💊 FDA Product Classification using Machine Learning and Deep Learning Models

This project presents a robust multi-class classification pipeline to predict the marketing category of FDA-listed pharmaceutical products. Leveraging a blend of classical ML models and deep learning, the pipeline handles structured metadata to categorize each product into NDA, ANDA, OTC, BLA, or UNAPPROVED.

---

📌 Project Highlights

• Task: Multi-class classification of FDA drug entries by marketing category
• Input Data: Structured metadata (product name, route, substance name, year, etc.)
• Models Used: Logistic Regression, Random Forest, XGBoost, Deep Learning (Keras)
• Techniques: Label Encoding, Normalization, Hyperparameter Tuning, Embedding Layers
• Evaluation: Confusion Matrix, ROC Curve, Accuracy, AUC
• Visuals: SHAP values, Feature Importance, Distribution plots

---

🧪 Dataset

• Source: Public FDA database (processed version available below)
• 📂 Dataset on Hugging Face

---

🛠️ Preprocessing Overview

The dataset includes text, categorical, and numeric fields. Preprocessing included:

• Grouping categories (e.g., combining subtypes under main classes like NDA)
• Label Encoding of categorical fields (e.g., route, substance name)
• Normalization of numeric features (year, name lengths)
• Missing value handling

---

✨ Feature Engineering

• Categorical variables encoded using LabelEncoder
• Numerical features normalized using mean-std scaling
• Feature importance visualized using SHAP and built-in model metrics:

---

🎯 Model Training

Logistic Regression

• Baseline model for reference

Random Forest

• Tuned using GridSearchCV

XGBoost Classifier

• Strong performance, especially for BLA and NDA

Deep Learning (Keras)

• Initial dense model + improved model with Embedding layers

Improved Deep Learning

• Uses embedding layers, tuned architecture, better generalization

---

📊 Distribution Plots

---

📁 Model & Dataset Access

• 📦 Dataset on Hugging Face
• 🤖 All Trained Models on Hugging Face

---

📈 Evaluation Summary

---

🔍 Key Takeaways & Future Work

• Deep learning outperformed traditional ML after sufficient tuning
• Category imbalance addressed via grouping; future work can explore SMOTE
• SHAP revealed strong impact of route and proprietary name features

Future Directions:

• Experiment with BERT-style tabular encoders or TabNet
• Integrate external drug description texts for enrichment
• Add explainability dashboard (Gradio or Streamlit)

---

🧠 Summary

This project showcases a complete, reproducible pipeline for real-world regulatory data classification using both traditional ML and deep learning. It highlights preprocessing, feature importance, visual evaluation, and production-ready model deployment.

🎯 Deployed models and data are hosted on Hugging Face for public access and further experimentation.

GitHub