Combination Of Feature Selection Methods Research Articles

Augmentation in performance and security of WSNs for IoT applications using feature selection and classification techniques

Energy consumption, secure connection of Sensor Nodes (SNs), and their performance analysis play a major role in Wireless Sensor Networks (WSNs) research. Increased case use of various Internet of Things (IoT) applications has led to the development of complicated networks. To address the threats and security issues faced in complicated WSNs, a novel combined feature selection method known as the Fast Correlation-based Feature Selection (FCBFS) with XG-Boost has been proposed for the NSL-KDD intrusion detection benchmark dataset. It helps in the selection of the best features in a cluster-based WSN before the classification step. Five popular machine learning-based classifiers, namely decision tree, random forest, Naïve Bayes, extra tree, and XG-Boost, are used for the development of a robust intrusion detection system in WSN and its IoT applications. The effectiveness and robustness of the developed ensemble methods have been checked using classification accuracy, precision, recall, and F-score. XG-Boost classifier along with the FCBFS process has performed best with a classic accuracy, precision, recall, and F-score up to 99.84%, 99.83%, 99.84%, and 99.82% on multiple runs. Upon comparison with the existing state-of-the-art work in this field, the proposed work has outperformed. Results show that, in contrast to the previous filter approaches, our proposed process can successfully minimize the number of features while maintaining a high classification precision and recognition rate. It further tends to lower the overall energy required by sensor nodes during attack detection, hence extending the network lifetime and usefulness to a sufficient time frame.

Radiomics model based on features of axillary lymphatic nodes to predict axillary lymphatic node metastasis in breast cancer.

Breast cancer (BC) is among the most common cancers worldwide. Machine learning-based radiomics model could predict axillary lymph node metastasis (ALNM) of BC accurately. The purpose is to develop a machine learning model to predict ALNM of BC by focusing on the radiomics features of axillary lymphatic node (ALN). A group of 398 BC patients with 800 ALNs were retrospectively collected. A set of patient characteristics were obtained to form clinical factors. Three hundred and twenty-six radiomics features were extracted from each region of interest for ALN in contrast-enhanced computed tomography (CECT) image. A framework composed of four feature selection methods and 14 machine learning classification algorithms was systematically applied. A clinical model, a radiomics model, and a combined model were developed using a cross-validation approach and compared. Metrics of the area under the curve (AUC), accuracy, sensitivity, and specificity were calculated to evaluate the performance of these models in the prediction of ALNM in BC. Among the 800 cases of ALNs, there were 388 cases of positive metastasis (48.50%) and 412 cases of negative metastasis (51.50%). The baseline clinical model achieved the performance with an AUC = 0.8998 (95% CI [0.8540, 0.9457]). The radiomics model achieved an AUC = 0.9081 (95% CI [0.8640, 0.9523]). The combined model using the clinical factors and radiomics features achieved the best results with an AUC = 0.9305 (95% CI [0.8928, 0.9682]). Combinations of feature selection methods and machine learning-based classification algorithms can develop promising predictive models to predict ALNM in BC using CECT features. The combined model of clinical factors and radiomics features outperforms both the clinical model and the radiomic model.

Evaluation of the dependence of radiomic features on the machine learning model

BackgroundIn radiomic studies, several models are often trained with different combinations of feature selection methods and classifiers. The features of the best model are usually considered relevant to the problem, and they represent potential biomarkers. Features selected from statistically similarly performing models are generally not studied. To understand the degree to which the selected features of these statistically similar models differ, 14 publicly available datasets, 8 feature selection methods, and 8 classifiers were used in this retrospective study. For each combination of feature selection and classifier, a model was trained, and its performance was measured with AUC-ROC. The best-performing model was compared to other models using a DeLong test. Models that were statistically similar were compared in terms of their selected features.ResultsApproximately 57% of all models analyzed were statistically similar to the best-performing model. Feature selection methods were, in general, relatively unstable (0.58; range 0.35–0.84). The features selected by different models varied largely (0.19; range 0.02–0.42), although the selected features themselves were highly correlated (0.71; range 0.4–0.92).ConclusionsFeature relevance in radiomics strongly depends on the model used, and statistically similar models will generally identify different features as relevant. Considering features selected by a single model is misleading, and it is often not possible to directly determine whether such features are candidate biomarkers.

