Artificial Intelligence for Predicting Responses to Thyroid Cancer Treatment

Thyroid cancer, one of the most prevalent endocrine malignancies, presents diverse treatment responses among patients, underscoring the necessity for personalized treatment strategies. This study aims to develop and evaluate AI models that can accurately predict the initial treatment response among patients with well-differentiated thyroid cancer. We trained and validated various machine learning and deep learning models using a dataset that comprises demographic and clinicopathological features. This dataset was collected from a retrospective cohort of 383 patients diagnosed with thyroid cancer at a single medical center. Models in all experiments achieved an average accuracy of 72.2%, average precision of 76.2%, average recall of 68%, average F1 score of 70.6%, and average AUC of 80.5%. Multi-layer Perceptron achieved the highest accuracy (85.7%), recall (75%), and F1 score (81.7%) in this study when it was used for predicting binary treatment response using all features, excluding those weakly correlated with the treatment response. However, Naive Bayes attained the highest precision (95.1%) and AUC (88.7%) in this study when it was used for predicting binary treatment response using all features. Models used for predicting binary treatment responses outperformed those used for predicting multi-class treatment responses. AI demonstrates satisfactory performance in predicting the response to thyroid cancer treatment, yet there is room for optimization. Healthcare providers should not solely rely on our models; combining them with other tools is advised until further studies validate their optimal performance. Future research should enhance predictive capabilities by including additional features, exploring alternative models, and utilizing larger and balanced datasets.