significantly impacts speech, making it a promising, non-invasive biomarker for diagnosis and monitoring. While deep learning models excel at detecting these subtle speech changes from raw audio, their "black box" nature hinders clinical adoption. Clinicians need to understand why a model makes a prediction to trust and utilize it. Existing methods for explaining these models either build interpretability into the design (which can be restrictive) or provide post-hoc explanations that are often noisy and difficult to interpret, especially for continuous data like speech. This paper aims to bridge this gap by developing a method to extract meaningful, interpretable insights from complex speech models for PD.
Methods
The authors adapt Sparse Autoencoders (SAEs), a technique from natural language processing for uncovering interpretable representations, to the domain of speech. Their approach involves two stages:
- Detector Training: A standard deep learning model (using a frozen Whisper encoder) is trained for PD detection using speech audio. This stage establishes a baseline performance.
- SAE Training: A SAE layer is inserted after the attention pooling layer of the trained detector. This SAE is trained to reconstruct the output of the attention pooling layer while enforcing sparsity. Key novelties include:
- Masking Structure: A novel masking mechanism is used to adapt SAEs for smaller biomedical datasets, allowing for disentangled dictionary representations.
- Small Dictionary Sizes: The SAEs are trained with relatively small dictionary sizes, suitable for the limited data available in biomedical speech research.
- Post-hoc Interpretation: The SAEs are trained after the main detector is frozen, allowing for interpretation without altering the original detection model's performance.
The SAE learns a set of "dictionary entries" that represent internal model activations. By analyzing which features activate these entries, the authors can understand what aspects of the speech the model is attending to.
Datasets
The primary dataset used is from the Quebec Parkinson's Network (QPN), comprising speech recordings from 208 patients with Parkinson's Disease and 52 healthy controls. The data includes various speech tasks like sustained phonation, reading, recalling a memory, and describing a picture. The "cookie theft" picture description task was found to yield the best discrimination performance.
Key Insights and Ideas
- Speech as a Biomarker is Validated: The study reinforces the idea that speech is a rich source of information for detecting PD, with their best-performing models achieving over 80% F1 score.
- Interpretable Features Emerge: The SAE successfully identified interpretable internal representations that correlate with known acoustic features associated with PD speech.
- Attention to Low-Energy/Weak Speech: The model's attention mechanisms tend to focus on regions of low energy in the audio, which can correspond to breath pauses or weak/breathy speech, both known indicators in PD.
- Spectral Flux is Key: A strong correlation was found between the activation of a specific SAE dictionary entry (#61) and spectral flux, a measure related to articulation dynamics. Healthy speech has higher spectral flux due to more dynamic articulation, while PD speech is often flatter.
- Linking Speech to Brain Structure: Crucially, the study demonstrated a significant correlation between spectral flux (and the SAE activation) and the volume of the putamen in the basal ganglia, a brain region known to be affected in PD and involved in motor control. This provides a direct link between the model's interpretation of speech and underlying neuroanatomy.
Key Results
- The trained SAE models correlate with hand-crafted acoustic features, indicating that the deep learning model relies on spectral flux and flatness in low-energy speech regions for PD detection.
- Dictionary entry #61 showed a strong correlation (ρ = 0.851) with spectral flux and a strong correlation (ρ = -0.943) with the final PD prediction score.
- Spectral flux was significantly correlated with the volume of the putamen (left putamen: r = 0.442, p = 0.019; right putamen: r = 0.429, p = 0.044), suggesting a neuroanatomical basis for the speech biomarker.
Limitations and Future Work
- Generalizability: The authors acknowledge that the effectiveness of the SAE technique might be condition-specific, and it's unclear how well it would generalize to other neurological disorders.
- Prior Knowledge Requirement: The approach requires some prior knowledge of how a condition might affect speech to identify meaningful correlations.
- Indirect Interpretations: SAE interpretations are indirect and might miss subtle patterns.
- Validation of Novel SAE Approach: The newly proposed mask-based SAE approach needs further validation across different architectures and datasets.
Future work should focus on applying this method to a wider range of disorders, exploring its potential for longitudinal tracking of disease progression, and investigating its utility for treatment planning.
Summary updated