Development of a classifier with analysis of feature selection methods for COVID-19 diagnosis

Abstract

Purpose The COVID-19 pandemic situation is increasing day by day and has affected the lifestyle and economy worldwide. Due to the absence of specific treatment, the only way to control a pandemic is by stopping its spread. Early identification of affected persons is urgently in demand. Diagnostic methods applied in hospitals are time-consuming, which delay the identification of positive patients. This study aims to develop machine learning-based diagnosis model which can predict positive cases and helps in decision-making. Design/methodology/approach In this research, the authors have developed a diagnosis model to check coronavirus positivity based on an artificial neural network. The authors have trained the model with clinically assessed symptoms, patient-reported symptoms, other medical histories and exposure data of the person. The authors have explored filter-based feature selection methods such as Chi2, ANOVA F-score and Mutual Information for improving performance of a classification model. Metrics used to evaluate performance of the model are accuracy, precision, sensitivity and F1-score. Findings The authors got highest classification performance with model trained with features ranked according to ANOVA FS method. Highest scores for accuracy, sensitivity, precision and F1-score of predictions are 0.93, 0.99, 0.94 and 0.93, respectively. The study reveals that most relevant predictors for COVID-19 diagnosis are sob severity, cough severity, sob presence, cough presence, fatigue and number of days since symptom onset. Originality/value Treatment for COVID-19 is not available to date. The best way to control this pandemic is the isolation of positive persons. It is very much necessary to identify positive persons at an early stage. RT-PCR test used to check COVID-19 positivity is the time-consuming, expensive and laborious method. Current diagnosis methods used in hospital demand more medical resources with increasing cases of coronavirus that introduce shortage of resources. The developed model provides solution to the problem cheaper and faster decreases the immediate need for medical resources and helps in decision-making.

Chi-Square and PCA Based Feature Selection for Diabetes Detection with Ensemble Classifier

“…Empirical findings of this study showed that combining two highly correlated feature sets obtained using Chi2 and PCA is better than using either of them. Recent studies show that feature combinations help to obtain higher classification accuracy [29,30].…”

Seleksi Fitur pada Supervised Learning: Klasifikasi Prestasi Belajar Mahasiswa Saat dan Pasca Pandemi COVID-19

“…Untuk itu diperlukan metode seleksi fitur yang baik dan sesuai dengan data. Adapun beberapa metode seleksi fitur berbasis filter yang umum digunakan adalah Chi-Square dan Mutual Information [43].…”

“…Chi-square adalah metode statistik yang melakukan uji signifikansi data terhadap hubungan antara nilai suatu fitur dengan kelas target [46]- [49]. Sedangkan Mutual Informaton adalah bentuk umum dari koefisien korelasi yang digunakan untuk mengukur ketergantungan antar variabel [37], [43], [46]. Penerapan chi-square dan mutual information pernah dilakukan oleh beberapa penelitian sebelumnya.…”

“…MI populer karena komputasinya yang cepat dan sederhana [54]. MI adalah bentuk umum dari koefisien korelasi yang digunakan untuk mengukur ketergantungan antar variabel [37], [43], [46]. Nilai MI merepresentasikan derajat korelasi antara fitur dan kelas [61].…”

A Novel COVID-19 Diagnosis Approach Utilizing a Comprehensive Set of Diagnostic Information (CSDI)

“…As presented in article Chauhan H. et al, an artificial neural network served as the diagnostic model for assessing coronavirus positivity [24]. The researchers trained this model using a combination of clinically evaluated symptoms, patient-reported symptoms, case histories, and human exposure data.…”