top of page

Medical AI and First Response Assistant - AI Project Idea

Optimizing and Fine-Tuning DeepSeek R1 Model Using LoRA on PubMed, MedQA, and BioASQ; Designing a Flutter-Based GUI Integrating Whisper Voice Input and Coqui TTS; Enabled Dual Operational Modes for Emergency Response and Medical Queries.




1. Project Overview

1.1 Project Title: Medical AI and First Response Assistant


1.2 Purpose

The purpose of this project is to develop an AI-powered assistant that serves dual purposes:

  • Emergency Response Mode: Provide rapid, accurate first-response guidance in critical medical situations.

  • Medical Query Mode: Answer detailed medical questions for educational or informational purposes.


The project involves optimizing and fine-tuning the DeepSeek R1 model using Low-Rank Adaptation (LoRA) on medical datasets (PubMed, MedQA, BioASQ) and building a user-friendly Flutter-based graphical user interface (GUI) with Whisper for voice input and Coqui TTS for text-to-speech output.


1.3 Learning Objectives

  • Gain hands-on experience in fine-tuning large language models (LLMs) using LoRA.

  • Understand medical dataset preprocessing and model optimization.

  • Develop skills in cross-platform app development with Flutter.

  • Integrate speech recognition (Whisper) and text-to-speech (Coqui TTS) into a functional GUI.

  • Apply AI in real-world medical and emergency response scenarios.


1.4 Target Audience

  • Students with intermediate knowledge of AI/ML, programming (Python, Dart), and GUI development.

  • Intended end-users: Medical professionals, first responders, and students.



2. Scope

2.1 In-Scope

  • Fine-tuning the DeepSeek R1 model using LoRA on PubMed, MedQA, and BioASQ datasets.

  • Designing and implementing a Flutter-based GUI with two operational modes:

    • Emergency Response Mode

    • Medical Query Mode

  • Integrating Whisper for voice input and Coqui TTS for audio output.

  • Testing the system for accuracy, usability, and responsiveness.


2.2 Out-of-Scope

  • Deployment of the application to app stores.

  • Real-time integration with live medical databases or hardware devices (e.g., wearables).

  • Training the DeepSeek R1 model from scratch (pre-trained model will be provided).

  • Multi-language support beyond English.



3. Functional Requirements


3.1 Model Optimization and Fine-Tuning

  • Dataset Preparation:

    • Preprocess and clean data from PubMed (research articles), MedQA (medical Q&A), and BioASQ (biomedical Q&A).

    • Ensure datasets are formatted for compatibility with DeepSeek R1 and LoRA.

  • Fine-Tuning with LoRA:

    • Use LoRA to efficiently fine-tune DeepSeek R1 for medical domain knowledge.

    • Optimize model performance for accuracy and response time.

  • Evaluation:

    • Test model accuracy on a validation subset of MedQA and BioASQ.

    • Achieve at least 80% accuracy on medical question-answering tasks.


3.2 GUI Development

  • Framework: Use Flutter for cross-platform compatibility (Android, iOS, and desktop).

  • Modes:

    • Emergency Response Mode:

      • Quick, concise responses for first-aid instructions (e.g., CPR, choking).

      • Voice-activated input and audio output prioritized for hands-free operation.

    • Medical Query Mode:

      • Detailed responses to medical questions (e.g., symptoms, drug interactions).

      • Support for both text and voice input/output.

  • Features:

    • Clean, intuitive interface with mode-switching toggle.

    • Real-time voice input using Whisper API/model.

    • Audio output of responses using Coqui TTS.

    • Error handling for poor voice input or model failures.


3.3 Integration

  • Connect the fine-tuned DeepSeek R1 model to the Flutter GUI via an API or local inference.

  • Ensure seamless interaction between Whisper (input), DeepSeek R1 (processing), and Coqui TTS (output).



4. Non-Functional Requirements


4.1 Performance

  • Model response time: < 3 seconds for emergency mode, < 5 seconds for query mode.

