🫁 LungAllytics
Two Independent AI Tools for COPD Assessment
LungAllytics is a research-oriented machine learning project for COPD assessment, combining clinical/spirometry data and chest X-ray imaging. It deliberately keeps the two modalities as independent, separately validated tools rather than a fused ensemble — see "Why not a fused ensemble?" below.
🔗 Live demo: https://huggingface.co/spaces/Aishwarya-2006/copd-insight
🚀 Key Features
Clinical severity staging from symptoms/vitals (voting ensemble) Chest X-ray screening for COPD-related abnormality (ResNet50 CNN) A second imaging pipeline exploring PCA-based dimensionality reduction on ResNet50 features, classified with XGBoost Explainable AI via Grad-CAM heatmaps on the imaging model Deployed as a Flask web app, containerized with Docker, hosted on Hugging Face Spaces
🧠 Models Used
Clinical severity model → Voting ensemble (Random Forest + KNN + XGBoost), SMOTE for class imbalance — 91% test accuracy on Mild/Moderate vs. Severe/Very Severe staging Imaging model (fine-tuned) → ResNet50, end-to-end fine-tuning — 97%+ test accuracy on Normal vs. Affected X-ray classification Imaging model (PCA pipeline) → ResNet50 (frozen feature extractor) → PCA (256 components) → XGBoost — 97.6% accuracy, 0.998 AUC-ROC; paired with Grad-CAM for interpretability
An earlier version of this project attempted to fuse the clinical and imaging model outputs into a single probability score. That fusion produced only 64% accuracy — worse than either model alone. Investigation showed why: the clinical model predicts COPD severity among already-diagnosed patients, while the imaging model predicts disease presence from an X-ray — two different questions, and the two source datasets (230 clinical records, ~1,500 X-ray images) have no shared patient IDs to begin with. Rather than force an invalid fusion, the two models are presented as independent diagnostic tools.
⚙️ Core Technologies
Python · scikit-learn · XGBoost · TensorFlow/Keras · imbalanced-learn (SMOTE) · Flask · Docker · Hugging Face Spaces
🔬 Methodology
Clinical severity staging via a heterogeneous voting ensemble, with SMOTE to address class imbalance Chest X-ray feature extraction and classification via transfer learning on ResNet50 (both fine-tuned and frozen-feature-extractor variants) Dimensionality reduction via PCA on ResNet50 feature vectors Explainable AI visualization via Grad-CAM Deployment as a containerized Flask web app on Hugging Face Spaces