AI Text Algorithm Classifies Sickle Cell Disease Subtypes Across Five French Hospitals

Sickle cell disease is the most common inherited blood disorder in France, yet researchers who want to study it using the vast digital archives of hospital records have long faced a deceptively simple problem: the records themselves do not say, in any standardized way, which form of the disease a patient actually has. Now, a team at the Greater Paris University Hospitals has shown that a lightweight text-mining algorithm can solve much of that problem, classifying thousands of patients into precise sickle cell subtypes from free-text clinical reports in minutes, while exposing just how unreliable the standard diagnostic codes used in medical research can be. The study, published in the Journal of Medical Systems, offers a template for how rare-disease research might unlock the potential of large medico-administrative databases without sacrificing accuracy.

Sickle cell disease is caused by mutations in the hemoglobin gene and comes in several distinct genotypes. The most common and often most severe form is hemoglobin SS, but patients may also carry hemoglobin SC disease or one of two sickle beta-thalassemia variants, Sβ+ and Sβ0, each with different clinical trajectories and treatment needs. Complicating matters further, many people carry only the sickle cell trait, a benign carrier state that must be reliably distinguished from actual disease in any epidemiological analysis. Hospital databases, which increasingly fuel retrospective research on these rare conditions, typically identify patients through ICD-10 diagnostic codes, but those codes frequently record only that a patient has sickle cell disease, not which subtype, and sometimes they flag patients who do not have the disease at all.

To address this, Loris Azoyan of Sorbonne Université, INSERM and the Assistance Publique – Hôpitaux de Paris, along with Olivier Steichen and colleagues across five French expert centers, developed an algorithm built from regular expressions: pattern-matching rules that scan unstructured clinical text for the specific phrases clinicians use when documenting a patient’s hemoglobin genotype. The approach deliberately avoided heavyweight machine learning. Instead, it relied on the fact that French clinical reports, from consultation notes to laboratory results, tend to mention sickle cell genotypes in formulaic ways, using standard genetic shorthand that can be captured with carefully crafted textual patterns. The researchers made the regular expressions, the R analysis script containing the classification algorithm, and fictional illustrative patient texts openly available in an Open Science Framework repository, so that other teams can adapt the method to their own data warehouses.

The scale of the evaluation was considerable. The team included every adult patient with at least one ICD-10 code for sickle cell disease who had consulted or been hospitalized at least once between August 1, 2017 and January 29, 2025 at the five participating expert centers of the Greater Paris University Hospitals. In total, 10,868 patients met the entry criteria, and their records contained 561,924 clinical documents issued between 2013 and 2025. The algorithm processed all of those reports in five minutes, a throughput that would be unthinkable with manual chart review and that demonstrates why text-mining tools are becoming essential for research on conditions whose clinical details live in narrative documents rather than structured fields.

The results painted a nuanced picture of the cohort. For 1,259 patients, no sickle-related term could be found anywhere in their clinical reports, immediately raising questions about why their coded diagnoses existed in the first place. Among the remaining 9,609 patients, 1,802, or 18.8 percent, had only non-specific sickle mentions that did not permit any classification. Another 1,502 patients, 15.6 percent, were classified as having sickle cell trait rather than disease. The algorithm assigned 6,234 patients, or 64.9 percent, to a definite sickle cell disease subtype: 4,443 of them, or 71.3 percent, had hemoglobin SS disease; 1,431, or 23.0 percent, had SC disease; 263, or 4.2 percent, had the Sβ+ form; and 97, or 1.6 percent, carried the more severe Sβ0 genotype. Just 71 patients, 0.7 percent, remained unresolved even after the full analysis.

To determine whether the algorithm’s assignments could be trusted, the researchers conducted a manual review of a stratified random sample of 1,510 patients, using that expert adjudication as the reference standard against which the automated classification was measured. The algorithm achieved an overall positive predictive value of 94.0 percent, with a 95 percent confidence interval of 92.6 to 95.4, and subtype-level performance ranged from 91.4 percent to a perfect 100 percent. In practical terms, when the algorithm says a patient has hemoglobin SC disease or Sβ0 thalassemia, that statement is very likely to be confirmed by a human expert reading the same records. For retrospective cohort studies, where the alternative is either laborious manual review or crude diagnostic codes, that level of precision represents a meaningful advance.

