Comment on: Identification of clinical phenotypes and disease trajectories in SLE using AI through a natural language processing framework

Dear Editor, We read with interest the study by Bosello et al. ‘Identification of clinical phenotypes and disease trajectories in SLE using AI through a natural language processing framework’, recently published in Rheumatology [1]. The authors present a novel natural language processing (NLP) pipeline designed to extract clinical characteristics and disease trajectories from unstructured electronic health records (EHRs) in SLE. Their AI-driven framework successfully categorized patients into low-, medium-, and high-complexity phenotypes at first contact and demonstrated that high-complexity patients experienced significantly more flares over time. This work represents a significant advancement in leveraging real-world data for SLE characterization, offering a feasible tool to study disease heterogeneity and progression. However, several methodological and clinical considerations warrant discussion before widespread implementation. First, while the hybrid AI-human intelligence (HI) approach achieved impressive accuracy (≥0.93 for domains, ≥0.94 for therapies), the reliance on rule-based NLP algorithms raises concerns about generalizability across healthcare systems. The pipeline was developed and validated in Italian EHRs from a single tertiary centre, requiring language-specific ontologies and terminologies. Recent improvements in transformer-based architectures, like the authors’ later ‘Lupus Alberto’ model, show that they work better across languages and institutions [2]. Large language models (LLMs) and generative AI for medical record profiling are also interesting options that could make us less reliant on rule-based frameworks that take a lot of time to set up while still being easy to understand [3]. Future versions should think about adding these more scalable methods to make it easier to validate across multiple centres, which is still a major barrier to clinical use [4].