(P13-012-23) Machine Learning Analysis for Prediction of Breast Cancer Risk From Diet, Life Style and Demographic Factors

Objectives: Data on the association between diet, life style and demographic factors with odds of breast cancer (BC) are conflicting. Machine learning analysis can predict BC risk with high accuracy based on features hidden in data. 

This study aimed to predict BC risk using different machine-learning approaches applying diet, life style and demographic data.

Methods: In this population-based case-control study, 350 newly diagnosed cases of BC were recruited and 700 age-matched control subjects were enrolled. Dietary intakes were examined by using a 106-item dish-based validated FFQ. This study has explored risk factors for predicting BC by using data mining techniques. Prior to training algorithms, feature selection and cross validation were essential to avoid “overfitting” the data. There were 64 attributes including nutrients, life style and demographic factors. ASC-J48DT, IKB and Naïve bias were applied on the data and performance metrics of all algorithms were presented.

Results: The main aim of this research was to analyze the algorithms performance for BC data based on the numerous input parameters. The average performances of these algorithms were measured. IKB, Naïve Bayes and ASC- J48DT algorithms belong with 10-fold cross validation have given maximum correctly identifying BC free subjects. Among them, ASC- J48DT was also the best for correctly identifying subjects with BC. ASC-J48DT achieved the sensitivity of 83.19 %, specificity of 90.28% and accuracy of 88.52%. According to ASC- J48DT including all data points (n=1050), age was the first rank parameter and the most important determinant of having BC. ASC- J48DT could evaluate age feature in individuals ≥ 75 years old, which were fallowed with dietary iron consideration. Dietary intake of total trance fat was the second important predictor for being in BC group. It is noteworthy mentioning that iron, red and process meat and fruit juice were among the strongly influencing features for prediction of BC using ASC- J48DT model. It identifies who are likely to be high risk or low risk for BC in accordance with other features.

Conclusions: We used some data mining algorithms for analyzing our dataset and prediction a model. These promising results may indicate that ASC-J48DT has the potential to help clinicians and also is beneficial to nutritionist for breast cancer prediction and prevention.

Funding Sources: No funding