Experiment on Deep Learning Models for COVID-19 Detection from Blood Testing


Fiqhy Bismadhika; Nunung Nurul Qomariyah; Ardimas Andi Purwita

Due to the equipment and expert shortages in diagnosing COVID-19 disease, detecting an individual infected with Coronavirus using hematochemical data could provide a cheaper and faster alternative. The quicker and less expensive alternative could be realized by utilizing deep learning to classify Coronavirus infection using complete blood count test results. Two architectures are developed and implemented in this study, which is custom-built DNN (Deep Neural Network) and TabNet. Also, three datasets from the hospitals in Italy, Brazil, and Indonesia are used for training the models. The deep learning models trained with the datasets from San Raphael Hospital in Italy, Albert Einstein Hospital in Brazil, and Pasar Minggu Hospital in Indonesia obtained average AUC scores of 0.87, 0.90, and 0.88, respectively. Based on the results obtained, this method of diagnosis could serve as an alternative in developing countries to diagnose COVID-19 disease without costly RT-PCR equipment and the expert to operate it.