A Survey on Applications of Artificial Intelligence in Fighting Against COVID-19

1 Introduction

• AI was being explored for medical image inspection, diagnosis, protein analysis, drug and vaccine development, genomic research, and epidemic modeling.Source: §1 · p. 2
• Existing surveys focused on narrower areas, such as forecasting or imaging, leaving the scope of AI research on COVID-19 difficult to see.Source: §1 · p. 2
• The survey maps AI applications across disease detection and diagnosis, virology and pathogenesis, drug and vaccine development, and epidemic and transmission prediction.Source: §1 · p. 2
• The survey also covers available research data and resources, challenges, and potential directions, with Sections 2–5 devoted to the four application areas.Source: §1 · p. 2
• The authors collected 2,471 COVID-19-related publications and resources and selected 443 papers explicitly using AI methods for Figure 1.Source: Figure 1 · p. 3

2.1 RT-PCR Detection

• 96.97% accuracy was obtained when an RF classifier based on 11 key blood-test variables classified 253 samples from 169 suspected COVID-19 patients.RT-PCR is described as the current standard detection technology, but it faces complicated sample preparation, low efficiency, and a high false-negative rate.Source: §2.1 · p. 4

2.2 Medical Image Inspection

• AI-based medical image inspection uses chest X-rays and lung CT to support image acquisition, organ recognition, infection-region segmentation, and disease classification.The survey says these methods shorten radiologists’ imaging diagnosis time and improve diagnostic accuracy.Source: §2.2 · p. 4

2.2.1 CT Image Inspection

• CT inspection pipelines segment regions of interest, extract lung features, detect candidate infection regions, and classify COVID-19; iSARF outperformed LR, SVM, and NN in comparative experiments.CT approaches described include segmentation with U-Net, V-Net, or VB-Net and classification using extracted image or clinical features.Source: §2.2.1 · pp. 4, 6
• A lack of large-scale CT datasets is identified as the main implementation challenge affecting diagnostic performance; augmentation, transfer learning, and radiologist-in-the-loop annotation were used to mitigate data scarcity.The cited workflow iteratively uses radiologist annotations and corrections to complete image annotations.Source: §2.2.1 · p. 6

2.2.2 Chest X-ray Image Inspection

• CXR images are easier to obtain but generally less sensitive than CT, and segmentation difficulty has led most deep-learning models to classify whole images; COVID-CAPS was designed for small datasets.Some early COVID-19 CXR images appear normal.Source: §2.2.2 · pp. 6, 7, 9

2.3 Other Detection Technologies

• Noninvasive approaches include cough and respiratory-pattern analysis, a 20,000-record crowdsourced cough dataset, a model classifying six respiratory patterns, and smartphone-sensor diagnosis frameworks.The smartphone framework collects features such as voice and temperature and sends them to an AI-supported cloud server for analysis.Source: §2.3 · pp. 10–11

3.1 Proteomics

• AlphaFold uses sequence alignments and residue features to predict protein structures, then optimizes generated potentials to produce structures.Its distance-prediction network estimates distances between residue pairs and its output layer predicts accessible surface area, distance maps, and secondary structure.Source: §3.1 · p. 11
• AlphaFold-predicted SARS-CoV-2 membrane-protein structures were published, while docking analyses indicated higher affinity between the spike protein and human ACE2.The predicted structures had not been verified by clinical trials.Source: §3.1 · p. 12

3.2 Genomics

• Six machine-learning classifiers applied to genomic sequence-distance data supported the hypothesis that COVID-19 originated in bats and is a β-coronavirus.The trained models were applied to 29 COVID-19 sequences.Source: §3.2 · p. 12
• AI-based genomics studies also predicted SARS-CoV-2 miRNA-like hairpins and analyzed potential interactions between viral and human genes.The analysis used DT, Naive Bayes, and RF methods to predict hairpin categories and assess similarity to human miRNAs.Source: §3.2 · p. 12
• Metsky et al. used genomic and AI technologies to design an RT-PCR test that detects 67 respiratory viruses, including SARS-CoV-2.The work aimed to improve the current SARS-CoV-2 RT-PCR test.Source: §3.2 · p. 12

4.1 Drug Development

• Beck et al.’s MT-DTI model predicted COVID-19 drug candidates using SMILES strings and amino-acid sequences, while AutoDock Vina predicted binding affinities for 3,410 drugs against SARS-CoV-2 3CLpro.Source: §4.1 · p. 13
• ADQN-FBDD with its Structure-Based Optimization Policy produced 47 potential SARS-CoV-2 3CLpro-inhibiting derivatives from lead compounds.The approach used a fragment library derived from 284 reported 3CLpro inhibitors.Source: §4.1 · p. 14
• AI drug-development studies also generated candidate compounds and screened biological products against viral protein targets [81].Reported approaches included generative models and virtual screening; one multi-task model selected the top 10 potential drugs for each of three principal targets.Source: §4.1 · p. 14

4.2 Vaccine Development

• The survey describes AI vaccine applications as commonly supporting sub-processes, including epitope prediction and mapping.netMHCpan predictions included 2,915 alleles across peptide lengths, and one associated tool provided genetic-variation and epitope-analysis functions.Source: §4.2 · p. 15
• Reverse vaccinology and machine learning were used to assess SARS-CoV-2 proteins and predict vaccine candidates, including T-cell and B-cell epitopes.The analysis used Vaxign and Vaxign-ML and evaluated nsp3 for phylogenetic and immunogenicity features.Source: §4.2 · p. 15
• DeepNovo used patient HLA alleles and immune-peptide information to identify candidate T-cell epitopes for peptide-based COVID-19 vaccines [161].The authors suggested designing an epitope-based vaccine for each person based on their HLA alleles.Source: §4.2 · pp. 15–16
• Other vaccine studies predicted immune stimulation and toxicity, including a multi-epitope vaccine combining S, M, and E protein epitopes and evaluation of 174 SARS-CoV-2 epitopes.The cited work used 19 epitope-HLA combination prediction tools, including IEDB, ANN, and PSSM algorithms.Source: §4.2 · p. 16

