IP-CRR: Information Pursuit for Interpretable Classification of Chest Radiology Reports (2505.00191v1)

Abstract: The development of AI-based methods for analyzing radiology reports could lead to significant advances in medical diagnosis--from improving diagnostic accuracy to enhancing efficiency and reducing workload. However, the lack of interpretability in these methods has hindered their adoption in clinical settings. In this paper, we propose an interpretable-by-design framework for classifying radiology reports. The key idea is to extract a set of most informative queries from a large set of reports and use these queries and their corresponding answers to predict a diagnosis. Thus, the explanation for a prediction is, by construction, the set of selected queries and answers. We use the Information Pursuit framework to select informative queries, the Flan-T5 model to determine if facts are present in the report, and a classifier to predict the disease. Experiments on the MIMIC-CXR dataset demonstrate the effectiveness of the proposed method, highlighting its potential to enhance trust and usability in medical AI.

An Overview of IP-CRR: Enhancing Interpretability in Radiology Report Classification

The paper "IP-CRR: Information Pursuit for Interpretable Classification of Chest Radiology Reports" presents a comprehensive approach to enhancing the interpretability of automatic methods used for classifying chest radiology reports. The impetus for this research stems from the necessity of interpretability in clinical decision-making where the transparency of diagnostic tools is paramount for establishing trust and facilitating safe outcomes. Machine learning (ML) models, despite achieving high accuracy, often lack the requisite transparency, thus hindering their adoption in medical settings.

Methodology and Innovations

The proposed framework, IP-CRR, is built on an interpretable-by-design paradigm for classifying radiology reports. This framework innovatively combines the principles of the Information Pursuit (IP) model with LLMs to create a system that provides interpretable and sequential query-answer chains leading to a diagnosis. The primary components of the IP-CRR methodology involve:

1. Query Generation: The authors develop a robust method for extracting a substantial set of queries from existing radiology reports. Utilizing a large-scale radiology dataset, queries are constructed by mining domain-specific facts that are deemed clinically relevant.
2. Query Answering via LLMs: Given the inability of traditional methods to answer queries lacking explicit supporting text in reports, the authors leverage a LLM, Flan-T5, to infer the existence or absence of facts. This natural language inference (NLI) mechanism deduces whether particular factual statements are applicable to a given report.
3. Interpretable Classification using Variational Information Pursuit (V-IP): The IP-CRR employs V-IP to iteratively and sequentially select queries based on their information gain. This approach ensures that only the most informative queries are chosen to reach a confident prediction, reducing the reliance on opaque post-hoc explanations.

Experimental Setup and Results

The framework was evaluated on the MIMIC-CXR dataset, focusing on the classification of conditions like Lung Opacity, Cardiomegaly, and Pneumonia, among others. The results were characterized by:

• High Interpretability with Competitive Performance: While IP-CRR provided fewer confident predictions than highly tuned black-box models such as CXR-BERT (fine-tuned on all layers), it outperformed baseline models in interpretability with comparable or superior predictive performance.
• Efficiency: The framework achieved high average precision with significantly fewer queries, indicating that it efficiently selected the most informative features for each specific classification task.
• Derivation of Trustworthy Explanations: IP-CRR could generate intuitive query-answer chains that were easily interpretable by medical professionals, aiding in transparent and trustworthy clinical decision-making.

Implications and Future Directions

The development of the IP-CRR framework has significant implications for clinical ML applications. By providing a mechanism for interpretable predictions, this method addresses a critical barrier preventing the integration of advanced ML tools into clinical settings. Practically, this framework may enhance clinician trust and improve diagnostic workflows by offering transparent predictions supported by empirical evidence directly derived from radiological reports.

Theoretically, this research lays the groundwork for future developments in applying interpretable AI in other high-stakes fields. Future research could focus on expanding IP-CRR to incorporate more advanced LLMs and integrating multimodal data sources to further improve accuracy and interpretability. Additionally, the framework could potentially be adapted for real-time decision support systems in clinical scenarios, paving the way for more dynamic patient-care strategies.

In conclusion, the IP-CRR framework emerges as a promising solution to the challenge of interpretability in medical AI, underscoring the synergy of combining traditional ML models with modern LLMs to produce innovative, transparent AI tools for healthcare.