Development and assessment of an assisted diagnosis model using machine learning for identifying adult-onset Still's disease in fever of unknown origin: a retrospective study in China.

Abstract

Background: Adult-onset Still's disease (AOSD) is a systemic autoinflammatory disorder lacking a gold-standard diagnostic criterion.

Objective: To develop and validate a clinically applicable model for identifying AOSD among patients with fever of unknown origin (FUO) who have clinical suspicion for AOSD.

Methods: Clinical data (2010-2020) were divided into training and internal test set (7:3) using stratified random sampling according to disease status (AOSD vs non-AOSD). Feature selection was performed using Boruta, recursive feature elimination and least absolute shrinkage and selection operator algorithms. Selected features were used to train logistic regression (LR), random forest and extreme gradient boosting models with fivefold cross-validation. Model performance was evaluated using area under the curve (AUC), receiver operating characteristic curves, sensitivity, specificity and accuracy. External validation was performed at another centre using the same adjudication procedure.

Results: A total of 847 patients were included, comprising a derivation cohort of 771 patients and an independent external validation cohort of 75 patients. Six features-age, neutrophil percentage, white blood cell count, infection indicator, ferritin and 'AOSD-related clinical presentation score'-were consistently selected by at least two algorithms and used to build the model. LR achieved the highest AUC in both training (0.969; 95% CI 0.956 to 0.983) and test sets (0.960; 95% CI 0.934 to 0.985). A nomogram based on the LR model demonstrated good real-world performance in the independent validation cohort, with an AUC of 0.906.

Conclusions: We developed a machine learning model using electronic medical records to identify AOSD among patients with FUO with clinical suspicion for AOSD. The model shows high accuracy and potential for early identification of AOSD in this specific clinical setting.