RESEARCH PUBLICATIONS

This study uses EEG signals to detect abnormal brain activity associated with schizophrenia. Traditional machine learning requires heavy preprocessing, but deep learning—specifically a CNN—can classify normal and schizophrenic EEG patterns more effectively. The proposed CNN model achieves 96.4% accuracy, outperforming existing methods and demonstrating strong potential for robust early diagnosis.

This study evaluates seven machine learning techniques to identify sinus bradycardia—a type of cardiovascular disease—using ECG signals. A comparative analysis on a small dataset of 94 records shows that the proposed lightweight ANN model achieves the highest accuracy at 95.2%, outperforming KNN, SVM, and Random Forest. The results highlight the strong potential of machine learning models in supporting clinicians with reliable bradycardia diagnosis, even with limited data.

This study analyzes air pollution trends in Lahore from 2003 to 2022 using eight pollutants and four weather factors. Multiple time-series models—including SARIMA, SARIMAX, LSTM, and NAR—were evaluated, with the NAR model achieving the best performance (RMSE 23.52, DTW 5023). Findings indicate a projected 13% increase in AQI by 2030 compared to 2022, emphasizing worsening air quality. The research supports strategic policymaking and aligns with the SDG goal of ensuring “Good Health and Well-Being.

This work introduces a hybrid breast cancer diagnosis model that combines DenseNet121 with explainable AI methods (Grad-CAM & Grad-CAM++) to provide accurate and interpretable predictions on histopathology images. With a customized classification head and strong data augmentation, the model achieves 95.56% accuracy on BreakHis and 91.7% on ICIAR 2018, outperforming VGG16 and ResNet50. Its lightweight design (33 MB, 8.1M parameters) and rapid inference make it suitable for real clinical deployment, especially in low-resource settings.z

This study introduces hybrid ML–DL models (CNN-RF, CNN-KNN, and CNN-LSTM) to improve diabetes diagnosis using the PIMA dataset. The proposed CNN-LSTM model—combining 1D convolutions with an LSTM layer—achieves 84% accuracy, surpassing standalone LSTM (78%) and Random Forest (76%). It also delivers strong clinical metrics (precision 0.88, F1-score 0.89), reducing false negatives by 15%. These results demonstrate the superior diagnostic potential of hybrid architectures for early diabetes detection.