Journal of Carcinogenesis & Mutagenesis

Sampling Statistical Errors in Big Data Research: 3 Cases of Breast Cancer Research

Han-Jun Cho^1* and Eui Seok Jeong^{2 1}Department of Biomedical Institute for Convergence at SKKU, Suwon, South Korea
²Department of Ecological Science, Kyungpook National University, Sangju, South Korea
^*Correspondence: Han-Jun Cho, Department of Biomedical Institute for Convergence at SKKU, Suwon, South Korea, Tel: 82-10-8574-8358, Email:

Received: 03-Nov-2021 Published: 24-Nov-2021, DOI: 10.35248/2157-2518.21.s20.001

Introduction

Breast Cancer (BRCA) is one of the most common cancers found in women. Also, according to the results reported in the National Cancer Center for 2021, one in four people dead from breast cancer [1]. Recently on the according to the results of the research, there has been great progress in the treatment technology of breast cancer. These methods are breast cancer research using big data. In addition, with the convenience and economy of the national health insurance system, as women's interest in breast cancer and health increases, more patients come to the hospital at an early stage and it is possible to detect it early [2]. Big data refers to the act of making data into valuable information with a specific technology or analysis tool while having the characteristics of high physical quantity and diversity of data [3]. In addition, big data analysis in the medical industry is becoming important due to the increase in medical data due to the development of the use of big data in the medical service development trend. According to IBM, 16,000 hospitals worldwide are collecting patient data, with 86,400 data being generated per patient per day [4]. In such an environment, in the case of breast cancer, which has a lower recurrence rate the earlier it is detected, it is a target disease model that can build the most effective precision medicine system in big data research [5]. Also, the use of big data health care is expected to have significant effects in cancer patient health tracking, remote patient monitoring, cost reduction and reduction of misdiagnosis rates at medical institutions, and precision medicine [6]. In this study, we report the results of analysis of recurrence characteristics using Machine Learning (ML) and analysis of usage behavior using data provided by the The Cancer Genome Atlas (TCGA) in USA and, Health Insurance Review & Assessment Service in Korea [7].

Materials and Methods

Mutation gene big data analysis using machine learning

The Cancer Genome Atlas-BRCA provided data for 652 BRCA patients with somatic non-silent mutations and clinical information. They divided into two Disease Free/Recurred groups. To identify recurrence-related mutations, four feature selection methods (Information Gain, Chi-squared test, MRMR, Correlation) and four classifiers (Naïve Bayes, K-NN, SVM, Correlation) were used [8]. We performed 5 fold-validations to find out the efficient algorithm.

Network analysis of hospital use behavior in breast cancer patients

The network analysis for medical utilization was conducted using Cytoscape version 3.7.2. Dataset: Health Insurance Review & Assessment Service total patient sample (HIRA-NPS-2016, 2017, HIRA-APS-2016, 2017) [9].

Results

Case 1: A study using machine learning and mutated genes

According to Kaplan Meyer statistics, the recurrence and survival prediction rates that would be expressed in all 7 specific mutant genes were rather closely related to the survival rate. In 40 vs 7, rather than using 7 genes, the predicted value is higher. The ACSF3 gene encodes a member of the acyl-CoA synthetase family that activates fatty acids by catalyzing the formation of thioester bonds between fatty acids and coenzyme A [10]. The ARID3B gene encodes a member of the ARID (AT-rich interaction domain) family of DNA binding proteins [11]. The KHSRP gene encodes a multifunctional RNA-binding protein involved in a variety of cellular processes including transcription, alternative pre-mRNA splicing, and mRNA localization [12]. The LUZP2 gene encodes a Leucine Zipper Protein. This protein is deleted in some patients with Wilms' tumor Aniridia Genitourial ornormal-mental Retardation (WAGR) syndrome. Alternate splicing results in multiple script variants [13]. The RPL18A gene encodes a member of the L18AE family of ribosomal proteins, which is a component of the 60S subunit [14]. The TPI1 gene encodes an enzyme composed of two identical proteins that catalyzes the isomerization of glyceraldehyde 3-phosphate (G3P) and dihydroxy-acetone phosphate (DHAP) in glycolysis and gluconeogenesis [15]. Von Willebrand Factor A Domain Containing 5B2 (VWA5B2) is a protein-coding gene. An important paralog of this gene is VWA5B1 [16]. It is inferred that the commonality of the genes is related to the process of fat synthesis. By inferring that most of the components of breasts in the human body are lipids, this is a possible decision. When 40 genes were used, the OS (overall survival) P-value: 0.0521 and DFS (Disease free survival) P-value: 0.107 as shown in Figures 1A and 1B P-value came out, but unlike this, all 7 genes as shown in Figures 1C and 1D that P-values were all significant.

Figure 1:Total 40 vs. 7 Recurrence of Kaplan-meier specific from Recurrence-related genes value. Among the 40 genes extracted by machine learning, 7 genes (ACSF3, ARID3B, KHSRP, LUZP2, RPL18A, TPI1, VWA5B2)(supplementary Figures 1-8) highly related to breast cancer patients(supplementary Tables 2-1, 2-2). According to Kaplan Meyer statistics, the recurrence and survival prediction rates that would be expressed in all 7 specific mutant genes were rather closely related to the survival rate. In 40 vs. 7, rather than using 7 genes, the predicted value is higher.

