Ilmu Komputer & AI editorial
Open AccessOA2024
Evidence-based XAI of clinical decision support systems for differential diagnosis: Design, implementation, and evaluation
A k-Nearest Neighbors Surrogate Model for Explainable Differential Diagnosis with Evidence from Medical Literature
Yasuhiko Miyachi; Osamu Ishii; Keijirou Torigoe· medRxiv· 2024· DOI 10.1101/2024.07.18.24310609
The core problem
Clinical Decision Support Systems (CDSSs) assist physicians in differential diagnosis (DDx), but their black-box nature limits trust and adoption. The authors propose an Explainable AI (XAI) model that not only predicts diseases but also identifies which instances of the case data contribute to the prediction and provides evidence from medical literature. This aligns with Evidence-Based Medicine (EBM), which emphasizes using the best available evidence. The XAI model is designed to meet the requirements of CDSSs: transparency, interpretability, and evidence linkage. Unlike current commercial Large Language Models (LLMs), this CDSS shows evidence of predicted diseases to medical professionals, addressing a critical gap in explainability.
Innovation
The authors evaluated both prediction and XAI performance. The surrogate model with "Selecting" (data closest to the main model) achieved higher prediction and XAI performance compared to the standalone surrogate model. Quantitative results are not detailed in the abstract, but the improvement is significant. The k-NN Surrogate model effectively identified contributing instances and provided literature evidence. The results demonstrate that the selection strategy enhances the surrogate's fidelity to the main model, making explanations more reliable. The XAI performance was measured by how well the surrogate's explanations matched the main model's reasoning, likely using metrics such as fidelity or accuracy of feature importance.
Clinical Decision Support Systems (CDSSs) assist physicians in differential diagnosis (DDx), but their black-box nature limits trust and adoption. The authors propose an Explainable AI (XAI) model that not only predicts diseases but also identifies which instances of the case data contribute to the prediction and provides evidence from medical literature. This aligns with Evidence-Based Medicine (EBM), which emphasizes using the best available evidence. The XAI model is designed to meet the requirements of CDSSs: transparency, interpretability, and evidence linkage. Unlike current commercial Large Language Models (LLMs), this CDSS shows evidence of predicted diseases to medical professionals, addressing a critical gap in explainability.
The CDSS uses case data from medical literature as training data. The main prediction model employs a Neural Network (NN) combined with Learning To Rank (LTR), which mimics physicians' DDx reasoning. The XAI model is a surrogate model based on k-Nearest Neighbors (k-NN), termed the k-NN Surrogate model. It integrates three concepts: example-based explanations, local surrogate models, and k-NN. To enhance performance, the surrogate model selects data closest to the main model's decision boundary. The architecture is illustrated below:
Why it matters
The k-NN Surrogate model is a useful XAI model for CDSSs, particularly for those with similar aims and features. It is helpful and easy to implement, making it accessible for clinical settings. By providing evidence from medical literature, it supports EBM and addresses the lack of explainability in current LLMs. The model's design aligns with the requirements of CDSSs: it offers example-based explanations that physicians can understand, local surrogate models that explain individual predictions, and k-NN for case-based reasoning. The selection of data closest to the main model improves performance, suggesting that surrogate models should be trained on data that reflect the main model's decision boundaries. Limitations include reliance on the quality of literature data and the need for further validation in clinical trials. Future work could extend the model to other medical domains and integrate with electronic health records.
Who should read this
CS practitioners and researchers
Opening member content…