Recommendation for Classification of News Categories Using Support Vector Machine Algorithm with SVD

Online news is a digital information media currently has a very easy and flexible updating process. The News Document grouping process is implemented in several stages, including Text Mining which includes Text Pre-processing which includes Tokenizing, Stopword removal, Stemming, Word Merging, TF-IDF and Confusion Matrix. Of the several techniques in Text Mining, the most frequently used for News Document classification is the Support Vector Machine (SVM). SVM has the advantage of being able to identify separate hyperplane that maximizes the margin between two or more different classes. The selection of features in SVM significantly affects the classification accuracy results. Therefore, in this study a combination of feature selection methods is used, namely Singular Value Decomposition in order to increase accuracy and reduce the Classifier Time Support Vector Machine. This research resulted in text classification in the form of categories Entertainment, Health, Politics and Technology. Based on the Support Vector Machines Algorithm, an accuracy rate of 81% was obtained with 360 Data Training and 120 Data Testing, after adding the Singular Value Decomposition feature with a K- Rank value of 50%, a significant increase in accuracy was obtained with an accuracy value of 94% and The time of Algorithm process is faster.

Artificial neural network tools for predicting the functional response of ultrafast laser textured/structured surfaces

Artificial Neural Networks (ANNs) are well-established knowledge acquisition systems with proven capacity for learning and generalisation. Therefore, ANNs are widely applied to solve engineering problems and are often used in laser-based manufacturing applications. There are different pattern recognition and control problems where ANNs can be effectively applied, and one of them is laser structuring/texturing for surface functionalisation, e.g. in generating Laser-Induced Periodic Surface Structures (LIPSS). They are a particular type of sub-micron structures that are very sensitive to changes in laser processing conditions due to processing disturbances like varying Focal Offset Distance (FOD) and/or Beam Incident Angle (BIA) during the laser processing of 3D surfaces. As a result, the functional response of LIPSS-treated surfaces might be affected, too, and typically needs to be analysed with time-consuming experimental tests. Also, there is a lack of sufficient process monitoring and quality control tools available for LIPSS-treated surfaces that could identify processing patterns and interdependences. These tools are needed to determine whether the LIPSS generation process is in control and consequently whether the surface’s functional performance is still retained. In this research, an ANN-based approach is proposed for predicting the functional response of ultrafast laser structured/textured surfaces. It was demonstrated that the processing disturbances affecting the LIPSS treatments can be classified, and then, the surface response, namely wettability, of processed surfaces can be predicted with a very high accuracy using the developed ANN tools for pre- and post-processing of LIPSS topography data, i.e. their areal surface roughness parameters. A Generative Adversarial Network (GAN) was applied as a pre-processing tool to significantly reduce the number of required experimental data. The number of areal surface roughness parameters needed to fully characterise the functional response of a surface was minimised using a combination of feature selection methods. Based on statistical analysis and evolutionary optimisation, these methods narrowed down the initial set of 21 elements to a group of 10 and 6 elements, according to redundancy and relevance criteria, respectively. The validation of ANN tools, using the salient surface parameters, yielded accuracy close to 85% when applied for identification of processing disturbances, while the wettability was predicted within an r.m.s. error of 11 degrees, equivalent to the static water contact angle (CA) measurement uncertainty.

Identification of important features and data mining classification techniques in predicting employee absenteeism at work

Employees absenteeism at the work costs organizations billions a year. Prediction of employees’ absenteeism and the reasons behind their absence help organizations in reducing expenses and increasing productivity. Data mining turns the vast volume of human resources data into information that can help in decision-making and prediction. Although the selection of features is a critical step in data mining to enhance the efficiency of the final prediction, it is not yet known which method of feature selection is better. Therefore, this paper aims to compare the performance of three well-known feature selection methods in absenteeism prediction, which are relief-based feature selection, correlation-based feature selection and information-gain feature selection. In addition, this paper aims to find the best combination of feature selection method and data mining technique in enhancing the absenteeism prediction accuracy. Seven classification techniques were used as the prediction model. Additionally, cross-validation approach was utilized to assess the applied prediction models to have more realistic and reliable results. The used dataset was built at a courier company in Brazil with records of absenteeism at work. Regarding experimental results, correlationbased feature selection surpasses the other methods through the performance measurements. Furthermore, bagging classifier was the best-performing data mining technique when features were selected using correlation-based feature selection with an accuracy rate of (92%).

