K-nearest-neighbor algorithm to predict the survival time and classification of various stages of oral cancer: a machine learning approach

Research background

Several studies on oral cancer indicate that it is one of the most prevalent cancers worldwide (Cervino Gabriele, 2019). Tackling the mortality rate is crucial as the percentage of people diagnosed with oral cancer is considerably growing. According to the GLOBOCAN 2008 estimate (Ferlay et al., 2010), 7.6 million deaths have already occurred owing to oral cancer; further, in India, approximately 20 per 100,000 are diagnosed with oral cancer every day, of which 64.8% are males and 35.2% females. In the US, as per the American Cancer Society’s (ACN) Cancer Facts and Figures 2020 (Siegel, Rebecca L,2020), at least one person is killed every hour because of oral cancer, and approximately 38,330 men and 14,880 women are being diagnosed with oral cancer every year. The estimate of oral cancer deaths in China is approximately 0.89/100,000, and hospitals encounter roughly around 52,500 new cases every year (Zhang, Shao Kai, 2015). In Saudi Arabia, the annual incidence is evaluated with >3.29% cases (Al-Jaber, 2016). These statistics demonstrate the necessity to improve oral cancer treatments. Han-Jun Cho et al. used machine learning techniques to draw the association between specific gene mutations and the survival factor in lung squamous cell carcinoma (Cho, Han Jun 2018); to this end, the RapidMinor tool was adopted for the implementation, and feature extraction was realized using the Chi-Squared test and correlation algorithms. Further, they used various classification algorithms such as Naïve Bayes, KNN, support vector machine (SVM), and decision trees to yield specific gene mutations efficiently; Fisher’s extract test and Kaplan–Meier analysis were adopted for the data analysis. The work is implemented on the cancer genome atlas (TCGA) lung adenocarcinoma (LUAD) with the clinical information of 471 patients. Matlak and Szczurek (2017) showed the type of mutations that influences the combativeness of cancer and its consequence on the patients’ survival time. The interactions between the mutated genes are accelerated by epistatic communication. The statistical likelihood-ratio test was proposed to recognize the biomarker tumor suppressor p53-binding protein (TP53BP1) that targets poly-ADP (Adenosine diphosphate) ribose polymerase in breast cancer gene-impaired tumors. Zhang et al. studied the association between racial disparities related to cancer survival and mutation (Zhang, Wensheng, 2017). They conducted a consolidative analysis of TCGA clinical samples and genomic data by establishing a relation between racial imbalance and patient survival time. The analyses were based on the Kaplan–Meier standards and achieved an accuracy of 0.69 area under curve (AUC) measure. Chen et al. conducted a meta-analysis of the connection between TP53 mutations and survival time in osteosarcoma patients using published data (Chen, Zhe, 2016). Their study suggested that TP53 mutations had no impact on a patient’s survival time, thus indicating that TP53 mutations could be effectively used as a prognostic marker to estimate patients’ survival rate suffering from osteosarcoma. Their study also demonstrated that altered TP53 had a poor response to chemotherapy and brought down the survival rate of cancer patients. Ling et al. constructed a cohort for metastatic breast cancer patients using natural language processing techniques such as semi-supervised classification (Ling, Albee Y., 2019). They classified the patients’ EMRs into the de-novo stage or recurrent metastatic breast cancer stage, proven with good sensitivity and specificity measures. The SVM technique was used to classify and categorize patients with diabetes based on their EMRs progress notes (Wright, Adam, 2013). This method achieved an F-score = 0.93 and AUC = 0.956.

Phase – I: Data acquisition

A total of 1,505 oral cancer data records were downloaded from the International Cancer Genome Consortium (ICGC) data portal. The ICGC is a global genomic data-sharing platform that provides the international community with a broad set of genomic data and related cancer types. The dataset contains 50, 178, 243, and 1,034 patients’ records from China, India, Saudi Arabia, and the United States.

