2026 ASH®
Annual Meeting & Exposition December 12-15, 2026 ▪ New Orleans, Louisiana

Paper Details

Abstracts, authors, affiliations, and disclosures.

Saturday, December 12, 2026
05:30 p.m. - 07:30 p.m.
MCCNO - Hall F
Access, Representation, Population Health, Technology and Procedures, Imaging, Gene Editing, Treatment Considerations, Supportive Care, Biological Therapies, Transfusion, Gene Therapy, Workforce, Adverse Events, Artificial Intelligence (AI), Natural Language Processing, Clinical Practice (Health Services and Quality), Clinical Guidelines, Health Economics, Patient Safety, Practice Models, Quality Improvement, Quality of Care, Systems-Based Hematology, Value, Diseases, Genetic Disorders, Hemoglobinopathies, Sickle Cell Disease, Thalassemia, Research, Clinical Research, Clinical Trials, Epidemiology, Health Outcomes Research, Health Disparities Research, Patient-Reported Outcomes, Real-World Evidence, Registries, Survivorship, Study Population, Human, Adult, Pediatric
PAPER

2593 Diagnostic accuracy of a natural language processing-based radiographic triage system for acute chest syndrome in adults with sickle cell disease

Shenise Gilyard, MD, Camille Motchoffo Simo2, Fatoumata Barry, Medical Student3
1 University of California, Los Angeles, Los Angeles, California, United States
2 University of California, Los Angeles, Los Angeles, California, United States
3 University of California, Los Angeles, Los Angeles, California, United States

Introduction and Objective: Acute chest syndrome (ACS) is the leading cause of disease-related mortality in adults with sickle cell disease (SCD) and is defined by the National Heart, Lung, and Blood Institute (NHLBI) as the acute onset of a new pulmonary infiltrate on chest imaging accompanied by at least one clinical criterion. Radiographic classification is central to ACS recognition but is labor-intensive and subject to inter-reader variability. We developed and validated a CheXpert-style rule-based natural language processing (NLP) pipeline for automated triage of chest imaging reports in adults with SCD, ACS-NLP. 

Methods: We conducted a single-center retrospective diagnostic accuracy cohort study at two hospitals in a single center across a 10-year window (October 1, 2015 to October 1, 2025). ACS-NLP was applied to the impression sections of all chest imaging reports from adults with confirmed International Classification of Diseases, 10th Revision, Clinical Modification (ICD-10-CM) SCD diagnoses, assigning each study to one of four ordinal triage categories: ACS-Positive (positive acute radiographic findings with confirmed etiology), ACS-Maybe (positive acute radiographic findings of etiology not fully assessed), Chronic Only (positive chronic radiographic findings), or ACS-Negative (negative acute and chronic radiographic findings for ACS). A stratified random sample of 301 studies was independently classified by two human reviewers on two axes, acuity and chronicity; 291 were successfully matched to NLP output by order-procedure identifier and constituted the reference-standard analysis sample. ACS-Positive and ACS-Maybe were combined as a single positive category for primary analysis. ACS-Positive definition retained as a secondary analysis. Diagnostic accuracy was assessed against the fully reconciled binary reference standard. 

Results: Among 395 adults with SCD, 346 (87.6%) entered the NLP analytic cohort, contributing 6,007 chest radiographic imaging across 1,841 admissions. Inter-rater reliability between human reviewers on the primary binary definition was κ = 0.82 (N = 291, 92.4% agreement); all residual disagreement (N=22, 7.6% disagreement) was resolved through blinded principal-investigator adjudication, yielding a fully reconciled reference standard (N = 291). Against this reference standard, ACS-NLP demonstrated sensitivity 78.0% (95% CI 68.5-85.3%), specificity 94.5% (95% CI 90.4-96.9%), positive predictive value 86.6% (95% CI 77.6-92.3%), negative predictive value 90.4% (95% CI 85.7-93.7%), ordinal area under the receiver operating characteristic curve 0.894 (95% CI 0.851-0.931), and F1 score 0.821. Independently recorded NHLBI clinical criteria and composite adverse clinical outcomes both tracked the radiographic triage in a monotonic gradient across all four categories, and clinical-criteria positivity did not differ significantly between ACS-Positive and ACS-Maybe (87.3% vs. 83.6%, p = 0.237), nor did nearly all time-matched vital signs and laboratory values, supporting a common underlying clinical phenotype across the two categories. A cohort-wide, triage-label-independent accounting of chronic radiographic findings showed that 82.1% of patients had at least one documented chronic finding, and 73.5% of ACS-positive radiographic impressions also contained co-occurring chronic findings. 

Conclusions: ACS-NLP demonstrates good discriminative accuracy (AUC 0.894) for radiographic identification of ACS finding when uncertain and unambiguous acute interpretations are combined into a single clinically coherent positive category, a grouping supported by both radiological and independently measured clinical evidence. The tool additionally distinguishes acute findings occurring against a background of chronic pulmonary and cardiovascular change from those occurring in an otherwise unremarkable chest radiographic impression, a distinction with potential relevance to risk stratification that will be developed further.

Camille Motchoffo Simo: No relevant conflicts of interest to declare.