Document Type : Original Paper
Ph.D. Student in Biomedical engineering, Tarbiat Modarres University, Tehran, Iran
Assistant Professor, Biomedical Engineering Group, Tarbiat Modarres University, Tehran, Iran
Associate Professor, Faculty of Biomedical Engineering, Amir Kabir University of Technology, Tehran, Iran
Introduction: Amajor problem in the treatment of cancer is the lack of an appropriate method for the early diagnosis of the disease. The chemical reaction within an organ may be reflected in the form of proteomic patterns in the serum, sputum, or urine. Laser mass spectrometry is a valuable tool for extracting the proteomic patterns from biological samples. A major challenge in extracting such patterns is the optimum selection of feature subset from mass spectrum data.
Materials and Methods: In this research, the data corresponding to proteomic patterns of serum from patients with ovarian cancer was analyzed in two independent groups. Using a mathematical model, the baseline and electrical noises were eliminated in the preprocessing stage with subsequent normalization of mass spectra. The proposed method uses a hybrid algorithm based on a statistical test and Bhattacharyya distance measure. Using the final prediction error criteria, the best feature subset was selected from 15154 data points while maintaining the resolution and the valuable information. The selected feature subset was then used for the detection of biomarkers within the mass spectrum.
Results: Using the method of k-fold cross validation, the samples under study were divided into two sets called the learning and test. Using the least threshold value, the points having significance difference (p-value < 0.05) were selected. The best subset was then extracted from the remaining points such that it would have the maximum information content. By doing so, the number of input variables was reduced from 15154 to 80 points. In the next step, 16 and 6 biomarkers were selected for the two independent dataset. The obtained results show accuracy, specificity as well as sensitivity of 100%.
Discussion and Conclusion: To diagnose a disease in medicine is an example of pattern recognition in engineering and physical science. In this paper, a filter approach is introduced for feature subset selection which extracts appropriate features in the input space by using the combination of statistical method and distance measure based on information criteria. The result of this study emphasizes that the use of combination approach in feature extraction and selection in high dimensional data can appropriately separate the pattern classes in addition to maintaining the information content.