projects

Developed a novel attention-based deep convolutional neural network that revolutionizes automated epileptic seizure classification. The system processes EEG signals from the Temple University Hospital Seizure Corpus, extracting 11 sophisticated features including time-based correlation, eigenvalues, power spectral density, and wavelet coefficients.

Key Achievements

• 98.4% testing accuracy on multiclass seizure classification
• Outperformed traditional CNN by 15.4% (CNN alone achieved 76.7% accuracy)
• Successfully classified 7 different seizure types with high precision
• Published in Epilepsy & Behavior, a Q1 journal in neuroscience
• Demonstrated potential for real-world clinical deployment

Technical Approach

The multi-head attention mechanism enables the model to focus on the most discriminative patterns across different seizure types. The architecture combines:

• Feature Extraction - 11 sophisticated features from EEG signals
• Multi-Head Attention - Captures diverse seizure patterns
• Deep CNN - Processes spatial-temporal relationships
• Transformer Encoders - Layer normalization for stability

Performance Metrics

Metric	Value
Training Accuracy	99.1%
Testing Accuracy	98.4%
Weighted F1 Score	90.2%
Seizure Types	7 classes
Dataset	TUSZ v1.5.2

Technologies Used

Python, TensorFlow, Keras, NumPy, MNE, PyEDF, Matplotlib, EEG Processing, Deep Learning, Attention Mechanisms

Impact

This breakthrough in automated epilepsy diagnosis has potential to assist neurologists worldwide in faster and more accurate seizure classification, ultimately improving patient care and treatment outcomes.

Timeline

February 2022 - November 2022

Collaborators

• Muhammad Ayaz Shirazi
• Syed Sajjad Haider Zaidi

Currently developing a state-of-the-art hybrid architecture that revolutionizes knee injury diagnosis from MRI imaging. The system combines the power of Vision Transformers with specialized convolutional networks, achieving exceptional performance in detecting ACL tears, meniscus injuries, and abnormalities.

Key Achievements

• Developed novel hybrid VGG-Transformer architecture
• Achieved AUC of 0.919 for abnormal detection after 50 epochs
• Strong performance with 0.809 AUC for ACL tears and 0.760 for meniscus tears
• Average AUC of 0.845 across all injury types
• Advancing automated orthopedic diagnosis through AI

Technical Approach

The hybrid architecture leverages:

• Vision Transformers - Global attention mechanisms for anatomical context
• VGG Features - Deep spatial feature extraction
• OverLoCK ConvNet - Specialized localization and classification
• Multi-Task Learning - Simultaneous detection of multiple injury types

Performance Metrics

Detection Type	AUC Score
Abnormal Detection	0.919
ACL Tear	0.809
Meniscus Tear	0.760
Average	0.845
Training Epochs	50

Technologies Used

Python, PyTorch, TensorFlow, Vision Transformers, VGG, ConvNets, Medical Imaging, NumPy, Scikit-learn

Impact

This work has potential to revolutionize orthopedic diagnosis and reduce healthcare costs through automated MRI analysis, enabling faster and more accurate injury detection.

Timeline

January 2025 - Present

Collaboration

University of Liverpool Team

Engineered a groundbreaking hybrid architecture that revolutionizes financial time series forecasting by combining the computational efficiency of Echo State Networks with the long-term memory capabilities of LSTM networks.

Key Achievements

• Achieved exceptional R-squared of 0.975 on stock price prediction
• 94.12% directional accuracy - industry-leading performance
• Lowest MAE (513.10) and RMSE (650.59) among compared models
• Published at ICRAI 2024 international conference
• Demonstrated superior performance over individual ESN and LSTM models

Technical Approach

The system processes 5 years of KSE-100 historical data with sophisticated feature engineering including:

• Moving Averages - Multiple time windows for trend detection
• Technical Indicators - RSI, MACD, Bollinger Bands
• Hybrid Architecture - Combining ESN efficiency with LSTM memory
• Advanced Preprocessing - Normalization and feature scaling

Performance Metrics

Metric	Value
R-squared	0.975
Directional Accuracy	94.12%
MAE	513.10
RMSE	650.59
Data Period	5 years (2019-2024)

Technologies Used

Python, LSTM, Echo State Networks, NumPy, Pandas, Scikit-learn, Financial Analysis, Time Series, Technical Indicators

Impact

Significant advancement in emerging market forecasting with practical applications for investment decisions in developing economies like Pakistan.

Timeline

2023 - 2024

Collaborators

• Syed Ibrahim Zahid
• Asia Petroleum Limited

Currently developing an advanced epidemiological surveillance system using TimeGPT for accurate influenza forecasting. The system leverages state-of-the-art transformer-based time series modeling to predict influenza trends and patterns, significantly outperforming traditional epidemiological models.

Key Achievements

• Developed TimeGPT model achieving RMSE of 873.27 for 1-week forecasts
• Consistently outperformed traditional epidemiological models
• Contributing to global influenza surveillance capabilities
• Working with University of Liverpool’s Nixon Research Group
• Advancing AI applications in public health and disease prevention

Technical Approach

The system utilizes:

• TimeGPT Foundation Model - Pre-trained on diverse time series data
• Multi-Scale Analysis - Capturing seasonal and trend components
• Ensemble Methods - Combining multiple forecast horizons
• Real-Time Processing - Continuous data integration for surveillance

Performance Metrics

Metric	Value
1-week Forecast RMSE	873.27
Model Type	TimeGPT
Forecast Horizon	1-4 weeks
Performance	Superior to traditional models

Technologies Used

Python, TimeGPT, Epidemiological Modeling, NumPy, Matplotlib, Pandas, Time Series Analysis, Public Health Data

Impact

Critical advancement in disease surveillance with potential to save lives through early outbreak detection and improved public health preparedness.

Timeline

May 2025 - Present

Collaboration

Nixon Research Group, University of Liverpool