Perhaps the most consequential finding of the study concerns the ICD-10 codes that medico-administrative research typically relies on. The team evaluated several commonly used selection strategies based on those codes, varying how many codes were required to confirm a patient’s status and how unclassified patients were treated. The positive predictive value of code-only strategies ranged from as low as 58.1 percent to 88.3 percent depending on those assumptions. In other words, depending on the strategy chosen, a substantial fraction of patients identified through coding alone may not actually have sickle cell disease, or may carry only the trait. The authors warn that this creates a real risk of misclassification when researchers rely solely on administrative coding, a concern that echoes previous validation work on hospital discharge diagnoses in France and on sickle cell case definitions in the United States.

The implications reach beyond sickle cell disease itself. As hospital data warehouses and national health databases become standard infrastructure for epidemiology, the gap between what is coded and what is clinically true has become a central methodological challenge. Previous efforts have used natural language processing to extract information from clinical text for populating registries, identifying patients with specific conditions, and detecting complications such as contrast-associated kidney injury. The Paris team’s contribution is to demonstrate that even a deliberately simple, transparent set of pattern-matching rules, rather than a complex deep-learning model, can achieve high accuracy for a genetically defined condition, provided clinicians document the relevant details consistently. The authors emphasize, however, that their performance figures may reflect the specific subtype demographics, the expertise of the five expert centers, and local documentation practices, so other institutions should expect to validate and adapt the approach on their own data.

The study also touches on the emerging role of generative artificial intelligence in medical informatics research, though in an unusual way. The authors disclose that they used OpenAI’s ChatGPT solely to generate fictional illustrative clinical texts demonstrating the algorithm’s functionality, allowing them to publish realistic French-language medical report examples without exposing any real patient data. No generative AI tool was used for the analysis itself, the design of the algorithm, or the writing of the manuscript, and the authors take full responsibility for the generated content. The underlying clinical reports and patient-level data, meanwhile, cannot be shared publicly because of their sensitive nature and French data protection regulations, with access restricted under the governance of the AP-HP Clinical Data Warehouse, whose scientific and ethics committee approved the retrospective study under the French MR-004 reference methodology for reusing healthcare data in research.

For clinicians and researchers, the message is twofold. First, high-quality clinical documentation matters: the algorithm’s success depended entirely on clinicians writing genotype information into reports, and the authors highlight the importance of such documentation for retrospective research. Second, any study that builds a sickle cell cohort purely from ICD-10 codes should treat its findings with caution, because a meaningful proportion of code-positive patients may be misclassified. By pairing an open, reusable text-classification algorithm with a rigorous manual validation and an honest assessment of coding accuracy, the Paris team has provided both a practical tool and a methodological warning, ones that could shape how sickle cell disease research is conducted across France and beyond as hospitals continue to digitize their clinical narratives.

Subject of Research: Validation of a regular-expression algorithm for classifying sickle cell disease subtypes from unstructured clinical reports and assessing ICD-10 coding accuracy

Article Title: Classifying Sickle Cell Disease Subtypes from Clinical Reports: Algorithm Validation and ICD-10 Accuracy Assessment in Five French Hospitals

Article References: Azoyan, L., Leblanc, J., Santin, A., Chantalat-Auger, C., Le Jeune, S., Affo, L., Arlet, J.-B., Bartolucci, P., & Steichen, O. (2026). Classifying Sickle Cell Disease Subtypes from Clinical Reports: Algorithm Validation and ICD-10 Accuracy Assessment in Five French Hospitals. Journal of Medical Systems, 50(1), Article 134. https://doi.org/10.1007/s10916-026-02461-7

Image Credits: AI Generated

DOI: 10.1007/s10916-026-02461-7

Keywords: sickle cell disease, sickle cell trait, natural language processing, electronic health records, ICD-10, clinical text mining, hemoglobin SS, hemoglobin SC, regular expressions, rare disease research, hospital databases, diagnostic coding accuracy