An AI chatbot to help medical professionals with medical reports summarization

We aim to develop a medical QA chatbot that can respond quickly and accurately to general medical and patient-specific queries.

1. Objectives
2. Architecture
3. Data Sources & Data Samples
4. Live Demo
5. Modeling Approach
6. Results
7. Conclusions
8. Recommendations & Future Work

• Patient History Retrieval
  • Interact with medical transcripts
  • Extract insights based on patient history
• Medical QnA
  • Ask general medical queries
  • Learn about drug interactions, medical conditions, etc.
• DocScribe
  • Report Summarization
  • Summarize reports from multiple visits
  • Identify relevant information
• Transcribe Medical Reports
  • Convert patient medical report into a more accessible format
  • Making it easier to interpret data from Medical Reports

• Medical Transcripts
  • We leveraged the GPT-3.5 model to generate 4.5k QA prompts from the medical transcripts (MTSamples dataset).
• WikiDoc
  • WikiDoc is a platform for medical professionals to contribute and edit medical content.
  • 10k QnA prompts.
• WikiPatient
  • WikiPatient is a platform for patients to access information and education about diseases.
  • 5K QnA prompts.

Approach:

• Model: Fine-tuning Vicuna-13B
• Training: LoRA, PEFT, bitsandbytes
• Hyperparameters: Same as Alpaca

Frameworks:

• LLM Components: LangChain
• Weights: HuggingFace
• Training: PyTorch
• Embeddings: all-mpnet-base-V2
• Vector database: Qdrant

Our model has successfully achieved its objectives of extracting patient information, generating report summaries, and answering general medical questions.

• Training base model on medical corpus and incorporating human feedback.
• Explore the use of the medical LLM for research and clinical trials, including medical image analysis.