Project 1: PulmoScope

Project 1: PulmoScope#

PulmoScope is a deep-learning–based assistive system for disease-centered respiratory sound analysis.
This project investigates the effectiveness of temporal deep learning models, with emphasis on a Hybrid Temporal Convolutional Network–Spiking Neural Network (TCN-SNN), for multi-class classification of lung diseases using auscultation sounds.

Final Manuscript#

The complete research paper detailing the background, methodology, experimental framework, results, and discussion is provided below:

PulmoScope – Final Manuscript (PDF)

This manuscript presents a comparative evaluation of TCN-SNN, Pure TCN, LSTM, and Vanilla RNN architectures under standardized preprocessing, balancing, and evaluation protocols. :contentReference[oaicite:1]{index=1}

Exploratory Data Analysis (EDA)#

EDA Notebook: EDA_Pulmoscope.ipynb

This notebook focuses on:

  • Clinical–demographic data exploration

  • Class distribution and imbalance analysis

  • Temporal and spectral characteristics of lung sounds

  • Statistical relationships between age, gender, and diagnoses

The EDA phase guided key modeling decisions such as class consolidation, balancing strategies, and feature design.

Model Development and Experiments#

Modeling Notebook: PulmoScope.ipynb

This notebook documents:

  • Audio preprocessing and feature extraction

  • Hybrid Mel-Spectrogram + MFCC feature stacking

  • Implementation of:

    • Hybrid TCN-SNN

    • Pure TCN

    • LSTM

    • Vanilla RNN

  • Architecture comparison, hyperparameter tuning, and evaluation

  • Confusion matrix, ROC-AUC, and Grad-CAM interpretability analysis

The notebook demonstrates an end-to-end deep learning workflow for respiratory disease classification.

Learning Outcomes#

Through this project, I learned how to:

  • Design disease-centered respiratory sound classification pipelines

  • Handle severe class imbalance in medical audio datasets

  • Compare temporal deep-learning architectures fairly

  • Interpret model decisions using explainability techniques (Grad-CAM)

  • Align deep-learning experiments with clinical relevance and safety

Contents