Adult-onset Still’s disease (AOSD) is a systemic, autoinflammatory disorder that was first described in the early 1970s [21]. Most patients with AOSD present with high fever, transient rash, arthralgia or arthritis, and sore throat [21, 22]. The clinical features of AOSD are extremely similar to those of sepsis, also known as “Subacute septicemia,” especially in the early stage when fever is the initial clinical manifestation, it is often difficult to make a differential diagnosis between the two. Severe complications of AOSD are often associated with poor early inflammation control. However, there is no specific method to distinguish AOSD from sepsis at an early stage, resulting in delayed diagnosis and treatment [8, 23]. Currently, although some biomarkers have been explored for differentiating between AOSD and sepsis, none of them possess specificity. So far, there is no reliable model for discriminating between AOSD and sepsis, and there is also no reported application of machine learning methods in establishing a discrimination model. Therefore, we established a differential diagnosis model by combining common clinical features and laboratory tests and screened features by comparing three machine learning methods, including RF, GBDT, and LR.

The establishment of the model is a process of gradual exploration. In the initial stage, white blood cell count, arthralgia, monocyte percentage, α1-acid glycoprotein, ferritin, and sore throat were selected. The prediction model established by GBDT had the best effect (AUC: 0.9755, ACC: 0.9324, Sens: 0.9600, Spec: 0.9017). This model has preliminarily achieved satisfactory results in the differential diagnosis of AOSD and sepsis and has high sensitivity and specificity. In order to use fewer indicators to obtain better prediction effects, we also selected 6 features after the product ratio processing of all indicators in the second stage. The AUC: 0.9573, ACC: 0.8781, Sens: 0.9270, Spec: 0.8678, by comparison, it was found that the prediction efficiency of the model was not as good as that of the first stage. Therefore, we continued to explore, combine, and further screen the features extracted in the first two stages, and finally established a model with fewer indicators (5 indicators: arthralgia, ferritin × lymphocyte count, white blood cell count, ferritin × platelet count, α1-acid glycoprotein/creatine kinase) and higher prediction efficiency (AUC: 0.9916, ACC: 0.9457, Sens: 0.9556, Spec: 0.9578).

Arthralgia is one of the common symptoms of AOSD, and the commonly affected joints are the knee, wrist, ankle, elbow, and proximal interphalangeal joints [8, 10, 24,25,26], which is characterized by mild symptoms in the early stage and easy to be ignored. Sepsis often causes arthralgia because of joint or muscle infection and is characterized by typical joint symptoms of redness, swelling, heat, and pain, although the incidence of arthralgia in sepsis is small [27]. Arthralgia symptoms of AOSD can be relieved with the decrease of body temperature, but arthralgia in sepsis has no such characteristics. White blood cell count, a commonly used predictor of inflammation, is increased in both AOSD and sepsis. So far, there is no report that white blood cell count can be used to distinguish AOSD from sepsis, but white blood cell count was selected in our model. In the report by Fautrel B. et al., ferritin and glycosylated ferritin can be used for the diagnosis of AOSD, and glycosylated ferritin ≤ 20% can be used as one of the diagnostic criteria for AOSD [28, 29]. However, ferritin is increased in diseases such as infectious diseases and tumors, and when ferritin is used alone as a diagnostic marker, the specificity for the diagnosis of AOSD is poor, regardless of the threshold used [30]. Glycosylated ferritin is not readily available in most Settings and is therefore not practical in clinical practice. Zhang M et al. found through a retrospective study that lymphocyte count may be used as one of the indicators for the differential diagnosis of AOSD and sepsis, but the AUC of lymphocyte count alone was only 0.6760, and when it was combined with thrombocytocrit and ferritin, the AUC was 0.8360, the specificity was 0.9230, but the sensitivity was only 0.6730 [4]. Ge S. et al. suggested that platelet count to thrombocytocrit ratio (PMR) could be used as one of the differential diagnosis indicators of AOSD and sepsis. However, in the validation set, the AUC, sensitivity, and specificity of PMR alone as a differential diagnosis indicator were only 0.712, 0.8889, and 0.4286, even if PMR and ferritin were combined, all the evaluation indexes were improved, but the effect of differential diagnosis between AOSD and sepsis was still not satisfactory [6]. In our model, the related indexes discussed above also appeared, but they appeared in the form of products or ratios, such as ferritin × lymphocyte count, ferritin × platelet count, etc. Either alone or in combination, the performance of the above indicators in differentiating AOSD from sepsis was lower than that of the model established by GBDT (AUC: 0.9916, ACC: 0.9457, Sens: 0.9556, Spec: 0.9578).

First identified in 1950, α1-acid glycoprotein is produced mainly by the liver and some extrahepatic tissues and is increased in disease states such as infection, inflammation, and cancer [31,32,33]. α1-acid glycoprotein is a commonly used diagnostic biomarker [34]. Connelly M. A. et al. suggested that α1-acid glycoprotein can be used as a useful indicator to assess the activity of some autoimmune diseases [35]. Sun Y. et al. found that urinary α1-acid glycoprotein levels were significantly higher in AOSD patients than in non-AOSD patients [36]. In a prospective study, Ipek IO et al. found that two consecutive α1-acid glycoprotein measurements had a high sensitivity in the early diagnosis of neonatal sepsis, but a single α1-acid glycoprotein measurement had limited diagnostic value [37]. All these evidences indicate that α1-acid glycoprotein plays an important role in the diagnosis of inflammatory and autoimmune diseases, but there is no report on α1-acid glycoprotein used in the differential diagnosis of AOSD and sepsis.α1-acid glycoprotein was included in our first model screening. In our second model screening, although the single index of α1-acid glycoprotein was removed, the α1-acid glycoprotein/creatine kinase feature appeared. Therefore, we suggest that α1-acid glycoprotein plays an important role in the differential diagnosis of AOSD and sepsis but further studies are needed to confirm this.

Creatine kinase (CK) is found primarily in cardiac muscle, skeletal muscle, and brain tissue, with smaller amounts also found in lung, gastrointestinal tract, and thyroid tissues, which release sufficient amounts to increase their activity when diseased [38]. Creatine kinase elevation may occur in disease states such as muscle injury, brain tissue injury or tumor, hypothyroidism, and toxic effects of statins. Current studies indicate that myocardial dysfunction and prolonged muscle weakness are major causes of critical illness and death from sepsis [39,40,41]. Elevations in creatine kinase are often observed in patients with sepsis during cardiac injury and during skeletal muscle ischemia caused by sepsis-related hypotension [42,43,44]. Although some AOSD patients may present with muscle pain, there is no evidence that AOSD can cause muscle and myocardial damage, and the correlation between AOSD and creatine kinase has not been reported. This index was not screened in the first single index model, but appeared in the form of α1-acid glycoprotein/creatine kinase in the second multiple index product ratio. Therefore, whether creatine kinase can be used alone as a differential diagnosis between AOSD and sepsis is uncertain.

The limitations of this study include the following: first, it is a retrospective study, and all patients’ medical history data were obtained from the internal electronic medical record system of our hospital, which may not be accurate in the collection and recording of medical history, so information bias cannot be avoided. Secondly, due to the limitation of clinical sample size in our hospital, although the conclusions obtained by the statistical methods used in this study have good accuracy, the reliability needs to be further investigated. We plan to continue to expand the sample size in the future, and hope to further verify the results in clinical practice to make the conclusions more accurate. Finally, the differential diagnosis between AOSD and sepsis has always been a difficult problem to be solved. There may be other influencing factors that have not been further explored in the study, and further exploration is needed in the future.