A scoping review and comparison of the current state of non-generative artificial intelligence in the diagnostics and prognostics of rare diseases – focusing on paediatric populations

Abstract Background The low prevalence and heterogeneous nature of rare diseases make them particularly difficult to diagnose and prognosticate. These issues are amplified in the paediatric population, who are disproportionately affected by rare diseases. Non-generative artificial intelligence (AI) is a powerful tool that can aid clinicians in diagnostics, risk stratification and patient subgrouping. The extent to which this has been applied to rare diseases, in particular paediatric rare disease cohorts, is not well defined. Objective This scoping review aimed to identify how AI techniques have been used to aid diagnosis and prognosis of rare diseases, with a focus on their use in paediatric populations. Methods Embase, Medline, Web of Science, IEEE Xplore and Scopus databases were searched for original articles using the search terms “(‘machine learning’ OR ‘artificial intelligence’) AND (‘rare’ OR ‘orphan’) AND (‘condition*’ OR ‘disease*’ OR ‘disorder*’)”. Results One hundred and thirty-six studies met inclusion criteria, 28 (20.6%) of which used paediatric cohorts, with a variety of study types [diagnostic ( n = 61, 44.9%), prognostic ( n = 39, 28.7%), classification ( n = 19, 14.0%) and screening ( n = 17, 12.5%)]. The number of participants per study ranged from 16 to over 3 million. Tree-based models were the most frequently used in the studies ( n = 62, 45.6%) followed by neural networks ( n = 49, 36.1%) and linear models ( n = 46, 33.9%); the most frequent methods used in paediatric cohorts were neural networks ( n = 11, 39.2%), tree-based ( n = 10, 35.7%), and linear ( n = 10, 35.7%) models. Pre-processing steps were only described in a small number of studies, including feature selection ( n = 41, 30.1%), handling data missingness ( n = 36, 19.1%), data cleaning ( n = 44, 32.4%), addressing data imbalance ( n = 21, 15.4%), and data augmentation ( n = 23, 16.9%). Only 27 (19.9%) models were externally validated. Clinician involvement was described in 56 studies (41.2%), and no models had been translated into regular use in clinical practice. Conclusions AI tools to aid diagnosis and prognosis in rare diseases are increasingly being investigated, but their use for paediatric rare disease cohorts remains scarce. Limited methodological description of model development, clinician involvement and lack of external validation means there has been poor translation of models into clinical practice.