Lung Cancer Prediction Using Curriculum Learning Based Deep Neural Networks

Abstract

The high incidence and low survival rate of lung cancers contribute to their high death count, and drive the development of lung cancer prediction models using demographic factors. The five year relative survival rate of small cell lung cancer in particular (6%) is four times less than that of non small cell lung cancer (23%), though no predictive models have been developed for it so far. This study aimed to expand on previous lung cancer prediction studies and develop improved models for general and small cell lung cancer prediction. Established machine learning models were considered, in addition to a novel curriculum learning based deep neural network. All models were evaluated using data from the National Cancer Institute's Prostate, Lung, Colorectal and Ovarian Cancer screening trial, with performance measured using the area under the receiver operator characteristic curve (AUROC). Random forest models were found to give the best performances in lung cancer prediction (bootstrap optimism corrected (BOC) $\text{AUROC}\ = {0.927}$), outperforming previous logistic regression models $(\text{BOC} \text{AUROC} ={0.859})$. Additionally, curriculum learning based neural networks were shown to outperform all other model types for small cell lung cancer prediction in particular (AUROCs of 0.873 and 0.882 across two feature sets). To conclude, high-performance models were developed for general and small cell lung cancer prediction, and could help improve non-invasive lung cancer prediction in a clinical setting.