PD-Detection-Models

Multi-modal Parkinson’s Disease detection pipelines featuring models for spiral drawing, tremor, and audio-based analysis.

Part 1: Spiral and wave drawings Model [DONE]

Model used:

DenseNet201 with modified classifier.

Dataset size:

• 2611: Training
• 653: Validation

Number of trained epochs:

• 15: initial model comparisons
• 43: fine-tuning with early stopping; planned 50.

Metrics:

• Validation Accuracy: 97.02%
• Validation Recall: 0.9787
• Validation precision: 0.9609
• Validation F1-Score: 0.9689

Part 2: Tremor Model [IN-PROGRESS]

Model used:

Movement-Aware TremorNetV9.

Model Architecture:

• Multi-Scale Time-Domain CNN: Three parallel CNN branches (fast, mid, slow) with different kernel sizes to capture tremor patterns at multiple frequencies. Includes Squeeze-Excitation (SE) for channels and Temporal Attention for important time steps.
• Dominant Hand Super-Pathway: Dedicated, 3× capacity CNN branch for the dominant hand, emphasizing diagnostically relevant signals.
• Frequency & Statistical Features: Computes log-power spectrograms per IMU channel and coarse band energies; extracts statistical moments (mean, variance, skew, kurtosis).
• Contrastive & Bilateral Features: Captures left-right asymmetry and bilateral coordination via attention mechanisms.
• Clinical Metadata Integration: Encodes age, BMI, family history, alcohol effects; integrates with signal features using cross-attention.
• Movement Embeddings: Encodes movement type to provide context to the model.
• Fusion & Classifier: Concatenates all features, including CNN, dominant-hand, frequency/statistics, asymmetry, bilateral, embeddings, and metadata; uses a deep MLP for final tremor severity prediction.

Dataset size:

• 355 total patients where:
  • 79 Healthy
  • 276 Parkinson’s Disease (PD)
• It consists of 11 different type of movements (each movement recorded twice for each wrist).
• According to the movement type, some are 10 seconds (1024 samples), and others are 20 seconds (2048 samples)
• The signal is structured with 6-channels (3-axis accelerometer + 3-axis gyroscope.)

Total: (355) x (11 x 2) x (1024 X 6)

Number of trained epochs:

• Best Validation: 6
• Best Training: 38

Metrics:

• Best Validation Accuracy: 0.8575

• Best Validation F1-Score: 0.7953

• Best Training Accuracy: 0.9456

• Best Training F1-Score: 0.9456

Confusion matrix at threshold = 0.3

---

Part 3.1: Audio Model (Tubular) [DONE]

Model used:

EfficientNet-B0.

Dataset size:

Generated data (used in pre-training):

• Model used: TVAE (Tabular VAE)
• Generated data accuracy:
  • Column Shapes Score: 83.06%
  • Column Pair Trends Score: 92.87%
  • Overall Score (Average): 87.97%
• Generated: 100k samples
  • 80K: Training
  • 20K: Validation

Real data (used in finetuning):

• Using 80-20 split:
  • 156: Training
  • 39: Validation
• Using 50-50 split:
  • 97: Training
  • 98: Validation

Number of trained epochs:

• 15: pre-training on generated data
• 5: finetuned on real data with early stopping; planned 50.

Metrics:

• Validation Accuracy: 100%
• Validation Recall: 1.0000
• Validation precision: 1.0000
• Validation F1-Score: 1.0000

---

Part 3.2: Audio Model (Audio)

---

Part 4: Subject's Metadata Model

Model used:

EfficientNet-B0.

Generated data (used in pre-training):

• Model used: TVAE (Tabular VAE)
• Generated data accuracy:
  • Column Shapes Score: 91.69%
  • Column Pair Trends Score: 87.36%
  • Overall Score (Average): 89.53%
• Generated: 100k samples
  • 80K: Training
  • 20K: Validation

Real data (used in finetuning):

• Using 80-20 split:
  • 284: Training
  • 71: Validation

Metrics:

• Validation Accuracy: 98.96%
• Validation Recall: 98.81
• Validation precision: 1.0000
• Validation F1-Score: 99.39