Epidemiology of lung cancer and approaches for its prediction: a systematic review and analysis

We found that currently the incidence and mortality patterns of lung cancer closely follow each other at the global level. In the US and the UK, advanced technology and awareness programs have helped decrease the mortality from lung cancer; however, this is not the case in India and Egypt, where more effective steps, such as development of special awareness programs, are required to decrease lung cancer mortality.

Cancer epidemiology is the study of causes and risk factors of a cancer for a given population. It can be helpful by allowing (i) the identification of health problems related to cancer, (ii) the measurement of the spread of the disease in a community, (iii) the expansion of knowledge about the risk factors of cancer, and (iv) a better understanding of the effects of cancer.

Cancer epidemiology can provide insights into the causes of cancer. However, the area under an epidemiologic investigation is often limited to a particular region and usually involves a small sample size. Since cancer epidemiology is analytical in nature, additional computational methods are required. Sample size can be increased easily, which can yield better classification results.

Data mining techniques can provide better classification and categorization of data, but these techniques may not efficiently cluster, classify, and predict the trends of sequential and time series data; hence, evolutionary algorithms are used to obtain optimal solutions in such cases. Evolutionary algorithms can produce high-quality analytical solutions and can simplify the problems during different iterative stages. Evolutionary algorithms such as ant colony optimization (ACO), particle swarm optimization (PSO), and artificial bee colony (ABC) are beneficial as these algorithms are capable of achieving a nearer solution in comparison to that achieved by the use of data mining techniques alone.

Pattern finding is very important in cancer detection. For this, data mining methods are needed. Data mining is a procedure by which pertinent patterns can be separated from large databases [40]. According to Jain et al. [41], data mining can be used for six specific tasks: classification, estimation, prediction, association rule mining, clustering, and visualization. Classification, estimation, and prediction are examples of supervised learning. The primary aim of these techniques is to prepare a model based on the available data, which can represent one or more attributes. Association rule mining, clustering, and visualization are examples of unsupervised learning. The primary aim of these techniques is to establish relationships between attributes. The six methods are used in nearly every area of healthcare databases for knowledge discovery, classification, and prediction. Of them, association rule mining, classification, and clustering are the most commonly used data mining techniques. These methods may provide a real solution for discovering similar types of groups, group patterns, the frequency of items present in the groups, the extraction of significant patterns, and pattern visualization [42]. Since lung cancer symptoms are not the same in every patient, it is essential to characterize their distinctive features and give unique treatments to different patients. In this regard, clustering or classification techniques may be useful because several factors, such as age, sex, genetics, alcohol consumption, smoking status, and weight may contribute to lung cancer.

According to Dass et al. [43], the two most important factors in cancer treatment are classification and characterization. They successfully achieved all the classification rules by using Apriori algorithm, which is helpful in the diagnosis of and the drug development for squamous cell cancer (SCC) and adenocarcinoma (ADC) [43]. Rajan et al. [44] suggested that the early diagnosis of lung cancer is mainly dependent on its historical data. Using association rule mining, Agrawal et al. [45] identified hotspots in lung cancer SEER data. A prototype mortality risk calculator was developed in this study; and the obtained rules satisfied biomedical knowledge. According to Yadav et al. [46], the detection of lung cancer is difficult at an early stage because it depends on multiple attributes. They used clustering approach for analyzing dataset from Sanjay Gandhi Post Graduate Institute of Medical Science, Lucknow, India. They compared the traditional clustering with foggy clustering method and achieved better results by using the latter. Piedra et al. [47] suggested text mining for a better understanding of the diagnostic process, classification accuracy, and disease facts; this can also be helpful in predictive model design, alert system, and decision-making process. Nahar et al. [48] used the association rule mining for identifying risk factors in different types of cancer. For this, they used three different types of association rule mining algorithms: the Apriori, predictive Apriori, and tertius algorithms. The Apriori algorithm outperforms the other algorithms. According to Wang et al. [49], Bayesian network is a very useful method for understanding cancer metastasis. The study included 50,000 cancer patients from Taiwan, China, between 1996 and 2010. Sensitivity and specificity measures were compared based on three different approaches, namely naive Bayes, logistic regression, and support vector machine (SVM), but the researchers did not find significant differences in terms of accuracy and specificity of the results. The interpretation capabilities of naive Bayes were superior to those of the other approaches, and it was also efficient in cases of missing information, modeling of non-linear situations, and stochastic medical problems. In their study, Krishnaiah et al. [50] examined decision tree, naive Bayes, and artificial neural network. One dependency-augmented naive Bayes classifier and one naive creedal classifier 2 were used for data pre-processing and decision making purposes, and the prediction results were better than the traditional methods. According to Phillips-Wren et al. [51], when decision tree and artificial neural network were used in combination, the chances of good prediction results became high. Debnath et al. [52] proposed a new evolutionary method for efficient classification of lung cancer genes. When a smaller number of genes were selected, this method provided better classification accuracy. According to Esfandiari et al. [53], data mining can determine the frequency of the task at a specified time. According to their study, data mining can be applied for disease prediction by data pre-processing and data modeling. Balachandran et al. [54] performed data mining to conduct a systematic study of lung cancer. The data were collected from medically confirmed and diagnosed patients. Their results showed that training-based approaches such as neural network performed better than cross-validation approaches. Fung et al. [55] proposed a new classifier that combined the impact factors (IFs) method and Golub and Slonim (GS) method with k-nearest neighbour (KNN). They achieved good classification performance for lung and prostate cancer data. Kushwah et al. [56] used neural network with random forest tree classifier for cancer gene selection. According to their results, classification capability could be increased with the help of trained neural networks. Guo et al. [57] used a network-based method on 164 smokers to identify the genes associated with smoking. They identified genes associated with lung cancer survival and genes that could distinguish smokers and non-smokers; the accuracy of the method was 73.0%. Ahmed et al. [58] prepared a database of 400 patients that comprised patients with or without cancer. For pre-processing, k-means clustering was used. The results proved that this method was efficient in lung cancer risk identification. Sun et al. [59] suggested that the SVM can be used for lung cancer classification, based on the comparison of different algorithms, such as boosting, decision tree, and KNN. Oztekin et al. [60] proposed a prediction model based on decision tree, neural network, and logistic regression. The study suggested that these algorithms were capable of accurate classification of the lung cancer dataset.

