The Potential Of Machine Learning Models in Predicting Recurrence Risk in Nasopharyngeal Carcinoma

Creative Commons License

Aksoy A., Doğan Karataş T.

Master of Head and Neck Surgery IFHNOS 2025, İstanbul, Türkiye, 13 - 15 Kasım 2025, (Özet Bildiri)

  • Yayın Türü: Bildiri / Özet Bildiri
  • Basıldığı Şehir: İstanbul
  • Basıldığı Ülke: Türkiye
  • Açık Arşiv Koleksiyonu: AVESİS Açık Erişim Koleksiyonu
  • Sivas Cumhuriyet Üniversitesi Adresli: Evet

Özet

Aim

The aim of this study is to predict the risk of post-treatment recurrence in patients with nasopharyngeal carcinoma treated at our center using machine learning models.

Material-methods

This retrospective study included 40 nasopharyngeal carcinoma patients diagnosed, treated, and followed at Sivas Cumhuriyet University Faculty of Medicine between 2014 and 2024. Demographic, clinical, molecular, laboratory, and survival data were obtained from hospital records. Machine learning analyses were performed in Python (v2.3) using the PyCaret library with Z-score normalization and SMOTE for preprocessing. The dataset was randomly divided into training and testing sets, and dimensionality reduction was applied. Random Forest, AdaBoost, and correlation-based models were tested with a 0.90 feature selection threshold, and performance was evaluated by AUC, accuracy, sensitivity, specificity, and F1 score.

Results

The median age of the 40 patients was 46 years (18–75); 80% were male. Comorbidities were present in 25%, smoking in 50%, alcohol use in 5%, and family history in 25%. Stage distribution was 27.5% stage II, 52.5% stage III, and 20% stage IV, with distant metastasis in 12.5%. Chemoradiotherapy was administered to 95%. EBV status was positive in 30%, negative in 25%, and unknown in 45%. During a median follow-up of 56.7 months, recurrence occurred in 50%. Median values were neutrophils 4.7, lymphocytes 1.1, monocytes 0.5, platelets 238, and hemoglobin 13.1. The Gradient Boosting model showed moderate predictive ability (accuracy 66.7%, ROC-AUC 0.74, F1-score 0.71). Hemoglobin was the strongest predictor of recurrence, followed by neutrophil count, age, lymphocyte count, metastasis, and stage, while other clinical factors had limited impact.

Conclusion

Machine learning models showed moderate effectiveness in predicting recurrence in nasopharyngeal carcinoma. Hemoglobin and inflammatory markers were identified as principal predictors, underscoring their potential utility for risk stratification and prognostic assessment.

Keywords: Nasopharyngeal carcinoma, machine learning, recurrence, prognosis