Supporting the Diagnosis of Fabry Disease Using a Natural Language Processing-Based Approach

Identifying rare diseases in clinical settings remains a challenge due to the non-specific nature of symptoms. Existing tools often lack the precision needed to guide early diagnosis. Prior research has shown that electronic health records (EHRs) contain valuable data that could support diagnostic processes. However, extracting meaningful patterns from unstructured clinical notes is difficult. This gap motivated the development of a new approach to improve diagnostic accuracy for Fabry disease (FD). No prior work had resolved how to systematically use natural language processing (NLP) for this purpose. The need for a reliable method to detect FD-suspected patients is clear. This study aimed to explore whether NLP could enhance the identification of FD cases. The potential for integrating NLP into clinical workflows is significant.

The goal was to create a decision-support system for identifying patients suspected of having Fabry disease. The specific problem addressed was the difficulty in recognizing FD due to its non-specific symptoms. The motivation came from the need to improve early diagnosis in clinical settings. The study focused on using NLP to extract relevant clinical features from EHRs. The researchers proposed that structured and unstructured data could be combined to generate a risk score. The system was designed to assist physicians in referring patients for further testing. The approach aimed to reduce diagnostic delays and improve accuracy. This method could support early intervention for FD patients.

The team used natural language processing to analyze electronic health records for FD-related features. Clinical features were identified based on literature and expert knowledge. The NLP system extracted patient-specific data from unstructured clinical notes. Laboratory results and ICD-10 codes were also included in the analysis. These data were grouped into pre-defined FD-specific categories. Each category was assigned a score based on its relevance to FD signs. The total score formed the FD risk score for each patient. Physicians reviewed records of patients with the highest scores to decide on referrals.

The NLP-based system achieved an AUC of 0.998 in identifying FD-suspected patients. One patient with a high FD risk score was confirmed to have FD after DBS assay. The system demonstrated strong discrimination power in detecting potential cases. The scoring method effectively combined structured and unstructured data. The highest-scoring patients were accurately flagged for further evaluation. The model's performance suggests it can support clinical decision-making. The integration of NLP improved the accuracy of FD identification. The results suggest this method could be used in routine clinical practice.

The authors propose that the NLP-based scoring system can support the early diagnosis of Fabry disease. The high AUC suggests the method has strong diagnostic accuracy. The system's ability to integrate structured and unstructured data is a key finding. The study shows that NLP can extract meaningful clinical features from EHRs. The researchers suggest that this approach could reduce diagnostic delays. The model's performance supports its potential use in clinical workflows. The study does not claim the system is essential for all FD cases. The findings suggest further testing in diverse patient populations.