Usually, when a gene is used as a biomarker, a small number of genes are preferred, and the characteristic of a mutant biomarker is an objective marker that can distinguish the normal or pathological condition of a target disease and predicts the treatment response. In 652 patients, when viewed with a ratio of 589 (Disease Free): 63 (Recurred/Progressed) (supplementary Table 1), a sample statistical error exists, but as shown in Figures 2A and 2B, three genes are strong biomarker candidates. However, when looking at Figures 2C and 2D only 6 patients overlapped.

Figure 2: Kaplan-meier rates(OS and DFS) for 3 survival-specific genes from 4 feature selection methods.The Kaplan Meyer curve, which measured the expression rate of mutations in breast cancer patients with 3 (KHSRP, LUZP2, VWA5B2) genes, showed a very high predictive rate. This indicates that it can be easier, and only three genes can predict the stage of breast cancer patients.

Case 2: A study on how to apply diagnosis and algorithm using mutation feature selection in machine learning

optimal algorithm combination was Information gain-Naïve bayes, and when diagnosis using 22-42 mutation-specific genes out of 40 genes, breast cancer can be detected early with an 88.79% probability. This problem arises because of the low proportion of relapsed patients and many non-recurring patients among all breast cancer patients in the data provided by the TCGA. When statistical errors were minimized, 144 genes were used as an appropriate number when re-experimented (Table 3). The best algorithm model to be used for breast cancer diagnosis was found, but there was no significant difference from the previously reported results, and the number of genes increased as the number of genes increased. As a result of re-experiment, it was found that the number of genes using about 144 genes was rather high. In addition, as more than 500 genes were used, the diagnosis rate tended to decrease.

Case 3: A study on analysing medical facility usage behavior using big data network technique.

Looking at hospital usage behaviors in breast cancer patients can contribute to improving medical services. but, it depending on the regional characteristics of the Republic of Korea, large medical facilities are concentrated in the capital city of Seoul, so the higher the stage of breast cancer patients, the more markedly the hospital use behavior eventually moved to Seoul. Therefore, in order to solve the hospital usage behavior, it is necessary to construct a system that enables early diagnosis in a distributed form.

Discussion

The use of big data in cancer research is increasing day by day [17]. However, setting the sample itself is very important for big data. In the case of Figures 1 and 2 and Tables 1 and 2 mentioned above, mutant genes that can be used as biomarkers show high predictive values of recurrence and survival rates, but in machine learning using real big data, as shown in Table 3, the appropriate number of mutant genes is determined, and The ratio and the expression amount of a particular gene are very important characteristics [18]. In addition, it is difficult to apply to cancer patient treatment because cancer patient's hospital used behavior and network analysis techniques show the regional characteristics of each country (Figure 3) and Supplementary Figure 1). In addition, since it depends on the population density shown in the sample, it is difficult to apply it to improving hospital use and service in countries with low population density [19].

Figure 3:The regional distribution of BRCA patients and hospitals, its network.19% of total hospitals are located in Seoul, whiles 42% of total breast cancer patients visit hospitals in Seoul. This indicates strong seoul-centerism. Busan, Incheon, Daegu provinces also have higher patient visits compared to the percentage of hospitals. It all indicates metropolitan cities domination in hospital utilization. The network indicates that only few metropolitan cities attract most of the breast cancer patients.

Table 1: Kaplan meier values overall survival rate according to the number of genes.

Table 2: Kaplan meier values disease free survival rate according to the number of genes.

Table 3: Optimization of 1-500 (Option: random selection number) derivation of genetic equivalence for machine learning gene titration.

It is good to try to utilize the big data that accumulates every day, but it is necessary to balance the data in order to be used in cancer research. In other words, even when using big data, the more data is used, the lower the accuracy, and the closer to disorder. However, filtering reduces the reliability of the data because the total amount in the sample is reduced [20]. Because there is such a prisoner's dilemma as the Nash equilibrium of big data research, when using big data,it is necessary to reset the sample that is combined with the ratio of the sample rather than the simple population being formed by the organization that provides it [21].

In order to overcome these shortcomings of research using big data, first, the use of cancer patient data must be openly open and a clear sample range must be established. Second, deviating from the research methodology, each journal needs an evaluation team to evaluate whether the use of big data is the right analysis method. Third, it was the utilization of big data increases, it can be applied to various fields, so essential big data education of experts in various fields is required.

Conclusion

In this study, genetic defects were important in the study of mutant genes using machine learning in breast cancer. In addition, when looking at the results of patients' hospital use behavior through network analysis, various studies using big data are possible. However, uncertainty in the data remains. An out-of-balance especially in the proportions of the sample warns of the danger. The statistic that supports the result is very important, but it can be a statistic that applies only to a specific sample. In addition, most importantly, a clear regulation is needed to maintain the equivalence of the experimental group and the control group in the sampling of big data research.

