← Back to Projects

Oncowise AI Chatbot

NLP Engineering • Medical AI • Full-Stack Development • 2024

BioBERT PyTorch Flask SwiftUI Multi-Task Learning

Project Overview

Oncowise AI is a multi-task healthcare assistant designed to provide intelligent decision support for cancer treatment. By fine-tuning large language models on medical literature (PubMed), the system translates complex clinical data into accessible insights for both medical professionals and patients.

Who is it for?

For Patients

Receive personalized medication suggestions tailored to your unique lab results. With automated result uploading, you no longer have to manually enter complex data; the AI interprets your reports and provides clarity on treatment steps and durations immediately.

For Doctors

Focus on high-level clinical decisions. The Doctor's Dashboard allows you to view all patient-AI interactions. This eliminates repetitive questions during visits, as you can see exactly what the patient has asked and what the AI has already clarified.

The Challenge

Medical knowledge is vast, technical, and often locked behind complex research papers. Building a reliable chatbot requires more than simple text generation; it necessitates accuracy in classification, precision in knowledge extraction, and coherence in natural language response to ensure user safety and trust.

The Innovation: Lab-to-Medication Pipeline

The core value of Oncowise lies in its ability to analyze raw laboratory data (e.g., blood markers, tumor profiles) and map them to standard treatment protocols. By providing data-grounded suggestions on medication type and duration, the system ensures higher adherence and better clinical results.

Technical Architecture

The system employs a sophisticated cascaded pipeline combining three distinct fine-tuned models:

  • Classification Module (BioBERT): Categorizes user queries to route requests effectively. Achieved 80% accuracy.
  • QA Module (BioBERT-QA): An extractive model trained on MedQuAD to pull precise answer spans from PubMed.
  • Refinement Module (GPT-2): Rephrases technical data into patient-friendly, natural language.

Infrastructure & Deployment

The backend is powered by a Flask API that handles the model inference logic via AWS Cloud Processing/Google Colab, while the frontend is a modern SwiftUI mobile application integrated with Firebase for secure authentication.

Model Performance Metrics

80%

Classification Accuracy

63%

QA Exact Match (EM)

68%

BLEU Score (Linguistics)

Development Stack

Built with AdamW optimization, L2 regularization, and Dropout (0.3). The project manages API security via environment variables to protect OpenAI and backend endpoints.

Future Roadmap

  • Enhanced Dataset: Expanding training to 10,000+ clinical samples.
  • Contextual Memory: Implementing Pinecone for patient history.
  • Edge Optimization: Local inference on mobile for improved privacy.