TLDR: This research proposes a novel method to enhance the interpretability of rule-based AI predictions, specifically for arm lymphedema risk assessment in breast cancer patients. By applying Information Retrieval techniques (like tf-idf) to semantically clustered attributes, the approach provides users with clear, factor-based explanations of risk predictions. A user study demonstrated that this method significantly improves the perceived interpretability and usefulness of AI outputs for both AI experts and healthcare professionals, making complex AI decisions more transparent and trustworthy without requiring model retraining.
Artificial intelligence (AI) is transforming many fields, including healthcare, by offering powerful tools for decision-making. However, the complex, “black-box” nature of many AI systems often makes their predictions difficult to understand, especially for critical applications like patient diagnosis and treatment. This lack of transparency can hinder trust and adoption by clinicians and patients alike.
A new research paper, “Enhancing the Interpretability of Rule-Based Explanations Through Information Retrieval,” addresses this crucial challenge. The authors, Alessandro Umbrico, Luca Coraci, Francesca Fracasso, Silvia Gola, Gabriella Cortellessa, and Guido Bologna, propose an innovative approach to make AI predictions more interpretable, particularly in the context of assessing arm lymphedema risk after breast cancer radiotherapy. Lymphedema is a significant side effect that impacts patients’ quality of life, making accurate and understandable risk prediction vital for personalized treatment planning.
The core of their method lies in improving the interpretability of an existing explainable AI model that uses rule-based predictions. While rule-based models are generally considered more transparent than other AI types, their raw output can still be too technical for non-AI experts like doctors or patients. The researchers introduce an attribution-based approach that statistically analyzes the attributes within the rule-based model using standard metrics from Information Retrieval (IR) techniques, specifically a modified version of tf-idf (term frequency-inverse document frequency).
Instead of directly interpreting complex rules, the proposed system aggregates individual AI model attributes into “semantically coherent factors.” For example, different attributes related to smoking status (SMOKER, FORMER SMOKER, CURRENT SMOKER) can be grouped under a single, understandable factor like “SMOKER.” This semantic clustering allows for flexible levels of detail in explanations, catering to different users—an AI expert might prefer a highly detailed view, while a clinician or patient would benefit from a more abstract, clinically relevant interpretation.
The innovative use of IR metrics helps determine the relevance of each factor to a specific prediction. Unlike traditional IR where a high tf-idf score indicates uniqueness, this approach “reverses” the meaning: factors that appear frequently across the activated rules for a prediction are assigned a higher relevance score. This signifies their greater influence on the outcome, providing clear insights into which risk factors are most impactful for a given patient’s prediction.
The research included an experimental evaluation to assess both the technical feasibility and the interpretability of the approach. For global explanations (analyzing the entire AI model), the system quickly identified factors like BMI, NODES INVOLVED, and RT TECHNIQUE as highly relevant, aligning perfectly with established clinical knowledge regarding lymphedema risk. For local explanations (individual patient predictions), the approach effectively highlighted how the importance of different factors varied from patient to patient, offering personalized insights.
A user study involving both AI experts and non-AI experts (healthcare professionals, caregivers) compared three ways of presenting the AI’s output: raw propositional rules, a radar graph displaying contributing factors, and a list of contributing factors with their percentage impact. The results were compelling: both the radar graph and the list of factors significantly improved the perceived interpretability and usefulness of the AI model’s output compared to the raw rules. This suggests that simplifying the presentation of information, even for already “explainable” models, is crucial for effective communication across diverse user profiles.
Also Read:
- Unlocking AI Transparency: A New Approach to Help Users Evaluate Model Predictions
- MEDGELLAN: Enhancing Medical Diagnosis with AI-Generated Clinical Guidance
This work offers a lightweight and flexible solution that does not require retraining the AI model, making it adaptable to various user needs without costly computational overhead. By bridging the gap between complex AI logic and human understanding, this research contributes significantly to building trust in AI technologies, especially in sensitive domains like healthcare. For more details, you can read the full research paper here.
