Random Forest for Predicting Treatment Response to Radioiodine and Thyrotropin Suppression Therapy in Patients With Differentiated Thyroid Cancer But Without Structural Disease.

BACKGROUND: We aimed to develop a machine-learning model for predicting treatment response to radioiodine (131I) therapy and thyrotropin (TSH) suppression therapy in patients with differentiated thyroid cancer (DTC) but without structural disease, based on pre-treatment information. PATIENTS AND METHODS: Overall, 597 and 326 patients with DTC but without structural disease were randomly assigned to "training" cohorts for predicting treatment response to 131I therapy and TSH suppression therapy, respectively. Six supervised algorithms, including Logistic Regression, Support Vector Machine, Random Forest (RF), Neural Networks, Adaptive Boosting, and Gradient Boost, were used to predict effective response (ER) to 131I therapy and biochemical remission (BR) to TSH suppression therapy. RESULTS: Stimulated and suppressed thyroglobulin (Tg) and radioiodine uptake before the current course of 131I therapy were mostly attributed to ER to 131I therapy, while thyroid remnant available on the post-therapeutic whole-body scan at the last course of 131I therapy and TSH were greatly contributed to Tg decline under TSH suppression therapy. RF showed the best performance among all models. The accuracy and area under the receiver operating characteristic curve (AUC) for segregating ER from non-ER during 131I therapy with RF were 81.3% and 0.896, respectively. The accuracy and AUC for predicting BR to TSH suppression therapy with RF were 78.7% and 0.857, respectively. CONCLUSION: This study demonstrates that machine learning models, especially the RF algorithm are useful tools that may predict treatment response to 131I therapy and TSH suppression therapy in DTC patients without structural disease based on pre-treatment routine clinical variables and biochemical markers.

Predicting T-Cell Lymphoma in Children From 18F-FDG PET-CT Imaging With Multiple Machine Learning Models.

This study aimed to examine the feasibility of utilizing radiomics models derived from 18F-FDG PET/CT imaging to screen for T-cell lymphoma in children with lymphoma. All patients had undergone 18F-FDG PET/CT scans. Lesions were extracted from PET/CT and randomly divided into training and validation sets. Two different types of models were constructed as follows: features that are extracted from standardized uptake values (SUV)-associated parameters, and CT images were used to build SUV/CT-based model. Features that are derived from PET and CT images were used to build PET/CT-based model. Logistic regression (LR), linear support vector machine, support vector machine with the radial basis function kernel, neural networks, and adaptive boosting were performed as classifiers in each model. In the training sets, 77 patients, and 247 lesions were selected for building the models. In the validation sets, PET/CT-based model demonstrated better performance than that of SUV/CT-based model in the prediction of T-cell lymphoma. LR showed highest accuracy with 0.779 [0.697, 0.860], area under the receiver operating characteristic curve (AUC) with 0.863 [0.762, 0.963], and preferable goodness-of-fit in PET/CT-based model at the patient level. LR also showed best performance with accuracy of 0.838 [0.741, 0.936], AUC of 0.907 [0.839, 0.976], and preferable goodness-of-fit in PET/CT-based model at the lesion level. 18F-FDG PET/CT-based radiomics models with different machine learning classifiers were able to screen T-cell lymphoma in children with high accuracy, AUC, and preferable goodness-of-fit, providing incremental value compared with SUV-associated features.