This study is the first to develop a prediction model for approaching the etiologies of FUO in children by integrating clinical parameters from medical histories, physical examinations, and initial laboratory results. The prediction model was based on seven key parameters: duration of fever, cough, arthritis, splenomegaly, lymphadenopathy, anemia, and thrombocytopenia. Following multinomial logistic regression and predictive margin analyses, the model demonstrated good performance, particularly when distinguishing autoimmune diseases and malignancies, as assessed by the AUC. Previous studies described clinical parameters and FUO etiologies but did not demonstrate significant differences among these parameters, likely due to differences in study design and variable selection [1, 4, 7]. They focused on demographic factors or laboratory findings obtained later in the diagnostic process. In contrast, our study emphasized early clinical symptoms, physical examination findings, and basic laboratory tests, which are routinely available at initial presentation. This approach aligns with our aim of developing an early prediction model to support clinical decision-making before initiating extensive diagnostic workups. This model may help reduce unnecessary investigations and inappropriate treatments, especially in resource-limited settings. Our findings emphasize the potential utility of combining these early clinical markers into a systematic tool to aid diagnosis and reduce diagnostic uncertainty.

We found notable similarities and differences when comparing our results to previous studies. Consistent with the findings of Hu et al., the shortest duration of fever was observed in the infections category [7]. Cough emerged as a significant predictor in our model, likely because respiratory tract infections constituted the top three diagnoses in the infections category. Unlike prior studies, which frequently identified leukocytosis in autoimmune conditions due to various diseases, such as Kawasaki disease and inflammatory bowel disease [1, 12], our study did not observe this result. This discrepancy is likely attributable to the predominance of SJIA and SLE in our autoimmune diseases category, with SLE often presenting with leukopenia instead of leukocytosis [4, 12]. Additionally, prolonged fever was consistently associated with autoimmune conditions across various studies, aligning with our results [4, 7]. For malignancy-related FUO, our findings, characterized by prolonged fever, severe anemia, and thrombocytopenia, are consistent with previous research, likely due to the high prevalence of acute lymphoblastic leukemia involving bone marrow suppression and cytopenia [7, 12].

The overall performance of this prediction model was good, particularly when identifying autoimmune diseases, followed by malignancies. However, a difference in AUC for infections was observed between the validation and development cohorts. This discrepancy may be attributed to the relatively small number of patients with infections in the validation cohort and slight differences in baseline characteristics between the two cohorts. Nevertheless, this prediction model is particularly beneficial for patients with autoimmune diseases and malignancies, as these conditions can be life-threatening, especially in those presenting with prolonged fever. When differentiating these two conditions is challenging, physicians must decide on the appropriate course for further investigation. In such cases, a simultaneous workup for both conditions may be necessary to prevent delays in diagnosis, which could otherwise lead to increased morbidity and mortality.

This prediction model was developed using historical data, physical examination findings, and basic laboratory investigations, such as CBC. Other laboratory markers, such as uric acid and lactate dehydrogenase for malignancies or procalcitonin and blood cultures for infections, were not included due to missing data, as these tests were performed only for specific etiologies. As a result, this prediction model does not provide definitive diagnoses but serves as a clinical tool to support informed decision-making, which will guide physicians in determining the need for further investigations to identify specific etiologies.

The limitations of this study should be noted. First, this was a single-center study conducted in a tertiary care hospital, which may limit generalizability to the broader population of children in other regions. The etiologies of FUO in this study reflect both local epidemiology and our role as a tertiary referral center. For example, tuberculosis is endemic in Thailand; Epstein-Barr virus is frequently detected during thorough FUO workups, and lower respiratory tract infections are common in children. Similarly, SJIA and SLE are among the most prevalent autoimmune diseases in Asian pediatric populations. Acute lymphoblastic leukemia is the most common childhood malignancy and is frequently referred to our institution, whereas solid tumors were underrepresented, likely due to their lower likelihood of presenting with FUO alone. Consequently, the model may be less sensitive for detecting solid malignancies. Second, a retrospective study design in nature caused missing data, particularly inflammatory markers such as C-reactive protein and erythrocyte sedimentation rate, which prevented the inclusion of potentially essential variables in the prediction model. Third, patients without a confirmed final diagnosis were excluded from the model development. Since our primary aim was to construct a classification model that categorizes pediatric FUO cases into three specific etiologic groups, including infections, autoimmune diseases, and malignancies, patients with an unresolved etiology would introduce classification uncertainty and compromise model validity. In our cohort, only two patients remained undiagnosed, both of whom experienced spontaneous resolution of fever, likely due to an unspecified viral illness. These cases were not included in the infectious group due to a lack of confirmed etiology. The low number of undiagnosed cases may reflect the comprehensive diagnostic capabilities available at our tertiary care facility. However, we recognize that in routine clinical settings, a proportion of FUO cases remain undiagnosed. Therefore, future multi-center studies with a prospective study design, including undiagnosed patients, are needed to perform external validation and refine this prediction model.

Despite these limitations, this study’s findings have significant implications for clinical practice. The prediction model developed here can become a complementary tool to guide clinicians toward informed, data-driven decision-making, which will facilitate earlier diagnoses and reduce unnecessary testing and treatments. However, this model is not intended to replace clinical judgment or the diagnostic reasoning of physicians. Instead, it serves as a supportive tool by estimating the probability of various diagnostic possibilities based on clinical and laboratory data. The final diagnosis and management should always be grounded in comprehensive clinical evaluation and physician expertise.

In conclusion, key clinical parameters, including duration of fever, cough, splenomegaly, arthritis, lymphadenopathy, hemoglobin, and platelet levels, were identified as significant predictors of the etiologies of FUO in children. These factors were integrated into a prediction model designed to guide clinicians in differentiating the underlying causes of pediatric FUO, encompassing infections, autoimmune diseases, and malignancies.

The prediction model has demonstrated promise as a practical clinical tool by aiding early diagnoses and supporting evidence-based decision-making in managing pediatric FUO. This model might reduce diagnostic delays, minimize unnecessary testing, and improve patient outcomes by providing a systematic approach to interpreting clinical and basic laboratory data. Future multi-center studies are warranted to validate and enhance the applicability of the model across diverse pediatric populations.