Presentation (CTI)

OBJECTIVES: In recent years, machine learning (ML) methods have been increasingly applied in real-world evidence studies. The objective of this scoping literature review is to summarise the applications of ML-based techniques for treatment patterns analysis using real-world healthcare and medical data.
METHODS: Observational studies based on real-world datasets (including patient charts or registries, administrative claims data, and electronic health records), using ML methods for treatment patterns analysis, registered in MEDLINE or Embase databases (accessed via the OVID platform) by the date of search conducted on the 21^st of May 2025, were eligible. No publication date limit was applied. Information on the data source, study design, study population, as well as machine learning methodology was extracted. Key characteristics of ML algorithms used in included studies, such as clustering methods, number of clusters, cluster selection criterion, validation, interpretation, and distance metric, were obtained and analysed.
RESULTS: Of the 233 abstracts screened, 16 met eligibility criteria and were included for data extraction. Included studies were published between 2018 and 2024. Eleven of them used clustering methods to analyse treatment patterns (time-sequence analysis through K-means clustering, spectral clustering using similarity network fusion, temporal clustering using latent Dirichlet allocation), while 5 applied different ML methods at the stage of constructing treatment lines or for predicting treatment switches, add-ons, and prescription duration.
CONCLUSIONS: The application of machine learning methods in treatment patterns analysis remains relatively limited. Unsupervised techniques, particularly clustering, are the most commonly employed, whereas the overall diversity of machine learning approaches in this context remains narrow.