This project presents a novel game-based cognitive assessment system designed for the early detection of dementia. The system integrates deep learning techniques with an interactive gaming environment to evaluate cognitive health through both physiological data and facial analysis.
The approach combines:
- A 1D Convolutional Neural Network (1D-CNN) for analyzing health metrics
- A 2D Convolutional Neural Network (2D-CNN) for analyzing facial images
- A rule-based fusion method to generate the final prediction
- Develop an engaging game-based platform for cognitive assessment
- Detect early signs of dementia using AI-driven models
- Combine multimodal data (health metrics + facial images) for improved accuracy
- Provide a scalable and accessible screening tool
The system consists of two primary modules:
- Model: Modified 1D Convolutional Neural Network (MOD-1D-CNN)
- Input: Health metric data (e.g., clinical and physiological indicators)
- Dataset Size: 1000 samples
- Labels:
demented,non-demented
Performance:
- Loss: 0.2692
- Accuracy: 70.50%
- Model: Modified 2D Convolutional Neural Network (MOD-2D-CNN)
- Input: Facial images
- Dataset Size: 1800 images
- Labels:
demented,non-demented
Performance:
- Loss: 0.1755
- Accuracy: 95.72%
A rule-based linear weightage method is used to combine outputs from both models:
- Assigns weights to predictions from MOD-1D-CNN and MOD-2D-CNN
- Produces a final classification decision
-
Health Metrics Dataset
- Source: Apollo Diagnostic Center and associated hospitals
- Samples: 1000
- Type: Structured numerical data
-
Facial Image Dataset
- Samples: 1800
- Type: Image data
- Labels: Demented / Non-demented
- Python
- HTML
- CSS
- JavaScript
- TensorFlow
- NumPy
- Pandas
- OpenCV (for image processing)
- Scikit-learn
- Interactive game-based assessment
- Multimodal deep learning integration
- High accuracy in facial-based dementia detection
- Lightweight and scalable architecture
| Model | Data Type | Loss | Accuracy |
|---|---|---|---|
| MOD-1D-CNN | Health Metrics | 0.2692 | 70.50% |
| MOD-2D-CNN | Facial Images | 0.1755 | 95.72% |
- Increase dataset size and diversity
- Improve health-metric model performance
- Integrate real-time assessment features
- Deploy as a mobile or web application
- Clinical validation with medical professionals
HOSEN ARAFAT
Bachelor of Software Engineering, China
GitHub: https://github.com/arafathosense
Research Interest: Image Computing and Perceptual Intelligence