Classification of ASD based on fMRI data with deep learning.

Autism spectrum disorder (ASD) is a neuro-developmental disorder that affects the social abilities of patients. Studies have shown that a small number of abnormal functional connections (FCs) exist in the cerebral hemisphere of ASD patients. The identification of these abnormal FCs provides a biological ground for the diagnosis of ASD. In this paper, we propose a combined deep feature selection (DFS) and graph convolutional network method to classify ASD. Firstly, in the DFS process, a sparse one-to-one layer is added between the input and the first hidden layer of a multilayer perceptron, thus each functional connection (FC) feature can be weighted and a subset of FC features can be selected accordingly. Then based on the selected FCs and the phenotypic information of subjects, a graph convolutional network is constructed to classify ASD and typically developed controls. Finally, we test our proposed method on the ABIDE database and compare it with some other methods in the literature. Experimental results indicate that the DFS can effectively select critical FC features for classification according to the weights of input FC features. With DFS, the performance of GCN classifier can be improved dramatically. The proposed method achieves state-of-the-art performance with an accuracy of 79.5% and an area under the receiver operating characteristic curve (AUC) of 0.85 on the preprocessed ABIDE dataset; it is superior to the other methods. Further studies on the top-ranked thirty FCs obtained by DFS show that these FCs are widespread over the cerebral hemisphere, and the ASD group appears a significantly higher number of weak connections compared to the typically developed group.

Combination of Feature Selection and CatBoost for Prediction: The First Application to the Estimation of Aboveground Biomass

Increasing numbers of explanatory variables tend to result in information redundancy and “dimensional disaster” in the quantitative remote sensing of forest aboveground biomass (AGB). Feature selection of model factors is an effective method for improving the accuracy of AGB estimates. Machine learning algorithms are also widely used in AGB estimation, although little research has addressed the use of the categorical boosting algorithm (CatBoost) for AGB estimation. Both feature selection and regression for AGB estimation models are typically performed with the same machine learning algorithm, but there is no evidence to suggest that this is the best method. Therefore, the present study focuses on evaluating the performance of the CatBoost algorithm for AGB estimation and comparing the performance of different combinations of feature selection methods and machine learning algorithms. AGB estimation models of four forest types were developed based on Landsat OLI data using three feature selection methods (recursive feature elimination (RFE), variable selection using random forests (VSURF), and least absolute shrinkage and selection operator (LASSO)) and three machine learning algorithms (random forest regression (RFR), extreme gradient boosting (XGBoost), and categorical boosting (CatBoost)). Feature selection had a significant influence on AGB estimation. RFE preserved the most informative features for AGB estimation and was superior to VSURF and LASSO. In addition, CatBoost improved the accuracy of the AGB estimation models compared with RFR and XGBoost. AGB estimation models using RFE for feature selection and CatBoost as the regression algorithm achieved the highest accuracy, with root mean square errors (RMSEs) of 26.54 Mg/ha for coniferous forest, 24.67 Mg/ha for broad-leaved forest, 22.62 Mg/ha for mixed forests, and 25.77 Mg/ha for all forests. The combination of RFE and CatBoost had better performance than the VSURF–RFR combination in which random forests were used for both feature selection and regression, indicating that feature selection and regression performed by a single machine learning algorithm may not always ensure optimal AGB estimation. It is promising to extending the application of new machine learning algorithms and feature selection methods to improve the accuracy of AGB estimates.

