Assertiveness-based Agent Communication for a Personalized Medicine on Medical Imaging Diagnosis: Assertiveness-based BreastScreening-AI

Document Title

Assertiveness-based Agent Communication for a Personalized Medicine on Medical Imaging Diagnosis

Document Information

• Subject Area: Design and evaluation of personalized clinical decision support systems in the fields of artificial intelligence and medical imaging
• Keywords: Human-computer interaction, explainable artificial intelligence, clinical decision support systems, breast cancer diagnosis, medical image analysis, deep learning, personalized medicine, dynamic communication, assertiveness-based communication

Research Background and Issues

• Problems and Challenges:

  • Artificial intelligence systems show great potential in the medical field, but existing systems often fail to account for the variability in characteristics of individual clinicians or specific groups.
  • The tone of expression used in AI recommendations may influence clinicians' acceptance and clinical workflows, yet there is limited research on designing personalized recommendation communication.
  • High-risk decision-making scenarios demand greater reliability and explainability in communication, which is especially critical in clinical diagnosis.
  • Addressing the "black box" problem of AI—understanding how and why a model produces a specific output.

• Significance:

  • Medical errors can directly impact patient health, making the reduction of misdiagnoses (e.g., false positives and false negatives) critically important.
  • Personalized and precision medicine has become a core direction in modern medical development, and AI systems should support and enhance clinicians' decision-making capabilities.

• Research Motivation and Related Work:

  • Medical imaging analysis systems typically present results as simple numerical outputs, which fail to meet clinicians' needs for diagnostic reasoning.
  • Psychological and HCI research indicates that assertiveness and explainability in communication significantly enhance user trust.
  • The authors drew on related literature concerning deep learning technologies, multimodal medical imaging diagnostic systems, and cue-based design in human-computer interaction.

Solution

• Methods and Solutions:

  • Propose an "Assertiveness-based Communication Agent" for the breast cancer diagnosis domain.
  • Design two communication strategies: non-assertive (suggestive, advisory) and assertive (mandatory, authoritative), enabling the AI agent to adopt dynamic communication capabilities.
  • The AI agent assists clinicians with detailed explanatory information, including the number of detected findings, a visual scale of cancer severity, sensitivity and specificity results, and sources of patient clinical variables.
  • Use DenseNet and ResNet models for classification and segmentation of mammography (MG), ultrasound (US), and magnetic resonance imaging (MRI).

• Innovations:

  • Dynamically adjust the tone of AI recommendations to suit the professional experience of clinicians (e.g., interns vs. senior doctors).
  • Incorporate human-interpretable clinical arguments and visual explanation components into traditional numerical predictions.
  • Apply "assertiveness-based communication theory" to deep learning models in medical scenarios for the first time.

• Implementation Steps:

  • Study Design: 52 clinicians participated in comparative experiments under different tones (assertive, non-assertive) and traditional model conditions.
  • Dataset: Breast images from 289 patients were extracted from real cases to train DenseNet and ResNet models.
  • Human-Computer Interaction System Design: Develop a prototype of assertiveness-based communication integrated with explanatory modules based on existing frameworks.

Research Outcomes

• Specific Results:

  • The assertiveness-based agent significantly improved diagnostic efficiency, reducing average diagnostic time by 25%.
  • While efficiency improved, diagnostic accuracy showed no significant decline.
  • Data revealed that adapting the communication tone had a significant impact on clinicians' decision-making: interns preferred assertive tones, while senior doctors favored advisory tones.
  • Quantitative results supported: the assertiveness-based agent enhanced perceived trust and understanding among clinicians, leading to higher user satisfaction.

• Advantages:

  • Streamlined workflows and reduced clinical misdiagnoses (false positives and false negatives).
  • Increased trust and adoption rates of AI recommendations among clinicians.
  • Enhanced educational value of AI-assisted tools, helping clinicians understand flexible diagnostic processes and learning mechanisms.

• Experimental and Evaluation Results:

  • False positives were reduced by approximately 26%, and false negatives by about 2%.
  • The advantages of assertive communication were particularly evident across different experience levels: diagnostic accuracy improved by approximately 17.4% for interns and 4.4% for senior doctors.

• Limitations and Future Directions:

  • Limitations: The current study focuses solely on breast cancer diagnosis, and its generalizability to other fields remains to be validated.
  • Experimental Control: Limited participation time by clinicians introduced variability in task completion times.
  • Legal Issues: The question of liability in the medical application of AI requires further exploration.
  • Future Directions: Expand the range of communication tones, study the impact of dynamically adjusted communication based on AI confidence levels, and validate findings in other domains (e.g., skin cancer, lung cancer diagnosis).

Additional Information

• Data and Code Repository: The authors have made the model code and statistical analysis data publicly available, linked to a relevant GitHub repository.
• Intellectual Property: This study involves patent applications and is protected by intellectual property declarations.