Machine learning approach to determine the diagnostic value and predictive factors of PET/CT in FUO and IUO patients

TL;DR

This study uses machine learning to predict which patients with unexplained fever or inflammation will benefit most from PET/CT scans.

Contribution

The study introduces machine learning models to guide PET/CT use in diagnosing complex cases of fever and inflammation.

Findings

Machine learning models like MLP and Logistic Regression effectively predicted PET/CT usefulness in FUO/IUO patients.

Lower procalcitonin and higher lymphocyte levels were linked to increased PET/CT benefit according to model analysis.

The study identified clinical factors that could help tailor diagnostic approaches for unexplained fever and inflammation.

Abstract

This study aimed to identify the clinical and laboratory variables determining the true diagnostic contribution of FDG-PET/CT in patients with fever of unknown origin (FUO) or inflammation of unknown origin (IUO), and to develop machine learning models capable of predicting which patients are most likely to benefit from PET/CT imaging.

A retrospective cohort of patients aged over 18 years who underwent FDG-PET/CT for FUO/IUO evaluation was analyzed. Machine learning algorithms—including Extreme Gradient Boosting (XGBoost), linear and radial basis function Support Vector Machines, Multilayer Perceptron (MLP), k-Nearest Neighbors (KNN), Random Forest, Decision Tree, Logistic Regression (LR), and Naïve Bayes (NB)—were trained to predict true positive and true negative PET/CT results. Feature selection was performed using the PowerSHAP method. Model performance was compared using area under the precision–recall curve (PR-AUC), Receiver Operating Characteristic – Area Under the Curve (ROC-AUC), accuracy, precision, recall, and F1-score metrics.

A total of 273 patients (151 men, 55.3%; mean age: 59 ± 16.9 years) were included. PET/CT provided diagnostic benefit in 203 patients (74.4%). All algorithms performed well in terms of PR-AUC (>0.79), with the highest scores achieved by MLP and XGBoost, reaching PR-AUC values of 0.86 and 0.85, respectively. All algorithms except NB demonstrated good accuracy. When both PR-AUC and accuracy were considered together, the best-performing models were MLP and Logistic Regression. LR achieved accuracy, ROC-AUC, PR-AUC, precision, recall, and F1-score values of 0.75, 0.74, 0.84, 0.85, 0.86, and 0.83, respectively, whereas MLP achieved 0.73, 0.73, 0.86, 0.85, 1.00, and 0.85. The PowerSHAP analysis suggested that lower procalcitonin and erythrocyte sedimentation rate levels, longer symptom duration, older age, generalized body pain, inpatient evaluation, and higher lymphocyte counts were associated with increased model-predicted PET/CT usefulness.

Machine learning models—particularly MLP and LR—may have potential to assist in identifying FUO/IUO patients who could benefit from PET/CT imaging. The clinical and biochemical predictors highlighted in this study might help guide PET/CT use and support more tailored diagnostic approaches in complex FUO and IUO cases, though further validation is needed.

Figures