Radiomic Machine Learning Classifiers in Spine Bone Tumors: A Multi-Software, Multi-Scanner Study

PurposeSpinal lesion differential diagnosis remains challenging even in MRI. Radiomics and machine learning (ML) have proven useful even in absence of a standardized data mining pipeline. We aimed to assess ML diagnostic performance in spinal lesion differential diagnosis, employing radiomic data extracted by different software. MethodsPatients undergoing MRI for a vertebral lesion were retrospectively analyzed (n = 146, 67 males, 79 females; mean age 63 ± 16 years, range 8-89 years) and constituted the train (n = 100) and internal test cohorts (n = 46). Part of the latter had additional prior exams which constituted a multi-scanner, external test cohort (n = 35). Lesions were labeled as benign or malignant (2-label classification), and benign, primary malignant or metastases (3-label classification) for classification analyses. Features extracted via 3D Slicer heterogeneityCAD module (hCAD) and PyRadiomics were independently used to compare different combinations of feature selection methods and ML classifiers (n = 19). ResultsIn total, 90 and 1548 features were extracted by hCAD and PyRadiomics, respectively. The best feature selection method-ML algorithm combination was selected by 10 iterations of 10-fold cross-validation in the training data. For the 2-label classification ML obtained 94% accuracy in the internal test cohort, using hCAD data, and 86% in the external one. For the 3-label classification, PyRadiomics data allowed for 80% and 69% accuracy in the internal and external test sets, respectively. ConclusionsMRI radiomics combined with ML may be useful in spinal lesion assessment. More robust pre-processing led to better consistency despite scanner and protocol heterogeneity.

Feature selection in credit risk modeling: an international evidence

This paper aims to discover a suitable combination of contemporary feature selection techniques and robust prediction classifiers. As such, to examine the impact of the feature selection method on classifier performance, we use two Chinese and three other real-world credit scoring datasets. The utilized feature selection methods are the least absolute shrinkage and selection operator (LASSO), multivariate adaptive regression splines (MARS). In contrast, the examined classifiers are the classification and regression trees (CART), logistic regression (LR), artificial neural network (ANN), and support vector machines (SVM). Empirical findings confirm that LASSO's feature selection method, followed by robust classifier SVM, demonstrates remarkable improvement and outperforms other competitive classifiers. Moreover, ANN also offers improved accuracy with feature selection methods; LR only can improve classification efficiency through performing feature selection via LASSO. Nonetheless, CART does not provide any indication of improvement in any combination. The proposed credit scoring modeling strategy may use to develop policy, progressive ideas, operational guidelines for effective credit risk management of lending, and other financial institutions. The finding of this study has practical value, as to date, there is no consensus about the combination of feature selection method and prediction classifiers.

Pushing the limits of solubility prediction via quality-oriented data selection.

SummaryAccurate prediction of the solubility of chemical substances in solvents remains a challenge. The sparsity of high-quality solubility data is recognized as the biggest hurdle in the development of robust data-driven methods for practical use. Nonetheless, the effects of the quality and quantity of data on aqueous solubility predictions have not yet been scrutinized. In this study, the roles of the size and the quality of data sets on the performances of the solubility prediction models are unraveled, and the concepts of actual and observed performances are introduced. In an effort to curtail the gap between actual and observed performances, a quality-oriented data selection method, which evaluates the quality of data and extracts the most accurate part of it through statistical validation, is designed. Applying this method on the largest publicly available solubility database and using a consensus machine learning approach, a top-performing solubility prediction model is achieved.

Sentiment analysis of facebook comments on indonesian presidential candidates using the naïve bayes method

