Improving a data mining based diagnostic support tool for rare diseases on the example of M. Fabry

\(\bf Background\) Rare diseases often present with a variety of clinical symptoms and therefore are challenging to diagnose. Fabry disease is an x-linked rare metabolic disorder. The severity of symptoms is usually different in men and women. Since therapeutic options for Fabry disease exist, early diagnosis is important. An artificial intelligence (AI)-based diagnosis support algorithm for rare diseases has been developed in preliminary studies. \(\bf Objective\) Our aim was to extend and train the questionnaire-based AI, capable of distinguishing patients with from those without rare diseases, to achieve satisfactory sensitivity for the detection of a single rare disease, Fabry disease, taking into account gender differences in disease perception. \(\bf Methods\) We collected 33 complete datasets from patients with confirmed Fabry disease. These records contained answered AI questionnaires, general information on disease progression, demographic information and quality of life (QoL) measures. The AI was trained to distinguish patients with Fabry disease from patients with relevant differential diagnoses. Its performance was assayed using stratified eleven-fold cross-validation and ROC curve calculation. Variables influencing the performance of the AI were examined with linear regression and calculation of the coefficient of determination. \(\bf Result\) We were able to show that a relatively small sample is sufficient to achieve a sensitivity of 88.12% for the presence of Fabry disease, taking into account gender-specific differences in the disease perception during the pre-diagnostic phase. No confounders of the tool’s performance could be found in the data collected concerning the patients’ quality of life and diagnostic history. \(\bf Conclusion\) This study illustrates on the example of Fabry disease that differences between female and male Fabry patients, not only in the expression of symptoms, but also with regard to disease perception, might be relevant influencing variables for improving the performance of AI-based diagnostic support tools for rare diseases.