Project Name: CANCER TREATMENT APPLICATION BASED ON PUBMED ARTICLES Documented & Submitted By: Raghavan SU
This project is a Medical AI Chatbot designed to provide Intelligent Healthcare Assistance for cancer treatment. Leveraging Deep Learning models fine-tuned on medical literature, the system aims to:
- Bridge the gap between complex medical knowledge (from sources like PubMed) and user-friendly communication.
- Assist medical professionals in quick information retrieval and decision support.
- Provide accessible medical insights for patients.
The system employs a multi-task learning approach combining three distinct BioBERT and GPT-2 models for a cohesive output.
| Component | Base Model | Task | Key Function |
|---|---|---|---|
| Classification Module | Fine-tuned BioBERT | Cancer Type Prediction | Categorizes user queries (symptoms, diseases, treatment) to route the request. |
| Question Answering (QA) Module | Fine-tuned BioBERT | Knowledge Extraction (Extractive QA) | Retrieves relevant answer spans from medical context (MedQuAD/PubMed). |
| Response Refinement Module | Fine-tuned GPT-2 | Text Generation | Rephrases AI-generated responses for natural language flow and coherence. |
| Client Side | Swift-based Mobile App | User Interface | Provides a mobile interface for real-time interaction (OncoChats). |
| Backend | Flask API | Deployment | Secure query transmission and accessibility of the model pipeline. |
The model was evaluated using a multi-task approach. While individual tasks performed strongly, the Overall Accuracy reflects the complexity of the integrated system.
| Metric | Component | Value | Interpretation |
|---|---|---|---|
| Classification Accuracy | BioBERT Classification | 80% | Strong performance in cancer type classification (Breast Cancer: 38% highest prediction). |
| Question Answering (EM) | BioBERT-QA on MedQuAD | 63% | Reliable and consistent question-answering capabilities. |
| Response Quality (BLEU) | Fine-tuned GPT-2 | 68% | High linguistic similarity and good quality of generated text. |
| Overall Accuracy | End-to-End Pipeline | ≈ 34% | Indicates the challenge of balancing and integrating three complex tasks effectively in a multi-task learning environment. |
- Optimizer: AdamW (
LR=0.0001,Weight Decay=1e-5) for stable updates. - Regularization: L2 Regularization, Dropout (0.3), and Early Stopping to mitigate overfitting.
- Dataset Split: 80% Training / 20% Testing.
| Model / Task | Dataset Source | Preprocessing Techniques |
|---|---|---|
| BioBERT QA | MedQuAD Dataset (from PubMed) | Tokenization, Special Token Addition ([CLS], [SEP]), Padding/Truncation. |
| BioBERT Classification | All cancer types model from Kaggle | BioBERT Tokenizer, Label Encoding, Cross-Entropy Loss. |
| GPT-2 Model | Health Counseling Conversations (Kaggle) | GPT-2 Tokenizer, Lowercasing, Merge conversations into structured format. |
Due to the split nature of the project (Swift Client + Python/Flask Backend), both environments must be set up.
-
Secure Setup:
- The project requires an OpenAI API Key for the
APIService.swift. The key must be stored as an environment variable, NOT hardcoded. The backend will not run without a valid key.
- The project requires an OpenAI API Key for the
-
Backend (Flask API):
- Clone the repository.
- Set up a Python environment and install dependencies (
pip install -r requirements.txt). - Run the Flask server.
-
Client (Swift Mobile Application):
- Open the project in Xcode.
- Ensure all necessary pods/dependencies are installed.
- Configure Firebase/Google Sign-In by adding a valid
GoogleService-Info.plistfile (if necessary). - Run the application on a simulator or physical device.
Home page
Chat page
Book appointments
Symptoms
Lab Result
Setting page
This project is licensed under the GNU General Public License v3.0 (GPL-3.0) - see the LICENSE file for details.