Live Demo: Click here to test the app
NeuroVoice AI is an end-to-end medical screening tool designed to detect early vocal biomarkers associated with Parkinson's Disease (PD).
Unlike traditional "Black Box" AI models, this system is built on Explainable AI (XAI) principles. It uses advanced signal processing to extract micro-tremors from voice recordings and provides a transparent, clinically interpretable risk assessment.
Most voice AI models fail outside the lab because they confuse Microphone Noise with Neurological Tremors.
- The Challenge: Consumer laptops have a high noise floor (Low Signal-to-Noise Ratio).
- My Solution: I engineered a Dynamic Calibration Engine that scales the risk probability based on the audio's HNR (Harmonics-to-Noise Ratio). This effectively "subtracts" the hardware noise, preventing false positives for users with cheap microphones.
The dashboard features a dynamic Risk Gauge, real-time Biomarker Cards (Jitter/Shimmer), and a downloadable Clinical Report.
The model was trained on the UCI Parkinson's Dataset, a gold-standard collection of biomedical voice measurements.
| Metric | Score | Interpretation |
|---|---|---|
| Accuracy | 89.74% | Highly accurate classification |
| ROC-AUC | 0.9655 | Excellent separation between Healthy/PD |
| Precision (PD) | 0.93 | 93% of identified cases were correct |
| Recall (PD) | 0.93 | Caught 93% of actual PD cases |
precision recall f1-score support
0 0.80 0.80 0.80 10
1 0.93 0.93 0.93 29
accuracy 0.90 39
macro avg 0.87 0.87 0.87 39
weighted avg 0.90 0.90 0.90 39
The model correctly identified 27/29 Parkinson's patients and 8/10 Healthy controls in the blind test set.
Before training, I used t-SNE (t-Distributed Stochastic Neighbor Embedding) to visualize the high-dimensional data in 2D.
- Insight: The clear separation between Healthy (Green) and PD (Red) clusters confirmed that vocal features contain a strong predictive signal, validating the feasibility of an ML approach.
- Signal Processing: Used
Parselmouth(Python wrapper for Praat) to extract 22 acoustic features. - Selection: Removed collinear features (e.g.,
Jitter:RAP,Jitter:PPQ) to prevent overfitting and improve model interpretability. - Extraction Strategy: Implemented a "Stability Scout" algorithm that scans the audio file window-by-window to find the single most stable 1-second segment, ignoring breath/start-up artifacts.
Medical AI must be transparent. I used SHAP (SHapley Additive exPlanations) to break down why the model made a specific decision.
- Top Predictor: PPE (Pitch Period Entropy) and Spread1 were identified as the strongest biomarkers for Parkinson's, aligning with clinical literature on vocal rigidity.
- Frontend: Streamlit, Plotly (Interactive Charts)
- Backend: Python
- Machine Learning: XGBoost (Extreme Gradient Boosting)
- Audio Processing: Praat / Parselmouth
- Explainability: SHAP (Shapley Additive Explanations)
-
Clone the Repo
git clone [https://github.com/K-Ashik/NeuroVoice-AI-Parkinson-Prediction.git](https://github.com/K-Ashik/NeuroVoice-AI-Parkinson-Prediction.git) cd NeuroVoice-AI -
Install Dependencies
pip install -r requirements.txt
-
Run the App
streamlit run app.py
For Educational & Portfolio Purposes Only. This tool is a proof-of-concept for tele-medicine screening. It is not a certified medical device and should not be used for self-diagnosis. If you have health concerns, please consult a neurologist.
Created by [Khalid Md Ashik] - LinkedIn Profile | Portfolio | Github