Predicting COVID-19 Infections in South Africa Using Deep Learning

Master of Science in Computer Science, North-West University,vanderbijlpark Campus

In thiswork,themainaimwastoemploydeeplearningtechniquesforpredictingcoro- navirusdisease2019(COVID-19)infectionsinSouthAfrica,withafocusonenhancing the predictivecapacityofCOVID-19virusnewinfections.SincethesurgeofCOVID- 19 infections,therehavebeenconcernsaboutthepotentialeffectsitmighthaveinthe healthcare sector,particularlyinefficientlymanagingresourceavailabilityandallocation. VariousresearchersconducteddifferentstudiestopredictthetransmissionofCOVID-19 infectionsusingtechniquessuchasstatisticalpredictiveanalysis,mathematicalmodel- ing, andmachinelearningapproaches.Studiesusingmathematicalandstatisticalmod- eling hadlimitationsinpredictingfutureinfectionwavesastheywereunabletopredict future wavesofinfections,unlikethoseusingmachinelearning.Hence,themainaimof undertakingthisstudywastoexplorethefeasibilityofemployingaspecificsubsetofma- chine learning,specificallydeeplearning,topredictCOVID-19infectionsinSouthAfrica. The COVID-19datasetusedtobuildthedeeplearningmodelswassourcedfromanopen data repository.Toachievetheaim,thefirststepwastotrainthebasiclongshort-term memory(LSTM),stackedLSTM,andbidirectionalLSTMmodels.LSTM-basedmodels havedemonstratedremarkablecapabilityincapturingtemporaldependencies,whilethe stackedLSTMandthebidirectionalLSTMarchitecturesenhancethiscapabilitybyincor- poratingadditionallayersandbidirectionalinformationflow,respectively.Thenextstep wastoempiricallytestthetrainedmodelsonaCOVID-19realdatasettochecktheirper- formancesbasedontherootmeansquareerror(RMSE).Apersistencemodelthatwas used asabaselinemodeltoevaluateperformancesforthecomplexmodelswasintro- duced. Comparingthepredictionsofthetrainedmodelswiththebaselinemodel,onlythe basic LSTMmodelhadanRMSElowerthanthepersistencemodel.Lastly,identifygaps and challengesinthisstudyandproviderecommendations.Theprimarychallengesand gaps identifiedincludethefactthatthepredictionsweremadeforSouthAfricaoveralland did notcatertosubpopulations.Thisstudyconcludedthattheproposeddeeplearning models exhibitedenhancedpredictabilitycapacityofCOVID-19newinfectionscompared to previousstudiesthatwerereviewed.Additionally,thebasicLSTMmodelwasfoundto be thebestmodelforpredictingCOVID-19infectionsinSouthAfricaasithadthelowest RMSE.

Masters