Acknowledgement

The authors thank Dong Hyeon Lee, Da Hyun Song, You Jeong Hong and Young Geon Ji for their technical assistance with collection of data. I wrote a thesis on a completely different topic than my initial hypothesis. It is difficult to give author permission because the content of the research report is completely different. I also thank Eui-Seok Jung, who inspired us to write this paper. Even now I miss him. I pray for the repose of the deceased.

Funding

This work has partially supported by the National Research Foundation of Korea (NRF) grant founded by the Korea government (MSIT) (NRF-2019R1F1A1058771).

REFERENCES

1. Rakovitch E, Sutradhar R, Nofech-Moze  S, Gu S, Fong C, Hanna W, et al.  21-Gene Assay and Breast Cancer Mortality in Ductal Carcinoma in Situ. J Natl Cancer Inst Monogr. 2021;113(5): 572-579.
2. Torres MA, van Maaren MC, Hendriks MP, Siesling, S, & Geleijnse G. (2021). Explainable machine learning can outperform Cox regression predictions and provide insights in breast cancer survival. Nature. 2021; 11(1): 1-13.
3. Gao R X, Wang, L, Helu M, Teti R. (2020). Big data analytics for smart factories of the future. CIRP annals, 2020;69(2): 668-692.
4. Franz C. Innovation for health: success factors for the research-based pharmaceutical industry. (In Evolving Business Models) 2021.
5. Pinker K, Chin J, Melsaether AN, Morris EA, Moy L. Precision medicine and radiogenomics in breast cancer: new approaches toward diagnosis and treatment. Radiology. 2018;287(3): 732-747.
6. Kang, M. Y, Park DH. The Age of Smart Healthcare, Prepare for the Data War. Issue Monitor, Samjong KPMG, 2011.
7. Yerrapragada G, Siadimas A, Babaeian A, Sharma V, Neill OJ. (2021). Machine learning to predict tamoxifen nonadherence among US commercially insured patients with metastatic breast cancer. JCO Clin Cancer Inform. 2021; 5(2): 814-825
8. Cho HJ, Lee S, Ji YG, Lee DH. (2018). Association of specific gene mutations derived from machine learning with survival in lung adenocarcinoma. PLoS One. 2018;13(11): e0207204.
9.  Shim EJ, Lee JW, Cho J, Jung HK, Kim NH, Lee JE, et al. (2020). Association of depression and anxiety disorder with the risk of mortality in breast cancer: a National Health Insurance Service study in Korea. Breast Cancer Res Treat. 2020; 179(2):  491-498.
10. Bowman CE, Wolfgang MJ. (2019). Role of the malonyl-CoA synthetase ACSF3 in mitochondrial metabolism. Adv Biol Regul. 2019; 71(1): 34-40.
11. Saadat KA, Lestari W, Pratama E , Ma T., Iseki S , Tatsumi M, et al. (2021). Distinct and overlapping roles of ARID3A and ARID3B in regulating E2F dependent transcription via direct binding to E2F target genes. Int J Oncol. 2021;58(4): 1-12.
12. Chou CF, Lin WJ, Lin CC, Luber CA, Godbout, R, Mann M, et al. DEAD box protein DDX1 regulates cytoplasmic localization of KSRP. PLoS One. 2013;8(9): e73752.
13. Li Y, Deng  G, Qi Y, Zhang H, Jiang H , Geng  , et al. Downregulation of LUZP2 Is correlated with poor prognosis of low-grade glioma. Biomed Res. Int. 2020; 20(12): 1-16.
14. Su X, Hou Y, Yuan S, Tian M , Sun B , Li J, et al. CDNA, Genomic sequence cloning and sequence analysis of ribosomal protein L18A gene (RPL18A) from the Giant Panda (Ailuropoda melanoleuca). In 2010 3rd International Conference on Biomedical Engineering and Informatics. 2010;5(10) : 2165-2169.
15. Lim C, Lin AL, Zhao H. Metabolic strategies for microbial glycerol overproduction. J Chem Technol Biotechnol. 2018; 93(3): 624-628.
16. Biernacka JM, Sangkuhl K, Jenkins G, Whaley RM, Barma P, Batzler A, et al. (2015). The International SSRI Pharmacogenomics Consortium (ISPC): a genome-wide association study of antidepressant treatment response. Translational psychiatry. 2015;5(4): e553-e553.
17. Rehman A., Naz S, Razzak I. Leveraging big data analytics in healthcare enhancement: trends, challenges and opportunities. Multimedia Systems.2021; 5(11):1-33.
18. Malta TM, Sokolov A, Gentles A, Burzykowski T, Poisson L, Weinstein JN, et al. Machine learning identifies stemness features associated with oncogenic dedifferentiation. Cell. 2021; 173(2): 338-354.
19. Herwartz H, Schley K. (2018). Improving health care service provision by adapting to regional diversity: an efficiency analysis for the case of Germany. Health Policy. 2018; 122(3): 293-300.
20. Li H, Kadav A, Durdanovic I, Samet H, Graf, HP. (2016). Pruning filters for efficient convnets. 2016; 16(8).
21. World Health Organization. WHO report on cancer: setting priorities, investing wisely and providing care for all: WHO, 2021.