A Machine Learning Algorithm for the Detection of Paroxysmal Nocturnal Haemoglobinuria (PNH) in UK Primary Care Electronic Health Records

Abstract Background Paroxysmal Nocturnal Haemoglobinuria (PNH) is an ultra rare acquired clonal haematopoietic disorder. It is difficult to diagnose due to a variety of common symptoms, heterogeneity between patients and lack of awareness and recognition in primary care settings. As a result many PNH patients experience a long diagnostic odyssey and misdiagnoses. To improve diagnostic rates, this study aimed to develop a machine learning-based clinical case-finding tool which identifies patients at high risk of having undiagnosed PNH.Methods The study uses data from the Optimum Patient Care Research Database, which contains electronic health records from general practitioner (GP) practices across the United Kingdom. A PNH group was identified as records with an existing code for a PNH diagnosis and non-PNH patients were identified as records without an existing code for a PNH diagnosis. Clinically driven features including symptoms, diagnoses, healthcare utilisation and exclusionary features, from 131 PNH patients and 593,838 non-PNH patients, were inputted to a tree-based XGBoost machine learning model to classify patients.Results 60.4% of cases classified as positive by the final model were PNH cases (positive predictive value) and 27% of PNH cases were correctly classified (recall) and specificity was 99.9%. The adjusted positive predictive value (PPV) was 19.59 and can be interpreted as nearly 1 in 5 cases flagged by the algorithm may be confirmed as PNH. The five most important clinical features in the model were aplastic anaemia, pancytopenia, haemolytic anaemia, myelodysplastic syndrome, and Budd-Chiari. Other features that contribute as top features in the model are haematology referrals, urology referrals, and haptoglobin tests.Conclusion This is the first study to combine expert understanding of the clinical presentation of PNH with machine learning to develop a clinically-useful algorithm which uses real-world electronic health records to identify possible undiagnosed PNH patients. This algorithm has the potential to expedite diagnosis of PNH and improve outcomes for both patients and clinicians.