Sentiment analysis is a very interesting research object to study. Sentiment analysis itself is one branch of textmining whose research focuses on the opinion of a text document. In this study, the author examines the sentiment analysis of facebook commentary of 2 Indonesian presidential real candidates in the 2014. Henceforth in this study, the two candidates are referred to as Presidential Candidate 1 and Presidential Candidate 2, where the numbering sequence is adjusted to the numbering of real election data of president in Indonesia. Here the author chooses to use Facebook comment data on several statuses posted on the 2 official accounts of the Indonesian presidential candidates, because Facebook is a social media that is widely used by Indonesians, it is evident that Facebook’s social media ranks 3rd in Indonesia. In the process of classifying this text using the Naïve Bayes method because this method is very simple, has good performance in many domains and this method is very simple. But the Naïve Bayes method itself, has the disadvantage of being very sensitive to too many features, which can lead to low classification accuracy. To overcome the problems that exist in the Naïve Bayes method, this study uses a combination of feature selection methods, namely information gain and genetic algorithm, the two additional methods serve to improve the accuracy in Naïve Bayes classifier. In addition, in this study the author uses smile-forming character conversion, pre-processing of documents such as transform case, tokenization, Indonesian stopwords, Stemming, Token weighting, then classification and confusion matrix testing. This research produces a positive or negative text classification from Facebook comments. Then the measurement is based on the accuracy of Naïve Bayes before and after the addition of the feature selection method. The evaluation process uses 10 fold cross validation. From the results of the implementation and testing, the Naïve Bayes method with feature selection has an accuracy level of sentiment classification of 83.67% from the previous results of 60.00% here the researcher also displays the ROC curve.

Classification of white blood cells using deep features obtained from Convolutional Neural Network models based on the combination of feature selection methods

White blood cells are cells in the blood and lymph tissue produced by the bone marrow in the human body. White blood cells are an important part of the immune system. The most important task of these cells is to protect the human body against foreign invaders and infectious diseases. When the number of white blood cells in the blood is not enough for the human body, it can cause leukopenia. As a result of this situation, the resistance of the human body against infections and diseases decreases. In this respect, determining the number of these cells in the human body is a specialist task. Detection and treatment of this symptom is a labor-intensive process carried out by specialist doctors and radiologists. Image processing techniques have recently been widely used in biomedical systems for the diagnosis of various diseases. In this study, it is aimed to use image processing techniques to improve the classification performance of deep learning models in white blood cells classification. To perform the classification process more efficiently, the Maximal Information Coefficient and Ridge feature selection methods were used in conjunction with the Convolutional Neural Network models. The Maximal Information Coefficient and Ridge feature selection methods extracted the most relevant features. Afterward, the classification process was realized by using this feature set. In this study, AlexNet, GoogLeNet, and ResNet-50 were used as feature extractor and quadratic discriminant analysis was used as a classifier. As a result, the overall success rate was obtained as 97.95% in the classification of white blood cells. The experimental results showed that the use of the convolutional neural network models with feature selection methods contributed to improving the classification success of white blood cell types.

SENTIMENT ANALYSIS ON CLOSURE OF ILLEGAL MOVIE STREAMING SITES USING NAÏVE BAYES ALGORITHM

The closure of illegal movie streaming sites IndoXXI has been a trending topic on Twitter at the end of 2019. The reaction of netizens on Twitter shows positive and negative sentiments. Until now, there have been many studies in the field of Sentiment Analysis using data in the form of Tweets from Twitter users. In sentiment analysis research, there are so many method used, and Naïve Bayes is one of it, because it is very simple and efficient. The method has advantages and disadvantages. Naïve Bayes is so sensitive in feature selection. Too many features not only increase calculation time but also reduce classification accuracy. In order to solve the disadvantages and increase the performance of the Naïve Bayes classifier, this method often being combined with many kind of feature selection methods. This research aims to classify tweets into positive and negative using the Naïve Bayes classifier combined with the Genetic Algorithm. The accuracy of Naïve Bayes before using the combination of feature selection methods reaches 79.55%. While after using feature selection methods, which is the Genetic Algorithm, accuracy increased up to 88.64%. The accuracy improved by up to 9.09%.

Can structured EHR data support clinical coding? A data mining approach

ABSTRACT Structured data formats are gaining momentum in electronic health records and can be leveraged for decision support and research. Nevertheless, such structured data formats have not been explored for clinical coding, which is an essential process requiring significant manual workload in health organisations. This article explores the extent to which fully structured clinical data can support assignment of clinical codes to inpatient episodes, through a methodology that tackles high dimensionality issues, addresses the multi-label nature of coding and optimises model parameters. The methodology encompasses transformation of raw data to define a feature set, build a data matrix representation, and testing combinations of feature selection methods with machine learning models to predict code assignment. The methodology was tested with a real hospital dataset and showed varying predictive power across codes, while demonstrating the potential of leveraging structuring data to reduce workload and increase efficiency in clinical coding.

