🧠 Disease Prediction from Symptoms using ClinicalBERT

This project fine-tunes a domain-specific transformer model, ClinicalBERT, to predict diseases from natural language symptom descriptions.

Unlike traditional models that rely on keyword matching, ClinicalBERT captures the deep semantic understanding of patient complaints — enabling a more accurate and context-aware diagnosis pipeline.

---

📊 Dataset

The dataset contains structured mappings between symptoms and diagnosed diseases:

• dataset.csv: Symptom descriptions and labels
• symptom_description.csv: Full definitions of symptoms
• symptom_precaution.csv: Suggested precautions for each disease
• Symptom_severity.csv: Severity levels for all known symptoms

✅ Diseases are evenly distributed to avoid bias during training.

---

🏗️ Project Pipeline

🔧 Training Notebook (disease_prediction_Bio_ClinicalBert.ipynb)

1. Convert symptom keywords into natural human-like sentences
2. Encode disease labels using LabelEncoder
3. Tokenize with Bio_ClinicalBERT tokenizer
4. Fine-tune ClinicalBERT using Hugging Face Trainer
5. Evaluate using Accuracy, Precision, Recall, F1-score
6. Save model, tokenizer, and label encoder

🚀 Real-time Inference Notebook (real_time_disease_prediction.ipynb)

1. Load fine-tuned ClinicalBERT, tokenizer, and label encoder
2. Accept natural symptom inputs like:
   "I've been throwing up and my joints hurt, I don’t feel like eating"
3. Predict Top 3 Diseases with confidence scores
4. Visualize predictions with:
   • ✅ SHAP graphs for symptom-level explainability
   • ✅ LIME explanations to interpret prediction logic
   • ✅ Cosine similarity heatmap of disease embeddings

---

🧪 Model Performance

After 10 epochs of training, the model achieved:

Metric	Value
Accuracy	100% ✅
Precision	100% ✅
Recall	100% ✅
F1 Score	100% ✅

📌 Reasons:

• Clear, natural symptom phrasing
• Balanced dataset
• Domain-specific medical language model (ClinicalBERT)

---

📈 Example Output (Real-Time Inference)

Input:
"Acidity, headache, and depression with loss of appetite"

Prediction:

Top Predicted Diseases:

Typhoid (21.6%) Migraine (15.4%) Malaria (11.5%

---

✔️ Followed by SHAP explanation plot + cosine similarity heatmap

---

🖼️ Visuals & Explainability

• SHAP Plot: Highlights which words contributed most to the prediction
• LIME: Shows token-level influence on classification
• Cosine Similarity Heatmap: Reveals disease embedding closeness (e.g., Typhoid and Malaria are similar)

---

🙏 Acknowledgements

• Model: Bio_ClinicalBERT
• Data: Publicly available symptom-disease datasets from Kaggle
• Frameworks:
  • Transformers
  • scikit-learn
  • SHAP
  • LIME
  • Seaborn

---