Interpretable Early Detection of Parkinson's Disease Through Speech Analysis

Lorenzo Simone, Mauro Giuseppe Camporeale, Vito Marco Rubino, Vincenzo Gervasi, Giovanni Dimauro

TL;DR

This paper introduces a novel deep learning approach for early Parkinson's Disease (PD) detection using speech analysis. By segmenting speech into word chunks and employing a temporal convolutional network, the model not only achieves competitive diagnostic accuracy but also provides interpretable insights into which specific speech segments and phonetic features are most indicative of PD, potentially aiding in understanding disease progression.

What Problem Does This Paper Address?

Parkinson's Disease (PD) is a neurodegenerative disorder that significantly impacts motor and non-motor functions. Speech impairments, such as a softer voice, monotony, and imprecise articulation, are among the earliest and most common symptoms. Early detection is crucial for effective intervention. While machine learning (ML) has shown promise in identifying these speech patterns, existing methods often rely on handcrafted features that may not fully capture the temporal dynamics of speech. Furthermore, many ML approaches lack interpretability, making it difficult to understand why a particular prediction is made and how it relates to the underlying neuromuscular impairments. This paper aims to bridge this gap by developing an interpretable deep learning model that can detect PD from speech and highlight the specific vocal segments driving the diagnosis.

Methods

The authors propose a deep learning architecture designed for temporal speech analysis. Key methodological aspects include:

  • Speech Segmentation: To handle the variability in speech and focus on meaningful units, audio recordings are segmented into word chunks. This is achieved through a hybrid approach combining word-level timestamps from speech recognition with amplitude-based cutting to ensure robust segmentation.
  • Temporal Convolutional Network (TCN): A TCN is employed to process these segmented word chunks. The architecture consists of convolutional blocks, each comprising a 1D convolutional layer, batch normalization, and a ReLU activation. This allows the model to learn temporal dependencies within the speech segments.
  • Interpretability Framework (Grad-CAM Adaptation): To understand the model's decisions, the authors adapt Gradient-weighted Class Activation Mapping (Grad-CAM) for one-dimensional temporal signals. This technique visualizes which speech segments are most influential in the PD prediction by analyzing the gradients of the final convolutional layer's output with respect to the class score. This allows for the identification of specific phonetic features and articulatory movements that are indicative of PD.

Datasets

The study utilizes the "Italian Parkinson's Voice and Speech Database". This dataset contains 831 audio files from 65 individuals, comprising 37 healthy participants and 28 with Parkinson's disease, aged between 19 and 80. The recordings include various speech samples such as phonetically balanced texts, vowel pronunciations, and syllable repetitions ("pa" and "ta").

Key Insights and Ideas

  • Speech Segmentation is Crucial: Breaking down continuous speech into meaningful units (word chunks) allows for more focused analysis and better capture of temporal dynamics, which is essential for detecting subtle speech changes in PD.
  • Interpretability Reveals Phonetic Markers: The adapted Grad-CAM approach successfully highlights specific speech segments and phonetic features that are most predictive of PD. This goes beyond a simple classification and offers a window into the physiological basis of the observed speech impairments.
  • Articulatory Challenges in PD: The identified key speech segments, particularly those involving front vowels and plosive/fricative consonants, require precise tongue and lip coordination. The fact that these are highlighted by the model suggests that PD-related muscle rigidity and tremor directly impact these fine motor control aspects of speech production.
  • Temporal Dynamics Matter: The use of a temporal network, as opposed to static feature extraction, is critical for capturing the progressive nature of speech deterioration in PD.

Key Results

  • Competitive Performance: The proposed deep learning model achieves competitive classification performance compared to simpler ML models like KNN, SVM, RF, and Gradient Boosting.
  • Statistically Significant Improvements: The model demonstrates statistically significant improvements in accuracy, recall, and F1-score (p < 0.001) over baseline classifiers when analyzing extended speech signals.
  • High Accuracy: For extended speech signals, the model achieved an accuracy of 99.14% ± 1.60, with 100.0% ± 0.0 precision, 98.15% ± 3.46 recall, and 99.03% ± 1.81 F1-score.
  • Identification of Indicative Words: Table 2 lists words like "Cocco," "Ciccio," and "Luna" as being most indicative of PD speech, with their IPA transcriptions and frequencies, suggesting specific phonetic patterns are more affected.

Limitations and Future Work

  • Language Specificity: The current study focuses on the Italian language. The authors suggest future work to examine the model's generalization across different languages, considering the influence of phonetic similarities and native language.
  • Broader Phonetic Analysis: While the study identifies key segments, future research could further explore the relationship between phonation patterns and PD progression.
  • Personalized Monitoring: The ability to identify specific speech markers offers potential for personalized monitoring strategies for individuals with PD.
  • Temporal Dynamics of Deterioration: The authors aim to examine how the model's insights can offer valuable information about the individual progression of the disease over time.

Summary updated

BibTeX
@inproceedings{simone2025interpretable,
  author = {Lorenzo Simone and Mauro Giuseppe Camporeale and Vito Marco Rubino and Vincenzo Gervasi and Giovanni Dimauro},
  title = {{Interpretable Early Detection of Parkinson's Disease Through Speech Analysis}},
  booktitle = {AIME},
  year = {2025},
  doi = {10.1007/978-3-031-95841-0\_69},
}
Figure