Classification of thyriod cancer patients in the UK Biobank database using supervised machine learning methods

Abstract: Thyroid cancer is the most common endocrine malignancy worldwide and has reached epidemic levels in recent years. Thyroid cancer is notoriously difficult to diagnose as it typically becomes detectable as thyroid nodules with only 5-15% of nodules being cancerous. The ubiquity of benign thyroid nodules means that further diagnostic techniques like high-resolution ultrasonography and fine needle aspiration biopsies must be used to diagnose the disease. These techniques have limitations in that they are labour intensive, invasive, dependent on the skill of the clinician, and have accuracies ranging from 70-97%. The widely described problems of thyroid cancer overdiagnosis and overtreatment have frequently been attributed to the limitations of current diagnostic techniques and it is clear there is a need for improved thyroid cancer diagnostic techniques.

The use of machine learning techniques in biomedical research is on the rise and diagnosis by machine learning models is becoming available for an increasing number of conditions. This work looked to utilise three machine learning models using phenotype data from UK Biobank to address the limitations of current diagnostic techniques and classify individuals with thyroid cancer. The random forest, support vector machine, and multilayer perceptron models evaluated were all found to perform poorly on the dataset. These results suggest that for thyroid cancer diagnosis, the phenotype features in the UK Biobank do not contain enough information about the thyroid for the predictive models to succeed.

Data Source

This research was conducted using data from UK Biobank, a major biomedical database: www.ukbiobank.ac.uk.

Specifically, this work used binary disease data for 394,884 individuals and 787 disease as well as phenotype data for 390 phenotypes and 454,119 individuals. The phenotypes used included lifestyle factors, biological samples e.g. Blood samples, socioeconomic factors, family history, physical measures e.g. Height, and genetic principle components.

Reference:

Sudlow, C. et al. (2015) ‘UK Biobank: An Open Access Resource for Identifying the Causes of a Wide Range of Complex Diseases of Middle and Old Age’, PLoS Medicine, 12(3), p. e1001779. Available at: https://doi.org/10.1371/journal.pmed.1001779.

Methodology

The figure below shows the machine learning workflow used for this work. Following data pre-processing and manual feature selection, recursive feature elimination (RFE) was used to select the optimal features for the predictive models. However, models were also tested with the full feature set. Random undersampling was used to reduce the class imbalance and the need for SMOTE oversampling. Hyperparameter tuning was carried out on the full pre-processing and classification pipeline which included:

1. Imputation with either Median or Most Frequent, KNN, or Multiple Imputation (Iterative Imputer).
2. One hot encoding and normalisation
3. Tomek Links undersampling
4. SMOTE oversampling
5. Model Training with either Support Vector Machine, Random Forest, or Multilayer Perceptron

After the optimal hyperparameters were found, each model was evaluated on the holdout test data and SHAP analysis was conducted on the best performing model.

This research utilised Queen Mary's Apocrita HPC facility, supported by QMUL Research-IT. http://doi.org/10.5281/zenodo.438045

Acknowledgements

I would like to thank Dr Eirini Marouli for her help and suppor throughout the duration of this project.