Analyses of Lifestyle and Environmental Factors for Cancer Prevention using Deep Learning and Conventional Machine Learning from UK Biobank Data

Abstract

AbstractCancer may be a fatal illness which brings patients extreme pain and cause death. Cancer is the second leading reason behind death globally. There has been substantial research showing that cancer may be prevented by changing healthy environmental factors. Previous statistical studies investigated each of the risk factors separately, and so it failed to test multiple factors in tandem. The studies which did examine multiple risk factors simultaneously exhibited insufficient sample sizes for the large number of hypothesis tests made therein. With the emergence of enormous biobanks like the United Kingdom Biobank (UK Biobank) in recent years, it’s possible to unveil this relationship between environmental factors and cancer occurrence from such big data using cutting-edge statistical and machine learning integration techniques. Deep learning processes data with multi-layer neural networks increases the predictive power. Deep learning techniques have made substantial advances in many domains including healthcare. Clinical application of deep learning has been most rapid in image-intensive fields. However, there wasn’t much research reporting applications of deep learning in analyzing environmental factors for cancer prevention. To explore risk factors among environmental factors associated with cancer incidence for cancer prevention, the data of environmental factors was extracted from the UK Biobank, and two deep learning techniques including deep neural networks (DNNs) and convolutional neural networks (CNNs) were applied to analyze the data. Meanwhile, conventional machine learning techniques including random forest, support vector machine, and logistic regression model were also applied for comparisons. All the accuracies and sensitivities were over 0.92 and 0.90 respectively from the analyses of deep neural networks, convolutional neural networks, support vector machine (SVM), random forest, and logistic regression. Overall, CNNs had the most effective prediction (sensitivity: 0.933; F1 score: 0.961). DNNs had the second-best prediction (sensitivity: 0.933; F1 score: 0.956). Eighty-four important features were selected by machine learning techniques and were analyzed in logistic regression with the lasso penalty model to further reduce the number of features. Twenty-seven important features were further screened and obtained from the logistic regression model. After applying these 27 features selected to logistic regression, the results showed that sensitivity was 0.900, F1 Score was 0.919, and the area under the curve was 0.974. “Age”, “Number of operations, self-reported”, “Other serious medical condition/disability diagnosed by doctor (Yes)”, “Long-standing illness, disability or infirmity (Yes)” were significantly associated with cancer risk (p-values<0.01 or 0.05; odds ratios: 8.79, 1.76, 1.54, 1.33, respectively). Females were significantly more likely to get cancer than males (p-value <0.01, odds ratio: 1.502). “Number of operations, self-reported” was found to have significant associations with the risk of cancer. However, the factor of “0 self-reported non-cancer illness” was a risk factor for cancer, while the factor of “1 self-reported non-cancer illness” was a protective factor for cancer, which is very interesting for further research and discussion with physiologists with respect to mechanisms. The factors of “Sleep duration of 6 hours”, “Water intake (2 glasses of water daily)”, “Overall health rating (Excellent)” and “Overall health rating (Good)” were found as protective factors for cancer (p-values: <0.01, 0.01, 0.06, 0.08 (close to 0.05); odds ratios: 0.834, 0.878, 0.534, 0.562, respectively). The factor of “Never/rarely sleeplessness /insomnia” was also helpful for cancer prevention (p-value: 0.196; odds ratio: 0.931). The findings in this study will be helpful for initiating early cancer screening and educating the general public about risk factors and protective factors among lifestyle- environmental factors in high risk populations for cancer prevention.