Machine Learning Predicts Postembolization Fever After TACE

An ensemble machine learning approach demonstrated good clinical utility for predicting postembolization fever (PEF) after transarterial chemoembolization (TACE), achieving an area under the curve (AUC) of 81.5% with acceptable classification performance, according to results from a study published in Computers Informatics Nursing.

These results support this approach’s potential role in anticipatory care and risk stratification for oncology clinicians. In this retrospective cohort of 1495 patients undergoing TACE at a tertiary center, PEF occurred in 13.6% of cases, underscoring the relevance of reliable prediction in routine practice.

Seven machine learning-based algorithms were developed using SPSS WIN 27.0 and Python, with the ensemble method outperforming individual models in discrimination. Logistic regression achieved the highest classification accuracy at 87.5%. 

According to established benchmarks, an AUC exceeding 75% represents good predictive value, placing the proposed model within a clinically meaningful range. Although direct comparisons are limited by the scarcity of machine learning-based PEF models, performance was comparable to or exceeded that of prediction tools for other treatment-related adverse effects, such as chemotherapy-induced nausea and vomiting.

Model explainability using SHAP values highlighted several laboratory and clinical contributors to PEF. Positively correlated variables included post-TACE aspartate aminotransferase (AST), alanine aminotransferase (ALT), C-reactive protein, bilirubin, international normalized ratio (INR), and platelet count, as well as pre-TACE AST, alpha-fetoprotein, and platelet count. These findings align with prior evidence linking hepatocellular injury and inflammatory activation to fever following embolization, likely reflecting tumor necrosis and ischemic insult.

Treatment-related factors also contributed meaningfully. Larger amounts of lipiodol and doxorubicin and the presence of tumors larger than 5 cm were associated with higher PEF risk. Greater embolization volume and chemotherapeutic dose likely intensify ischemia and inflammatory responses, while larger tumor burden has been consistently associated with postprocedural symptoms. All patients in this cohort had multiple tumors, reinforcing the relevance of cumulative treatment exposure.

Conversely, higher lymphocyte and monocyte counts and higher albumin levels before and after TACE were negatively correlated with PEF. Albumin appeared protective, consistent with its role as a marker of hepatic synthetic function, whereas lymphocyte and monocyte findings may reflect the balance between antitumor immunity and inflammatory response. Radiologic features also contributed: LR-4 lesions, classified as probably hepatocellular carcinoma, were inversely associated with fever, a novel observation that warrants replication.

Overall, PEF reflected treatment response rather than infection and was not associated with survival. Monitoring laboratory trends, chemoembolic dose, tumor characteristics, and LI-RADS classification before and after TACE may help clinicians anticipate fever, optimize supportive management, and improve patient well-being without unnecessary antibiotic use.

“These findings underscore the need for comprehensive monitoring of laboratory, clinical, and radiologic variables to enhance patient well-being following TACE,” stated the authors. “Future research should focus on implementing a calibration process to enhance the utility of predictive analytics for supporting clinical decision-making in managing PEF.”