I developed a novel Depth of Anesthesia (DoA) index using EEG-derived features to enhance intraoperative patient monitoring in clinical settings. The model was designed to produce BIS-like index values ranging from 0 (deep anesthesia) to 100 (awake).

The project involved rigorous feature selection and model validation using Pearson correlation, R², and Bland-Altman plots to ensure robustness on withheld test data. Supervised machine learning techniques including Linear Regression and Random Forest were employed to model the index from EEG features (x1–x7).

For a separate dataset containing 4,965 EEG segments, I implemented unsupervised learning algorithms (K-means clustering and Gaussian Mixture Models) to identify anesthesia states without direct supervision, labeling them as states A/B to infer depth.

The final solution integrated both supervised and unsupervised approaches into a hybrid ensemble model that combined linear regression and random forest using weighted averaging predictors, resulting in enhanced predictive accuracy and system resilience.

Developed and trained deep learning models (DenseNet-121 & custom CNN) for multi-label classification of chest X-ray images using over 100,000 samples for comprehensive medical image analysis.

Implemented a custom 4-block CNN architecture with SE attention mechanisms, batch normalization, and channel recalibration, specifically optimized for extracting radiology features from medical imaging data.

Fine-tuned DenseNet-121 pretrained on ImageNet with a custom classification head and addressed class imbalance using BCEWithLogitsLoss with positional weighting to improve model performance across all disease classes.

Applied advanced image preprocessing techniques including CLAHE, Gamma correction, ColorJitter, and comprehensive data augmentation (RandomCrop, Flip, Rotation) to enhance model generalization and robustness.

Optimized training pipeline using differential learning rates, AdamW optimizer, and ReduceLROnPlateau scheduler for stable and adaptive training on Google Colab GPU/TPU (T4 & v2–8) with efficient memory usage and dynamic batching for high-throughput processing.

Developed and optimized deep learning models (CNN, LSTM-RNN, CAR-Transformer) for Environmental Sound Classification (ESC) using the UrbanSound8K dataset, achieving high accuracy in noisy, real-world audio environments.

Extracted and engineered audio features (MFCC-40, log-Mel spectrograms, Δ & Δ² features) to enhance temporal and spectral representation for classification tasks.

Applied attention mechanisms and Transformer architectures (CAR-Transformer with Coordinate Attention & Residual Convolutions) to improve robustness, interpretability, and global context modeling.

Implemented data preprocessing and augmentation pipelines (time shifting, pitch shifting, noise addition, time stretching) to address dataset imbalance, increase diversity, and improve generalization.

Conducted performance benchmarking of multiple architectures (CNN, CRNN, Transformer, Generative augmentation) on UrbanSound8K cross-validation folds with comprehensive model inspection methods.

Evaluated model performance using rigorous training, validation, and testing protocols with metrics including accuracy, precision, recall, F1-score, and confusion matrix analysis to assess classification performance across all environmental sound categories.

Integrated ethical considerations into model design, addressing privacy, bias, and consent issues in speech, emotion, and environmental audio domains, aligned with AI ethics frameworks.

Leveraged transfer learning and lightweight architectures (ResNet, DenseNet, Audio Spectrogram Transformer) to scale models efficiently for both research-grade accuracy and real-time applications (IoT, smart cities).