Population-Based Screening for Endometrial Cancer: Human vs. Machine Intelligence

Abstract

Incidence and mortality rates of endometrial cancer are increasing, leading to increased interest in endometrial cancer risk prediction and stratification to help in screening and prevention. Previous risk models have had moderate success with the area under the curve (AUC) ranging from 0.68 to 0.77. Here we demonstrate a population-based machine learning model for endometrial cancer screening that achieves a testing AUC of 0.96. We train seven machine learning algorithms based solely on personal health data, without any genomic, imaging, biomarkers, or invasive procedures. The data come from the Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial (PLCO). We further compare our machine learning model with 15 gynecologic oncologists and primary care physicians in the stratification of endometrial cancer risk for 100 women. We find a random forest model that achieves a testing AUC of 0.96 and a neural network model that achieves a testing AUC of 0.91. We test both models in risk stratification against 15 practicing physicians. Our random forest model is 2.5 times better at identifying above-average risk women with a 2-fold reduction in the false positive rate. Our neural network model is 2 times better at identifying above-average risk women with a 3-fold reduction in the false positive rate. Our machine learning models provide a non-invasive and cost-effective way to identify high-risk sub-populations who may benefit from early screening of endometrial cancer, prior to disease onset. Through statistical biopsy of personal health data, we have identified a new and effective approach for early cancer detection and prevention for individual patients.

Keywords: cancer screening; early detection; endometrial cancer; machine learning; statistical biopsy.

Conflict of interest statement

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Copyright © 2020 Hart, Yan, Huang, Liang, Nartowt, Muhammad and Deng.

Figures

FIGURE 1

A) The sensitivity and specificity of the random forest for both the training and testing data as a function of the threshold value and (B) The sensitivity and specificity of the neural network for both the training and testing data as a function of the threshold value.

FIGURE 2

A) Area under the ROC curve for the random forest on both the training and testing data. Similar performance on both datasets indicates that the random forest has no overfit and (B) Area under the ROC curve for the neural network on both the training and testing data. Similar performance on both datasets indicates that the neural network has no overfit.

FIGURE 3

(A) Kaplan-Meier plot of the below- (green), at- (yellow), and above- (red) average risk groups created from the testing data by our random forest model. Also shown are the p-value and hazard ratio (HR) between each group. Those in the above-average risk group clearly have the highest chance of developing cancer and (B) Kaplan-Meier plot of the below- (green), at- (yellow), and above- (red) average risk groups created from the testing data by our neural network model with 95% confidence intervals (shaded). Also shown are the p-value and hazard ratio (HR) between each group. Those in the above-average risk group clearly have the highest chance of developing cancer.