  • GUI load time: < 2 seconds on standard hardware (e.g., mid-range smartphone or laptop).


4.2 Usability

  • Interface should be simple and accessible for non-technical users.

  • Voice commands must be recognized with >90% accuracy in quiet environments.


4.3 Scalability

  • System should handle up to 10 concurrent users in testing without performance degradation.


4.4 Reliability

  • Model should provide medically accurate responses in >80% of test cases.

  • GUI should not crash during voice input/output operations.



5. Technical Requirements

5.1 Tools and Technologies

  • AI/ML:

    • DeepSeek R1 (pre-trained model).

    • LoRA for fine-tuning.

    • Python, PyTorch, or TensorFlow for model training.

  • Datasets: PubMed, MedQA, BioASQ (provided or accessible online).

  • GUI:

    • Flutter (Dart) for cross-platform development.

    • Whisper (open-source speech-to-text).

    • Coqui TTS (open-source text-to-speech).

  • Hardware:

    • GPU-enabled system for model training (e.g., NVIDIA GPU).

    • Standard laptop/smartphone for GUI testing.


5.2 Dependencies

  • Pre-trained DeepSeek R1 model availability.

  • API keys or local setup for Whisper and Coqui TTS.

  • Stable internet for dataset downloads and potential API calls.



6. Deliverables

6.1 Code

  • Python scripts for dataset preprocessing and model fine-tuning.

  • Flutter codebase for the GUI application.

  • Integration scripts (e.g., API or local inference setup).


6.2 Documentation

  • Technical Report:

    • Overview of model fine-tuning process and results.

    • GUI design and implementation details.

    • Challenges faced and solutions implemented.

  • User Manual: Instructions for installing and using the application.


6.3 Demonstration

  • A working prototype showcasing both modes (emergency and query).

  • Video or live demo of voice input/output functionality.



7. Timeline

Phase

Tasks

Duration

Deadline

Phase 1: Research

Study DeepSeek R1, LoRA, Flutter, Whisper, Coqui TTS

1 week

Apr 16, 2025

Phase 2: Data Preparation

Preprocess PubMed, MedQA, BioASQ datasets

1 week

Apr 23, 2025

Phase 3: Model Fine-Tuning

Fine-tune DeepSeek R1 using LoRA

2 weeks

May 07, 2025

Phase 4: GUI Development

Build Flutter GUI with mode-switching

2 weeks

May 21, 2025

Phase 5: Integration

Integrate model, Whisper, and Coqui TTS

1 week

May 28, 2025

Phase 6: Testing & Refinement

Test system, fix bugs, optimize performance

1 week

Jun 04, 2025

Phase 7: Documentation

Write technical report and user manual

1 week

Jun 11, 2025

Phase 8: Submission

Submit deliverables and demo

-

Jun 12, 2025

Note: The timeline date is provided for your convenience in planning; you can adjust it as needed and begin working on this project accordingly.


8. Assumptions

  • Students have basic knowledge of Python, Dart, and AI/ML concepts.

  • Access to necessary hardware (GPU for training, laptop/phone for GUI).

  • Datasets and pre-trained models are available or downloadable.



9. Constraints

  • Limited to open-source tools and datasets (no proprietary software).

  • Project must be completed within the 8-week timeline.

  • Team size: 1-4 students (adjust scope based on team size).


10. Success Criteria

  • Fine-tuned model achieves >80% accuracy on medical Q&A tasks.

  • GUI successfully switches between emergency and query modes.

  • Voice input/output works seamlessly in a demo environment.

  • All deliverables (code, documentation, demo) submitted by Jun 12, 2025.



11. Support and Resources

  • Mentor/Instructor: Available for weekly check-ins and troubleshooting.

  • References:

    • DeepSeek R1 documentation.

    • LoRA research papers/tutorials.

    • Flutter, Whisper, and Coqui TTS official docs.

  • Community: Online forums (e.g., Stack Overflow, GitHub issues).


Comentários


bottom of page