Validation of Risk Models for Predicting Febrile Neutropenia Among Breast Cancer Patients Receiving Chemotherapy: A Real-World Study

Background: Breast cancer patients receiving chemotherapy may develop a serious complication called febrile neutropenia (FN). We aimed to validate and compare three existing FN prediction models for breast cancer patients receiving chemotherapy in Taiwan.

Patients and methods: This was a retrospective observational real-world study. Data were acquired from the clinical research databases of three study hospitals. Breast cancer patients who have received at least one antineoplastic chemotherapy drug were chosen for the analysis. For evaluating the occurrence of FN, we used both broad (a body temperature above 38°C with an absolute neutrophil count (ANC) below 0.5 × 10⁹/L or a body temperature above 38°C with a diagnosis of neutropenia) and narrow definitions (having both fever and neutropenia diagnoses or having both neutropenia and infection diagnoses). Sensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV) were calculated for each selected FN model.

Results: Among the 1903 patients identified, when the broad and narrow definitions of FN were applied, 70 (3.7%) and 60 (3.2%) patients developed FN in the first cycle, respectively. Using the broad FN definition, Aagaard's model was the highest in sensitivity (90.0%), followed by Chantharakhit's (40.0%) and Chen's (7.2%); in specificity, Chen's (93.6%) was the highest. In addition, the accuracy was highest with the Chen model (90.4%). All three models' PPVs were low, ranging from 0.5% to 4.2%, but all three models' NPVs were over 96.3%. When the narrow FN definition was used, Chantharakhit's model showed a relatively high improvement in sensitivity (53.3%) and PPV (3.9%) while negligible increases or even slight decreases were seen in the other two models and in the other performance indicators of Chantharakhit's model.

Conclusion: The results of this study provide important information for clinicians when selecting models to identify patients at high-risk of FN. As the model performance observed was less than satisfactory, improving the prediction ability of the models is needed.