Evolutionary algorithms are population-based meta-heuristic optimization algorithms that are inspired by nature. The principal evolutionary algorithms are genetic algorithm (GA), ACO, PSO, ABC, and memetic algorithm. Now, evolutionary algorithms, in combination with the previously discussed methodologies, are being discussed. These algorithms can locate the closest solution even when dealing with complex issues.

Li et al. [61] proposed a bionic enhancement calculation-based system, termed ant colony optimization-selection (ACO-S) for high-dimensional datasets. The outcomes demonstrated that ACO-S could produce a high-quality subset with a small size and better characterization. Yu et al. [62] recommended ACO sampling to address the issue of class unevenness. The methodology resulted in greater grounded speculation capacity as compared with the traditional methods. Sowmiya et al. [63] suggested neural network and fuzzy logic to train data. Then, by using ACO, classification accuracy was improved. Alba et al. [64] compared PSO and GA. They used SVM in combination with either of the algorithms on high-dimensional microarray data for classification. The combination of PSO and SVM was capable of finding interesting genes. Minimum redundancy maximum relevance (MRMR)-GA was compared with GA-SVM wrapper and MRMR filter. In terms of selection and classification performance, MRMR-GA produced better results. Qasem et al. [65] presented a new multi-objective algorithm based on swarm optimization for classification problems, termed multi-objective particle swarm optimization RBF network (MPSON). The results indicated that this method had good generalization capability along with compact network structure. Runkler et al. [66] made efforts to minimize fuzzy c-means model using ACO, alternate optimization (AO), and PSO. They suggested two different forms of PSO: the first was PSO–V for representing particle as a component of a cluster center; the second was PSO–U for representing particle as a non-scaled and non-normalized membership value. PSO–V and PSO–U were compared with AO and ACO. They were compared with two different datasets: single outlier and lung cancer. The results of ACO, PSO–V, and PSO–U were slower than AO, but PSO variants outperformed significantly after each round of iteration. Liu et al. [67] suggested discrete particle swarm optimization (DPSO) and rule pruning for lung cancer diagnosis and achieved 68.3% classification accuracy. Liu et al. [68] used the PSO-based simultaneous learning framework for clustering and classification (PSOSLCC). PSOSLCC was applied to a real-world application, namely texture image segmentation, and good performance was obtained, showing that it could potentially classify problems on a large scale. Chen et al. [69] proposed an approach based on PSO with a decision tree classifier for statistical analysis. They found that this method outperformed other popular classifiers (i.e., SVM, self-organizing map, back propagation neural network, and C4.5 decision tree) by conducting experiments on 11 gene expression cancer datasets. Subbulakshmi et al. [70] proposed an efficient hybrid approach based on PSO with an extreme learning machine classifier. It had self-regulated learning capability that showed good generalization performance. These studies above suggested that the data mining and evolutionary algorithms both are efficient in lung cancer detection; while the evolutionary algorithms have the capabilities of handling complex problems, the data mining algorithms alone may fail. Therefore, combining both approaches at different levels of classification and clustering may produce better outcomes.

The use of tobacco products causes approximately 5 million deaths worldwide annually, with 2.41 million deaths in developing countries and 2.43 million deaths in developed countries [71–73]. Of the 5 million deaths that occurred annually in India, approximately 1 million could be attributed to cancer [71, 74]; by 2020, this figure is estimated to reach 1.5 million [40]. Smoking is responsible for 80.0% of the lung cancer incidences worldwide [71–74]. In India, cigarette or beedi smoking causes the majority of the deaths in the 25–69 age group [71, 75]. Some studies have reported that 15.0% of lung cancer cases were caused by genetic factors, air pollution, or exposure to radon gas, asbestos, and pesticides [71, 76, 77]. These studies also showed that, Indian non-smokers have almost the same chance of getting lung cancer as smokers because of exposure to pesticides and other carcinogens (Fig. 8) [36, 77]. Our results suggest that there must be strict restrictions on the use of tobacco products.