Early prediction of lung cancer based on the combination of trace element analysis in urine and an Adaboost algorithm

Early detection of cancer is the key to effective treatment and long-term survival. Lung cancer is one of the most frequently occurring cancers and its early detection is particularly of interest. This work investigates the feasibility of a combination of Adaboost (ensemble from machining learning) using decision stumps as weak classifier and trace element analysis for predicting early lung cancer. A dataset involving the determination of 9 trace elements of 122 urine samples is used for illustration. Kennard and Stone (KS) algorithm coupled with an alternate re-sampling was used to realize sample set partitioning. The whole dataset was split into equally sized training and test set, which were then reversed to yield a second operating case, we called them case A and case B, respectively. The prediction results based on the Adaboost were compared with those from Fisher discriminant analysis (FDA). On the test set, the final Adaboost classifiers achieved a sensitivity of 100% for both cases, a specificity of 93.8%, 95.7%, and an overall accuracy of 95.1%, 96.7%, for case A and case B, respectively. In either case, Adaboost always achieves better performance than FDA; also, it is less sensitive to the composition of the training set compared to FDA and easy to control over-fitting. It seems that Adaboost is superior to FDA in the present task, indicating that integrating Adaboost and trace element analysis of urine can serve as a useful tool for diagnosing early lung cancer in clinical practice.