Phase – II: Data analysis

i) Data cleaning: data were extracted in a.tsv file for all four countries. The dataset included several entities; however, not all the fields were required for the implementation. The parameters needed for predicting survival days are age, gender, mutations, and mutated genes. The remaining fields are dropped; ii) Data splitting: Dataset has to be split into training and test set to train an ML model efficiently. However, if the data is inappropriately broken, it may lead to problems such as; 1) overfitting—the model is trained exceptionally well on the training dataset but, the performance is poor on new/test data, and 2) under-fitting—the model does not acquire sufficient information to build the relationship between input datasets, and thus, exhibits a poor performance on training data (Zhang, Haotian, 2019). Tests were conducted for three categories—70:30, 80:20, and 90:10 to decide on the split ratio. The MAE score was equal to 7.32, 0.012, and 0.089, respectively. A model is considered good if the MAE score is small. Hence, the standard split ratio used in this experimental study is 80:20; iii) normalization: The values of the different fields in the dataset are diverse. For instance, age is denoted by integers and tumor stage by alphanumeric values; further, multiple fields had null or no weights. A standard scale to represent all the fields was required to address this disparity. Normalization was performed before loading the model with the training and test data; iv) implementation platform: The code for the proposed methodologies have been written using python programing (Python 3.9.0) on a Sypder v5.1.1platform.

Phase – III: Survival time prediction and classification of oral cancer into various stages

i. K-nearest neighbors for regression and classification

Description: KNN is the most straightforward machine learning algorithm used for both classification and regression tasks. It was initially developed in the early 1950s by Fix and Hodges (Fix, Evelyn, 1989) from the US Air Force School of Aviation Medicine. Classification involves grouping a given dataset into predefined classes, such as classifying records into one of the four stages of oral cancer (Chowdhury, Shovan, 2020). In contrast, regression is used for predicting the continuous values such as age, weight, and survival time (Huang et al., 2020b). The KNN algorithm requires a feature space that contains training data points. This ‘feature similarity’ is used to predict the new data point based on its similarity to the existing data points in the feature space. The algorithm determines the distances between an unknown data point and the nearest ‘k’ training data points and classifies the novel point to that particular class belonging. The value of ‘k’ is based on the number of data points selected from the training dataset. The algorithm primarily starts searching for a numerical feature to best draw the test data point by selecting a metric to calculate the distance (Gusti Prahmana, 2020). The distances between the new/unknown data point and the ‘k’ points are calculated using a distance measuring metric such as Euclidean distance, Manhattan distance, and Minkowski distance (Ali, Najat, 2019).

ii. KNN algorithm for classification and regression task

Comparison of actual and predicted values through error metric – Mean absolute error

Statistical representation is required for estimating the predictive capacity of the model. For this purpose, mean absolute error (MAE) is used (Botchkarev, 2018). The MAE is the simplest form of calculating the error metric that utilizes the typical magnitude of the residue values. In this method, the residue value is the difference between the actual and predicted values. This value is computed for each input, and the absolute value of these residues is considered such that the positive and negative values will not counteract and nullify. The MAE for each of the ‘k’ scores and the associated input values is calculated as follows:

The graphical representation of MAE and the value for ‘K’ are shown in Figure 2. Small MAE values indicate that the difference between the actual and predicted values is low, specifying that the model is good; further, large MAE values suggest that the model is a poor estimator (Sarker, Iqbal H, 2019). As observed in the figure, the scatter plot of the y-axis represents the coordinates for ‘i’ points of the predicted MAE. The x-axis shows the variations in ‘k’ values. When x = $x\prime$, the MAE will show the average vertical distance from the actual and predicted values. The graph between MAE and ‘k’ instances (k = 2 to 20) indicates that the MAE values decrease with an increase in the number of neighbors selected for ‘k’ values. When k = 1, the MAE has a high score, indicating that the error rate is high. The value of ‘k’ was varied from 2 to 19 to determine a suitable ‘k’ value for our experiments. When k = 6, MAE = 0.2, the least in the entire validation curve; thus, for calculating the survival time, ‘k’ was set to 6

Figure 2. Relationship between ‘k’ value used in the experiment and MAE score for predicting survival time of oral cancer patients.

Note that the MAE reaches a minimum (~0.02) score when k = 6; on the contrary, the error value is maximum (~7) when k = 2.

Classification of records to a definite oral cancer stage

The dataset now contained the predicted survival time values and the previously used parameters and was input into the KNN classifier. Again, to determine the best ‘k,’ the value was varied from 2 to 20. Experimentally, when k = 7, the MAE score attained the most negligible value of 0.1. The cloud of seven neighboring points and the new point is observed in the ‘feature similarity’ space. The test data point is marked with a class label with the highest number of instances in the feature space. If just one example of each class label exists in the feature space, the input data were marked to the closest class. The tumor stages are divided into four variations: T1, T2, T3, and T4. A total of 429 records were efficiently marked into their class belonging using the remaining 1076 data records used as the training data. Figures 3(a) –(f) shows the graphical analysis for different data points. The graph indicates other data points relative to the feature space. Each scatterplot displays the variation of cancer stages and related records with non-identical values at the x- and y-axis. A linear relationship is exhibited between any two variables selected.

