NeuroScreen is a multi-modal clinical screening application designed to facilitate the early detection of neurological stress and motor kinematic anomalies, particularly those associated with neurodegenerative conditions like Parkinson's disease. By analyzing both vocal biomarkers and fine motor control through drawing tasks, the platform provides a comprehensive, non-invasive, and data-driven risk assessment.
The system is engineered with a strong emphasis on clinical validity, diagnostic integrity, and explainability, ensuring that its outputs are not only accurate but also transparent and interpretable for healthcare professionals.
NeuroScreen integrates two distinct analysis modules to create a holistic patient profile.
Utilizes parselmouth, a Python interface to the Praat phonetics software, to perform a clinically-standard acoustic analysis on voice recordings. The system extracts key vocal biomarkers that correspond to the widely-used UCI Parkinson's dataset:
- MDVP:Fo(Hz), Fhi(Hz), Flo(Hz): Measures the mean, maximum, and minimum fundamental frequency (pitch), indicating overall vocal range and control.
- Non-Linear Dynamics (spread2, D2, RPDE): Quantifies complex, non-linear vocal cord variations and fundamental frequency dispersion, which are highly predictive of Parkinsonian dysarthria.
- SHAP-Pruned Architecture: Noisy legacy features (like specific Jitter and Shimmer variances) were mathematically pruned to reduce background noise and improve generalization.
Employs a deep learning model built on MobileNetV2 via transfer learning. The model is specifically trained on a dataset of physical, pen-and-paper spiral drawings to identify subtle micro-tremors, dysmetria, and kinematic irregularities characteristic of Parkinsonian motor deficits.
To move beyond "black box" predictions, the dashboard provides dynamic, explainable feedback for every analysis:
- Grad-CAM Heatmaps: The system integrates Gradient-weighted Class Activation Mapping (Grad-CAM) to generate heatmap visualizations. These highlight the specific regions of a drawing that most influenced the model's prediction.
- Acoustic Mel-Spectrograms: Voice analysis provides detailed Mel-Spectrogram "heatmaps" generated via
librosato visually represent the acoustic frequency distribution over time. - OpenCV Contour Isolation: Augmented with an OpenCV contour detection pipeline to precisely frame the most anomalous area identified by the heatmap, providing clinicians with actionable feedback on the patient's motor control and a localized "Kinematic Variance" score.
Several critical architectural decisions were made to ensure the platform's clinical validity and robustness.
To ensure clinical validity and prevent false positives arising from environmental noise, a rigid OpenCV preprocessing pipeline is enforced on all input images, both during model training and live inference. This eliminates training-serving skew, a common failure point in production ML systems.
The pipeline consists of:
- Grayscale Conversion & Alpha Blending: Handles various input formats (RGBA, BGR, etc.) and correctly blends transparent canvas drawings onto a white background.
- Gaussian Blur (5x5 Kernel): Smooths high-frequency noise before thresholding.
- Otsu's Global Thresholding: Robustly separates the ink from the background, creating a clean binary image without the noise amplification common with adaptive thresholding.
- Aggressive Cropping & Padding: The system locates all ink contours, calculates an absolute bounding box, and crops the drawing tightly, adding a strict 10-pixel padding to standardize the input.
The MobileNetV2 model was trained using a two-phase strategy to maximize sensitivity and specificity.
- Phase 1 (Warm-up): The entire MobileNetV2 base was frozen, and only the custom classification head was trained. This allows the new layers to learn the task-specific feature space.
- Phase 2 (Fine-Tuning): The top 50 layers of the MobileNetV2 base were unfrozen, and the model was re-compiled with a significantly lower learning rate (
1e-5). This allows the model to make precise adjustments to its feature extractors, adapting them to the unique characteristics of spiral drawings.
NeuroScreen's dual-modality triage system boasts an average accuracy of ~84.6%, with engineering decisions heavily prioritizing clinical sensitivity (Recall) to minimize False Negatives:
- Voice Acoustic Model (XGBoost): Achieves 86.95% Accuracy (5-Fold Cross-Validated) and a massive 98.92% Clinical Sensitivity (Recall) by using a manually shifted decision boundary (threshold: 0.39) on a 17-feature pruned dataset.
- Kinematic Vision Model (MobileNetV2): Achieves 82.31% Final Accuracy and 81.91% Clinical Sensitivity via deep transfer learning and kinematic-safe data augmentation (strict ≤10° rotation, flipping, nearest fill).
The platform intentionally prohibits the use of digital drawing inputs and strictly requires photo uploads of physical drawings. This critical design choice ensures the model analyzes an unfiltered representation of the patient's true motor function without software masking.
The platform is wrapped in a robust, secure, and user-friendly Flask application built with a modern Tailwind CSS frontend. Key system-level features include:
- Patient Profiling & Medication Tracking: Patients can configure age, gender, and their medication state ("ON", "OFF", "UNMEDICATED") to provide rich clinical context.
- Clinician Portal: Specialized access for healthcare providers to review patient cohorts sorted by risk.
- Longitudinal Patient Tracking: The History Dashboard uses Chart.js to plot a longitudinal trend of the patient's final risk score over time.
- Side-by-Side XAI Visualization: The clinical review interface directly contrasts the true original, full-color uploaded drawing alongside the generated Grad-CAM heatmap.
- Comprehensive PDF Reporting: Using the
fpdflibrary, the application generates formal, downloadable medical reports containing biomarkers, patient demographics, medication states, and XAI imagery.
- Backend: Flask, Python
- Machine Learning: TensorFlow, Keras, Scikit-learn, XGBoost
- Data Processing: OpenCV, Parselmouth, Librosa, NumPy, Pandas
- Frontend: Tailwind CSS, Jinja2, Chart.js
- Database: SQLAlchemy with SQLite
Follow these steps to run the NeuroScreen application on your local machine.
-
Clone the Repository
git clone <your-repository-url> cd NeuroScreen
-
Create and Activate a Virtual Environment
# For macOS/Linux python3 -m venv venv source venv/bin/activate # For Windows python -m venv venv .\venv\Scripts\activate
-
Install App Dependencies
Navigate to the
appdirectory and install the required packages:cd app pip install -r requirements.txt -
Run the Flask Application
python app.py
-
Access NeuroScreen Open your web browser and navigate to
http://127.0.0.1:5000.
Group No.: 11
Batch & Div.: CS(SE) - C
Academic Year: 2025 - 2026 (Semester 2)
| Member Name | Github Username | Email ID | Role |
|---|---|---|---|
| Surabhi Yadav | @surabhiyadav11 | surabhi.1251080184@vit.edu | Group Leader |
| Siddhesh Chavan | @Sid72020123 | siddhesh.1251080264@vit.edu | Assistant Group Leader |
| Sparsh Zalwade | @sparshzalwade | sparsh.1251080260@vit.edu | Member |
| Punam Surana | @suranapunam07 | punam.1251080097@vit.edu | Member |