Yunqi He

Yunqi He

Phd. Candidate in Computer Engineering

Project Duration

2022 - Present

Collaboration

Northwestern University, Medical Domain Experts

Keywords

Medical AI, Bayesian Networks, Deep Learning, Multimodal Diagnosis

Code Repository

GitHub

AI for Healthcare

Published:

Project Overview

This project demonstrates the transferability of machine learning expertise across domains by applying advanced AI techniques to medical diagnosis. Our work combines deep neural networks with Bayesian networks for enhanced diagnostic accuracy, showcasing how ML skills developed in hardware domains can address real-world healthcare challenges.

Research Focus

Cross-Domain ML Application

Core Innovation: Applying ML expertise from hardware/EDA domains to solve critical healthcare challenges, demonstrating the universal applicability of advanced machine learning techniques.

Key Achievement: Significant improvement in diagnostic accuracy through novel integration of deep learning with probabilistic reasoning methods.

Technical Innovations

Multimodal Medical Diagnosis

  • Bayesian Networks + Deep Learning: Novel combination achieving 19.3% accuracy improvement over pure deep learning approaches
  • Clinical Data Integration: Seamless incorporation of diverse medical information types
  • Extensible Architecture: Supporting diverse clinical information integration across multiple disease categories

System Architecture

  • Expert Knowledge Integration: Incorporating dermatologists’ domain expertise into AI decision-making
  • Interpretable AI: Providing transparent reasoning paths for medical decisions
  • Modular Design: Flexible framework enabling easy extension to new diseases and clinical scenarios

Performance Optimization

  • Multi-task Learning: Simultaneous optimization across multiple diagnostic objectives
  • Transfer Learning: Leveraging pre-trained models for efficient knowledge reuse
  • Scalable Processing: Handling large-scale medical datasets efficiently

Research Applications

Skin Disease Detection

  • Image Analysis: Advanced deep learning for dermatological image classification
  • Clinical Metadata: Integration of demographic information, symptoms, and medical history
  • Multimodal Fusion: Combining visual and textual medical information for enhanced diagnosis

Diagnostic Decision Support

  • Automated Screening: AI-powered diagnostic capabilities for resource-limited settings
  • Accuracy Enhancement: Achieving specialist-level performance with improved consistency
  • Educational Tools: Transparent reasoning for medical training applications

Performance and Impact

  • Diagnostic Accuracy: 88.1% overall accuracy on ISIC 2019 dataset (23.6% improvement over pure DL)
  • Balanced Performance: 19.3% balanced accuracy improvement across multiple datasets
  • Extensible Design: Easy integration of new diseases and clinical information
  • Industry Relevance: Deployable solutions for real-world healthcare settings

Methodological Approach

  • Domain Transfer: Applying ML expertise from hardware to healthcare challenges
  • Hybrid Methods: Combining deep learning with probabilistic reasoning
  • Practical Validation: Extensive evaluation on medical benchmark datasets
  • Collaborative Development: Working with medical domain experts for clinical validation
  1. A Combination of DNN and BN for Automatic Skin Disease Diagnosis - ISBI 2023
  2. Multimodal Bayesian Networks for Automatic Skin Disease Diagnosis - BIBM 2024

Future Research Directions

  • Domain Expansion: Extension to other medical specialties (radiology, pathology)
  • Healthcare AI Security: Applying hardware security concepts to protect medical AI systems
  • Federated Medical AI: Privacy-preserving collaborative learning for healthcare
  • Real-time Deployment: Development of clinical decision support systems
  • Cross-Domain Innovation: Applying medical AI insights back to hardware/EDA domains