Figure 3. (a). Indication of a gender (0-female, 1-male) and age with the tumor stage. As observed, aged people are more likely to be in the tumor stage-4. (b) Indication of gender and mutations with the tumor stage. As observed, the number of males (1) has a higher number of mutations than females (0). Besides, most of the males are in tumor stage-4. (c) Indication of age at diagnosis and mutations with the tumor stage. As observed, in tumor stage-4, the number of mutations is high and peaks with the increase in a patient’s age. (d) Indication of age at diagnosis and mutated genes with the tumor stage. The number of mutated genes at tumor stage-4 is high, and this pattern is observed in aged patients rather than middle-aged patients. (e) Indication of gender and mutated genes with the tumor stage. As observed, the male (1) experiences more mutated genes than the female (0). The observation also suggests that most of the males are in tumor stage-4. (f) Indication of mutations and mutated genes. As observed, tumor stage-4 has many mutations and mutated genes. Further, for a high number of mutations, the mutated genes are also high in tumor stage-4.

Error estimation for prediction of survival time using the hold-out cross-validation method

This is the most naïve and straightforward approach of the cross-validation method. The entire dataset is split up only once into training data and test data. The training dataset is chosen randomly, and the fit function is used to train the KNN classifier. The test data, usually 1/3 of the original data, is later predicted and compared with the actual values (Xu, Yun, 2018). The errors obtained are stored and returned as a mean absolute test error - a risk metric, for evaluating the regression model. The MAE score attained in this method is equal to 9.3, which is relatively high for the model. The high MAE score indicates that the model is performing poorly for the dataset. Interestingly, the cross-validation method consumes minimal time to compute the model’s efficiency, leading to the minimal time complexity of the model. This is potentially due to high variance leading to over-fitting.

Error estimation for prediction of survival time using the k-fold cross-validation method

To improve over the hold-out method, the data must be split into equally sized observations (Wong, Tzu Tsung, 2020). The dataset comprised 1505 data entries. For the k-fold error estimation, only 1500 were considered, and five samples were skipped based on selective sampling. A single fold dataset was used for testing, and the remaining k-1 folds were used for training (Yadav, Sanjay, 2016). According to kfold cross-validation, the entire dataset D_s (s = 1, 2, 3, … n) was positioned uniformly. Then dataset D_s was divided at k-folds of equal sizes such that each division D_i will have ‘k’ number of data points to be evaluated. For experimentation, we used the 10-fold cross-validation method. The dataset contained 1500 records; therefore, F={f₁,f₂,f₃, … .f₁₅₀₀}, and each feature f_i itself contained attributes such as age, gender, tumor stage, survival time, mutations, and mutated genes. Thus, Eq. (4), explains how the 10fold cross features are selected and provided for the experiment. Since k = 10, the dataset ‘F’ is divided into ten groups such that each group consists of 150 data points.

The accuracy is calculated at each fold, and the mean of all the 10-fold accuracies is estimated. The test data are kept isolated at all the iterations at each fold to obtain an unbiased approximate model performance. The test data were never used to fine-tune the model. The entire data set is divided into 10-fold ‘training’ and ‘test’ data. Once the data are divided into 10-folds training and test data, the accuracy is calculated using the standard metric MAE score. The values obtained for the validation error rate are as follows: -0.036, -0.030, 0.014, 0.019, 0.040, 0.035, 0.061, 0.020, 0.028, -0.0001; further, the mean score of the validation error rate is 0.015. The score is perfect as the error rate is negligible and could be rounded to zero. Thus, the k-fold method outperformed the hold-out method in terms of the MAE score when practically applied.

