Possibility of predicting the probability of thyroid cancer recurrence by machine learning methods

Objective. To develop a machine learning model for predicting the fact of recurrence in patients with thyroid cancer after surgical intervention.

Materials and Methods. According to the aim of the study, the case histories of 300 patients who had undergone surgical intervention for thyroid cancer were analyzed. The average age was 43.54 years. All patients included in the study underwent a comprehensive examination according to the clinical recommendations on the diagnosis and treatment of patients with thyroid cancer. Selection of the most appropriate model in machine learning is critical as it directly affects the accuracy and efficiency of prediction. Selection of the best model was done through comparing the performance of different algorithms on the same training sample using cross-validation. Each model was evaluated on such metrics as average accuracy and standard deviation to determine which model demonstrates the best results.  The random forest model performed best in terms of average accuracy and was used hereafter. The model was trained using a matrix of predefined features. Using param grid, we can efficiently adjust hyperparameters such as the number of trees, maximum depth and minimum number of samples for separation, which will help us to find the optimal settings for our task. RandomizedSearchCV method was used to select the hyperparameters. During the hyperparameter search process, the model was trained on training data selected as 70% of the original dataset. The search resulted in the following best hyperparameters for the random forest model for our data specifically: n_estimators = 161; min_samples_split = 5; max_leaf_nodes = 39; max_depth = 12; bootstrap = True.

 Results. A model that demonstrated high target feature accuracy was trained during the study. The proportion of patients with postoperative recurrence correctly identified by the model was 98 % of all patients with recurrence, and the proportion of patients without recurrence correctly classified by the model «as patients at no risk of recurrence» was 95 % of all patients without recurrence. This shows that the developed model effectively handles the task of classification based on medical parameters, which may be particularly important for decision making in clinical practice. The high accuracy indicates the reliability of the model and its ability to identify cases of recurrence correctly, this may contribute to the improvement of diagnostics and treatment.

Conclusions. A machine learning model to predict a high probability of thyroid cancer recurrence based on the analysis of medical parameters was developed while carrying out the study. The development process began with careful data preprocessing, which is a critical step in reliable models’ construction. During preprocessing, outliers and columns containing monotonic values were removed to improve the data quality and avoid distortions in the model training. Categorical variables were also coded to ensure that they could be used correctly in machine learning algorithms, and correlated features were excluded to minimize multicollinearity and increase the interpretability of the model.