In clinical practice, the consideration of non-specific symptoms of rare diseases in order to make a correct and timely diagnosis is often challenging. To support physicians, we developed a decision-support scoring system on the basis of retrospective research. Based on the literature and expert knowledge, we identified clinical features typical for Fabry disease (FD). Natural language processing (NLP) was used to evaluate patients’ electronic health records (EHRs) to obtain detailed information about FD-specific patient characteristics. The NLP-determined elements, laboratory test results, and ICD-10 codes were transformed and grouped into pre-defined FD-specific clinical features that were scored in the context of their significance in the FD signs. The sum of clinical feature scores constituted the FD risk score. Then, medical records of patients with the highest FD risk score were reviewed by physicians who decided whether to refer a patient for additional tests or not. One patient who obtained a high-FD risk score was referred for DBS assay and confirmed to have FD. The presented NLP-based, decision-support scoring system achieved AUC of 0.998, which demonstrates that the applied approach enables for accurate identification of FD-suspected patients, with a high discrimination power.

Background and Objectives: Primary myelofibrosis (PMF) is a rare hematologic malignancy with non-specific symptoms, causing diagnostic delays and missed diagnoses. Automated screening in heterogeneous electronic health records (EHRs) is challenging due to class imbalance, data sparsity, and incomplete labeling. We investigated two complementary objectives: (1) developing a screening algorithm using routine EHR data to identify PMF-risk patients for hematology consultation, and (2) assessing the applicability of positive-unlabeled (PU) learning. Methods: Using EHR data from 10 Polish hospitals, we evaluated several ensemble models and found that LightGBM achieved the best performance. Results: LightGBM achieved AP 20.83% (95% CI: 19.18–24.35%), sensitivity 45.52% (95% CI: 39.48–52.72%), and precision 14.72% (95% CI: 13.80–17.01%)—substantially exceeding expected prevalence (over 340-fold enrichment). The model captured interactions among RDW and PLT parameters revealing high-risk unlabeled patients with confirmed diagnoses (n = 5) or clinical profiles resembling PMF cases (n = 31), confirming PU problem. Conclusions: Although PU methods enhanced sensitivity, they reduced precision to levels exceeding real-world healthcare capacity. Ensemble learning enables disease identification in clinical settings.