Validating the classification of patients’ records into specific oral cancer stage through f-measures

Classification accuracy is the number of correct predictions from all the predictions made in the proposed model. With the accuracy measure, the robustness of the model is determined. In most cases, the classifier results are presented in the form of a confusion matrix, which is generally expressed in terms of four measures: true-positive (TP), true-negative (TN), false-positive (FP), and false-negative (FN) (Lever, Jake, 2016). TP is when the cancer is classified into its correct stage; TN is when a patient’s record is not classified into any wrong stage; FP is when the patient’s record is classified into the wrong stage, e.g., when the patient’s record is ranked as cancer stage-4, but it should have been recognized as cancer stage-3. Finally, FN is when the patients’ record is wrongly classified into a specific stage or sometimes not recognized as a cancer stage. FN leads to disastrous results for cancer datasets as a patient would be detected as not having cancer. Further, no measures would be taken, cancer would aggravate, and the patient will still not be diagnosed until severe symptoms are observed. The subsequent metrics related to accuracy are precision and recall. Precision measures the number of positive predictions over all the positive classes predicted. A recall is a true positive rate that indicates the ratio of all cancer samples that the model accurately predicted. The f-score represents the balancing weight between the recall and precision. Eqs. 5a, 5b, and 5c explain the recall, precision, and Fmeasure, respectively.

The accuracy achieved for the classification of cancer stages is shown in Figure 4. The model efficiently classified 429 data records into one of the four oral cancer stages.

Figure 4. Screenshot of the accuracy metrics obtained by the proposed model.

The figure illustrates the following; i) the confusion matrix at the top left panel. There were 429 records used for the classification task. The confusion matrix suggests the recall, f1-score, and iii) accuracy, macro, and weighted average accuracy. Same, both row-wise and column-wise, ii) calculation of precision.

Initially, the top left panel of Figure 4 indicates the confusion matrix of the four stages of oral cancer obtained with the classifier. As observed, row-wise and column-wise addition of the confusion matrix results in a total of 429 as indicated by support records (106 + 119 + 113 + 91 = 429). Using the formulae, recall, precision, and f-score are calculated for each of the cancer stage. For T1, precision (column-wise representation) = 97 / (97 + 15 + 3 + 8) = 0.78; recall (row-wise representation) = 97 / (97 + 3 + 1 + 5) = 0.91; and f-score = (2 * 0.78 * 0.91) / (0.78 + 0.91) = 0.84. For T2, precision = 99 / (3 + 99 + 2 + 2) = 0.93; recall = 99/ (15 + 99 + 2 + 3) = 0.83; and f-score = (2 * 0.93 * 0.83) / (0.93 + 0.83) = 0.87. For T3, precision = 95 / (1 + 2 + 95 + 4) = 0.93; recall = 95/ (3 + 2 + 95 + 13) = 0.84; and f-score = (2 * 0.93 * 0.84) / (0.93 + 0.84) = 0.88. Finally, for T4, precision = 77 / (5 + 3 + 13 + 77) = 0.78; recall = 77/ (8 + 2 + 4 + 77) = 0.85; and f-score = (2 * 0.78 * 0.84) / (78 * 0.84) = 0.80. The overall accuracy is equal to the number of correctly classified samples (97 + 99 + 95 + 77) divided by total number of samples (429) = 0.85. To verify the overall f-score of the classifier, macro and weighted average arithmetic mean are considered. The macro-averaged score is the arithmetic mean of all the recall, precision, and f1-scores obtained thus far. Therefore, the macro-average score for precision = (0.78 + 0.93 + 0.93 + 0.78) / 4 = 0.85; recall = (0.91 + 0.83 + 0.84 + 0.84) / 4 = 0.85; and f1-score = (0.84 + 0.87 + 0.88 + 0.80) / 4 = 0.84. The weighted average score is calculated as the product of the weighted sum of each of the recall, precision, and f1-scores and individual support records over the total number of support records for each cancer stage (429). The weighted average for precision = (0.78 * 106 + 0.93 * 119 + 0.93. * 113 + 0.78 * 91) / 429 = 0.86; recall = (0.91 * 106 + 0.83 * 119 + 0.84 * 113 + 0.84 * 91) / 429 = 0.85; and f1-score = (0.84 * 106 + 0.87 * 119 + 0.88 * 113 + 0.80 * 91) /429 = 0.85. Finally, the accuracy is calculated by looking at all the samples at once and identifying the correctly predicted cancer stages (TP) present in the mid-diagonal of the confusion matrix over the sum of correctly predicted samples (TP) and falsely predicted samples (FP). Therefore, the overall accuracy obtained by the proposed classifier is calculated as follows: (97 + 99 + 95 + 77) / [(97 + 99 + 95 + 77) + (3 + 1 + 5 + 15 + 2 + 3 + 3 + 2 + 13 + 8 + 2 + 4)] = 0.8578. Thus, the proposed classifier for efficiently segregating oral cancer data records into various stages in a multi-class classification is 85.7% accurate, and the same demonstrated in the experimental results.