Improving the performance of support-vector machine by selecting the best features by Gray Wolf algorithm to increase the accuracy of diagnosis of breast cancer

One of the most common diseases among women is breast cancer, the early diagnosis of which is of paramount importance. Given the time-consuming nature of the diagnosis process of the disease, using new methods such as computer science is extremely important for early detection of the condition. Today, the main emphasis is on the science of data mining as one of the computer methods in the field of diagnosis. In the present study, we used data mining as a combination of feature selection method by Gray Wolf Optimization (GWO) and support vector machine (SVM), which is a new technique with high accuracy compared to other methods in this classification, to increase the accuracy of breast cancer diagnosis. The UCI dataset and functional parameters and various statistical criteria were applied to evaluate the proposed method and assess the validity of the results in MATLAB, respectively. Application of the proposed method increased the improvement of the evaluated criteria, which increased the accuracy of diagnosis by 27.68%, compared to former works in the field. As such, it could be concluded that the proposed method had a higher ability to diagnose breast cancer, compared to previous techniques.

KOMPARASI METODE KOMBINASI SELEKSI FITUR DAN MACHINE LEARNING K-NEAREST NEIGHBOR PADA DATASET LABEL HOURS SOFTWARE EFFORT ESTIMATION

The methods for Software Effort Estimation are divided into two, these methods are grouped into Non Machine Learning (non-ML) and Machine Learning (ML) methods [1]. The k-NN method has the disadvantage of being unable to tolerate irrelevant features and greatly affect the accuracy of k-NN. The k-NN method is also difficult to deal with missing data problems and feature categorization problems such as features that are not relevant, weight features that are not optimal, and the same features [2]. Whereas the dataset of Software Effort Estimation still has some serious challenges such as the characteristics of the data set, which are irrelevant features and the level of influence of each feature in the estimated data of the software effort [3]. This study compared the k-NN individual method with the combination of feature selection method with k-NN to find out which method was the best. The results showed that the Forward Selection (FS) method and Median Weighted Information Gain with k-Nearest Neighbor can overcome the problem of irrelevant features so as to increase accuracy in the RMSE Software Effort Estimation dataset, which is smaller in the Albrecht dataset of 5,953 using the Median method -WIG k-NN, the Miyazaki dataset is 55,421 and Kemerer is 123,081 using the FS k-NN method. The combination of kNN with Feature Selection is proven to be able to improve the estimation results better than kNN individuals. With the FS k-NN method being the best by winning in 2 datasets Miyazaki and Kemerer.

Student’s Success Prediction Model Based on Artificial Neural Networks (ANN) and A Combination of Feature Selection Methods

The improvements in educational data mining (EDM) and machine learning motivated the academic staff to implement educational models to predict the performance of students and find the factors that increase their success. EDM faced many approaches for classifying, analyzing and predicting a student’s academic performance. This paper presents a model of prediction based on an artificial neural network (ANN) by implementing feature selection (FS). A questionnaire is built to collect students’ answers using LimeSurvey and google forms. The questionnaire holds a combination of 61 questions that cover many fields such as sports, health, residence, academic activities, social and managerial information. 161 students participated in the survey from two departments (Computer Science Department and Computer Information Systems Department), college of Computer Science and Information Technology, University of Basra. The data set is combined from two sources applications and is pre-processed by removing the uncompleted answers to produce 151 answers used in the model. Apart from the model, the FS approach is implemented to find the top correlated questions that affect the final class (Grade). The aim of FS is to eliminate the unimportant questions and find those which are important, besides improving the accuracy of the model. A combination of Four FS methods (Info Gain, Correlation, SVM and PCA) are tested and the average rank of these algorithms is obtained to find the top 30 questions out of 61 questions of the questionnaire. Artificial Neural Network is implemented to predict the grade (Pass (P) or Failed (F)). The model performance is compared with three previous models to prove its optimality.

