Abstract

The incidence and mortality of cancer are increasing at an accelerating rate due to a lack of awareness of the basic risk factors and how spotting these factors early can make a real difference. The interaction of genes, lifestyle, chronic diseases (comorbidities), and environment can influence a person’s risk of cancer. Hence, early detection of the risk factors through the application of machine learning will enhance personalized predictive healthcare, and significantly reduce the incidence and mortality of cancer.

A quantitative analysis was carried out based on the primary dataset of 580 records with 58 features obtained through a survey questionnaire from participants (aged 18 and above). Clustering analysis was deployed using KMeans and KModes to extract knowledge, and patterns, which are subsequently grouped into two class labels: Low and High cancer risk. The dataset was divided into training (n=371), validation (n=93) and testing (n=116) for building, training, and testing the model respectively. These techniques adopt the 80/20 train-test split. The data were used to construct six machine learning models including decision tree (DT), random forest (RF), logistic regression (LR), support vector machine (SVM), Naive Bayes (NB) and K-Nearest Neighbour (KNN) to predict and classify the risk of cancer in adults. A total of 10 relevant variables were input into these models. The performance of the models was measured using the area under receiver operating characteristic curve (AUC-ROC) and the confusion metrics.

Two cancer risk classes (Low: 0 and High: 1) were predicted. Of the six machine learning models, the Random Forest ensemble learning methods had the best performance in predicting cancer risk (AUC: 0.93) and outperformed conventional logistic regression (AUC: 0.83). While the decision tree, support vector machine, Naive Bayes and KNN had AUC of 0.89, 0.87, 0.80, and 0.80 respectively. With the Random Forest model, the number of people considered as having a high risk of cancer (that is, RECALL, otherwise known as Sensitivity) was 93% with a miss rate (false negative) of 7%. Gender, sugar intake, frequency of exercise, smoke per pack year, level of alcohol intake, comorbid: frequent cold, exposure to industrial pollution, domestic pollution, sun and co2 emissions were the 10 relevant risk factors used by the RF model.

The results of the study showed that baseline information of a person can be used by machine learning to accurately predict and classify the risk level of cancer. Beyond the lifestyle and environmental risk factors, the study included chronic diseases (comorbidities) in a person in building the predictive model which has been overlooked in many previous studies. In the context of this study and the demographic of the participants, risk factors such as gender, alcohol intake, chronic diseases, exposure to the sun, pollution resulting from domestic, co2 emissions, and industrial practices contributed significantly to classifying the cancer risk level for Black African, Asian-Indian, Asian Pakistani, White British and White Irish.

Applications of Machine Learning in Predicting the Risk of Cancer in Adults: A Quantitative Risk Assessment Model