5.1 Patient Mortality and Survival Rate Prediction

• An XGBoost model using blood samples from 404 Wuhan patients predicted individual survival with 90% accuracy; higher LDH levels seemed important in distinguishing critical cases.The model selected three important biomarkers.Source: §5.1 · p. 17

5.2 Outbreak and Transmission Prediction

• BlueDot predicted the COVID-19 outbreak and high-risk cities nine days before the official outbreak announcement.It combined heterogeneous data sources with filtering, machine-learning, and natural-language-processing techniques.Source: §5.2 · p. 17
• Studies modeled COVID-19 cases and spread across China using deep-learning forecasts, provincial clustering, and SEIR and LSTM approaches [82].The deep-learning models used the previous five days of case features to predict cases in the following days.Source: §5.2 · p. 17
• Both SEIR and LSTM models predicted a daily infection peak of 4,000 in the first week of February; the LSTM was pretrained on 2003 SARS data.Epidemiological parameters were selected as input features for the LSTM.Source: §5.2 · p. 18
• AI forecasting approaches outside China included case prediction, spatial-temporal modeling, and simulations of individual contacts and transmission routes [170].The survey describes applications in the United States, Italy, Spain, Iran, and Switzerland.Source: §5.2 · p. 18
• Global transmission studies used travel, mobile-phone, and online-media data to estimate spread risks and transfer epidemic models between countries.Other studies applied machine-learning and deep-learning regression models to real-time case prediction.Source: §5.2 · p. 18
• Statistical analyses reported peak viral shedding before or during symptom onset and that some transmission may occur before initial symptoms.Other statistical and epidemiological models assessed epidemic turning points and COVID-19 spread [171] [200] [204].Source: §5.2 · p. 19

5.3 Social Control

• AI-enabled social-control applications included temperature and mask detection, video and contact tracking, and analysis of travel trajectories.Studies also modeled the effects of contact tracing, quarantine, and other interventions on transmission.Source: §5.3 · pp. 19–20

6.1 Medical Images

• The survey gathers 17 groups of medical image resources, including CXR and CT collections from researchers, Kaggle, and imaging organizations.Cohen et al.’s CXR collection is described as widely cited.Source: §6.1 · p. 20

6.2 Biological Data

• Ten biological resources provide gene sequences, proteins, drug molecules and compounds, and miRNA sequences.Examples include NCBI, PDB, UniProt, CAI, DTC, and VHDB.Source: §6.2 · p. 20

6.3 Cough Sound Datasets

• Eight cough-sound datasets are collected, but such datasets are scarce and only five include confirmed COVID-19 patient coughs.The survey hopes these resources will facilitate cough-based diagnosis.Source: §6.3 · p. 21

6.4 Informatics Resources

• Informatics resources include situation reports, dashboards, case data, and demographic data from public agencies and online providers.WHO, NHC, and CDC provide real-time reports; Baidu, Microsoft Bing, Google, and JHU CSSE provide spread-tracking dashboards.Source: §6.4 · p. 21

7 Challenges and Future Directions

• The survey characterizes AI’s application and contribution to COVID-19 research as still relatively limited and summarizes challenges with corresponding suggestions.It also identifies 11 potential directions for further AI contributions.Source: §7 · pp. 21–22, §7.2 · p. 23

7.1 Challenges

• Large-scale annotated training data are insufficient after the rapid outbreak, and annotation is time-consuming and requires medical professionals.The survey identifies this as a challenge for AI methods that rely on large-scale training data.Source: §7.1 · p. 22
• Positive COVID-19 samples are limited relative to negatives, affecting the accuracy and robustness of AI algorithms in diagnosis.Negative samples are easier to obtain from historical publications than positive samples.Source: §7.1 · p. 22
• Unreviewed noise and fake news online limit AI accuracy and application, especially in epidemic prediction and transmission analysis.The survey says AI algorithms seem unable to judge and filter noisy or erroneous data effectively.Source: §7.1 · p. 22
• COVID-19 AI work requires collaboration across computer science, medical imaging, bioinformatics, virology, and other disciplines.The survey calls for integrating knowledge from multiple subjects.Source: §7.1 · p. 22
• Personal location, travel, and patient-disease data raise privacy and human-rights concerns during information acquisition and AI-based processing.Source: §7.1 · p. 22

7.2 Future Directions

• Potential directions include non-contact imaging, few-shot and transfer learning, remote diagnosis, prognosis monitoring, and AI-based biological research.The survey describes non-contact image acquisition as a way to avoid infection risk between radiologists and patients; few-shot and transfer learning address limited or imbalanced data.Source: §7.2 · p. 23
• Other proposed directions cover drug and vaccine development, fake-news filtering, impact simulation, contact tracking, intelligent robots, and AI-enabled Internet of Things monitoring.Source: §7.2 · p. 23

• The survey groups AI applications into disease detection and diagnosis, virological research, drug and vaccine development, and epidemic and transmission prediction.Source: §8 · p. 24
• Medical image analysis, drug discovery, and epidemic prediction are identified as AI’s main COVID-19 research battlefields.Source: §8 · p. 24
• The survey summarizes available medical imaging, biological, and informatics resources for AI-based COVID-19 research.Source: §8 · p. 24
• The survey highlights challenges and potential directions and aims to give medical and AI researchers a comprehensive view of current and potential contributions.Source: §8 · p. 24