A Machine Learning Approach for Identifying Gene Biomarkers Guiding the Treatment of Breast Cancer.

Genomic profiles among different breast cancer survivors who received similar treatment may provide clues about the key biological processes involved in the cells and finding the right treatment. More specifically, such profiling may help personalize the treatment based on the patients’ gene expression. In this paper, we present a hierarchical machine learning system that predicts the 5-year survivability of the patients who underwent though specific therapy; The classes are built on the combination of two parts that are the survivability information and the given therapy. For the survivability information part, it defines whether the patient survives the 5-years interval or deceased. While the therapy part denotes the therapy has been taken during that interval, which includes hormone therapy, radiotherapy, or surgery, which totally forms six classes. The Model classifies one class vs. the rest at each node, which makes the tree-based model creates five nodes. The model is trained using a set of standard classifiers based on a comprehensive study dataset that includes genomic profiles and clinical information of 347 patients. A combination of feature selection methods and a prediction method are applied on each node to identify the genes that can predict the class at that node, the identified genes for each class may serve as potential biomarkers to the class’s treatment for better survivability. The results show that the model identifies the classes with high-performance measurements. An exhaustive analysis based on relevant literature shows that some of the potential biomarkers are strongly related to breast cancer survivability and cancer in general.

Machine Learning to Predict Postradical Prostatectomy Pathology Outcomes in Intermediate Risk Prostate Cancer

Intermediate risk prostate cancer (IRPC) is a heterogeneous disease with various treatment options. Existing efforts to further risk stratify this disease entity were developed using a small a priori determined subset of predictors. Here, we use statistical machine learning that pulls from a larger pool of demographics and clinical features to better predict post-radical prostatectomy (RP) pathologic outcomes in IRPC. Knowing the probability of high risk pathology and thus the likelihood of adjuvant radiation may inform patients deciding between upfront surgery and radiation. We retrospectively reviewed patients who underwent RP from 6/2005 to 5/2015 at our institution and included 1560 patients with clinical IRPC, defined per NCCN. Predictors included demographics, PSA and location-specific biopsy findings. The post-RP surgical pathology variables analyzed were positive margins, Gleason upgrading and pathologic upstaging.MATLAB was used. We imputed missing data and used minority oversampling to balance data. 69 predictors were eligible after pre-processing. We used a combination of feature selection methods to decrease multicollinearity and overfitting. Learning methods included logistic regression (LR) with or without lasso regularization and support vector machine with a Gaussian kernel (SVM) with or without forward feature selection. 10-fold cross validation was used to estimate the performance of our model building methods. Statistical measures are reported with syntax: sensitivity/specificity/positive predictive value (PPV)/negative predictive value (NPV). Positive margins: SVM had excellent results with 98%/95%/95%/98% (AUC ∼1.0). Feature selection chose 11 predictors with 71%/75%/74%/72% (AUC 0.83). LR with or without lasso had poorer outcomes except specificity and NPV with <10%/99%/<34%/87%.Gleason upgrading: SVM performance was again excellent with 95%/89%/90%/95% (AUC 0.97). Feature selection narrowed to 21 predictors with good performance at 85%/73%/75%/83% (AUC 0.87). LR was inferior except in specificity with 24%/93%/47%/82% (AUC 0.68). Lasso selected 6 features with 6%/100%/100%/80% (AUC 0.73).Pathologic upstaging: SVM again was superior except for specificity, with 79%/77%/77%/79% (AUC 0.85). Feature selection resulted in 30 variables with 74%/74%/73%/75% (AUC 0.82). Using LR with or without lasso (39 features) resulted in almost identical performance with 50%/84%/65%/74% (AUC 0.73). We are able to build statistical machine learning models of varying complexity to accurately estimate post-RP pathological outcomes. While overall SVM performance was superior—likely due to its ability to handle high dimensional data—LR had unique advantages, suggesting that an ensemble method may be optimal. Further development of these models, including deep learning, may allow us to inform patients and colleagues whether upfront surgery or radiation would be the optimal choice.