The role of machine learning in healthcare responses to pandemics: maximizing benefits and filling gaps

Introduction

The COVID-19 pandemic numbers have finally declined¹, but not without causing tremendous human and financial losses. Healthcare responses varied across countries, with many health systems becoming overburdened and overwhelmed². At the outset of the crisis, uncertainty surrounding symptoms and prognoses as well as a lack of pharmaceutical treatments or vaccines presented significant obstacles in enacting timely yet efficient responses, which underscores the need for improved preparedness plans in our fight against future pandemics².

One potential strategy to combat future pandemics is to utilize the most recent technologies, particularly machine learning (ML), a subfield of artificial intelligence (AI)³. ML utilizes complex algorithms and statistical models to analyze vast amounts of data, detect patterns and anticipate outcomes³. ML has significantly progressed with the introduction of more advanced approaches such as Deep learning, which is a subfield of machine learning that has gained significant attention and advancements in recent years. Deep learning models, inspired by the structure and functioning of the human brain, are designed to automatically learn and extract meaningful patterns and representations from data. These models consist of multiple layers of interconnected artificial neurons, enabling them to capture intricate relationships and hierarchies within the input data. By training on large datasets, deep learning algorithms can uncover hidden patterns, identify potential risk factors, predict disease outcomes, and assist in drug discovery and development.

During the COVID-19 pandemic, ML was applied in various domains such as computational epidemiology; diagnosis; hospital and intensive care unit admission predictions; drug discovery; and repurposing³. Having said that, ML also has limitations such as the need for advanced computer clusters and unbiased data for processing (which can be partially overcome by careful data collection and preprocessing techniques to minimize biases and ensure a diverse and inclusive dataset). Furthermore, there are still some challenges to overcome, such as the “black box” problem associated with deep learning methods, which makes it difficult to explain the rationality of the results. Therefore, there is a scarcity of research on the real-world effectiveness of ML in improving the responses of healthcare systems to newly-emerging pandemics. This editorial investigates the potential benefits of ML implementation via exploring its effect across various healthcare domains and proposing practical solutions to overcome challenges faced by health systems to adopt efficient and swift pandemic response.

Discussion

Machine Learning (ML) has emerged as a promising technology for improving pandemic response, offering potential contributions across various fields including epidemiology, clinical practice, and drug discovery and repurposing (Figure 1).

Figure 1. Applications of Machine Learning in Pandemic Response.

Machine learning in epidemiology and clinical practice

ML models can utilize not only standard infectious disease data but also unstructured information issued by governments, social media, and international health organizations. The applications of ML in tackling a pandemic include disease diagnosis and prognosis⁴. ML models can analyze and interpret large quantities of lab tests, leading to reduced human error and the speed and accuracy of diagnosis. ML algorithms can also identify patterns in medical images and disease symptoms. From more than 300 clinical characteristics, a group of researchers built an ML model and identified three crucial clinical features: high-sensitivity CRP, lymphocyte count and lactic dehydrogenase. With an accuracy rate exceeding 90%, the model was able to predict COVID-19 patient survival rates⁵.

The unique ability of ML to “learn” how to combine data from diverse and complex data sources can provide more accurate predictions of the pandemic spread and assessments of disease severity. Al-Raeei et al.⁶ applied the susceptible-infected-recovered-dead (SIRD) model using data from the United States, Russia, and Syria, and found some markers regarding the spread of the COVID-19 pandemic, mortality, and recovery rates. The development of a long short-term memory (LSTM) model to forecast transmission dynamics of the pandemic in Canada by Chimmula et al.⁷ represents another example for ML implementation. Furthermore, ML assists in disease control and slowing the spread of the pandemic by rapidly detecting potential infection networks.⁴ It involves using the information on an infected individual’s diagnosis time and close contacts’ locations to predict how and when the virus might spread. Doing this enables quicker actions to contain outbreaks by identifying likely sources of infection or super-spreaders. Wang et al.⁸, developed an ML model which classified Taiwan’s population into low, moderate and high-risk groups based on health and social histories; those with higher risks were quarantined at home; ultimately, infection rates were much lower than expected.

ML In mental health

ML can be employed to help mitigate the psychological effects of a pandemic⁴. The ability of the ML to identify high-risk individuals who could benefit from mental healthcare services or create tailored interventions to address specific mental health issues represents another advantage from its implementaion⁴. Furthermore, ML can be helpful in the development of anonymous support platforms, personalizing interventions, and automating mental health screening, which have the potential to reduce stigmatization. Anonymous online platforms enable individuals seek help without being judged, personalized interventions customize treatment plans according to each individual’s needs, and automated mental health screening helps in early detection to reduce the stigma associated with late-stage diagnosis.

However, in mental health, there are several issues that need to be addressed to achieve maximum utilization in machine learning. These include data privacy, representativeness, interoperability, quality and reliability. For example, it is challenging to integrate and analyse mental health data that often scattered across various healthcare systems and platforms

ML In drug discovery and drug repurposing

Recent advancements in technology have greatly impacted the field of drug design and discovery, allowing for the reduction of associated costs, time, and work demands. The virtual screening of ligand libraries has seen a significant increase in efficiency, with the ability to screen one billion ligands in a matter of days or weeks⁹. This improvement in efficiency is due to the rather increased availability of data, the development of memory capacities, the introduction of supercomputing power such as GPUs and parallelization, and the advancement of theories behind ML methods¹⁰. In fact, ML has been successfully employed in various stages of drug discovery (e.g. virtual screening), allowing for the screening of larger chemical space and reducing false positives. These methods can be utilized as filters prior to docking and can provide valuable information on which scaffolds to retain in the ligand library. For instance, AI-based virtual screening has been developed using “deep docking” to significantly reduce the size of ligand libraries. Gentile et al. recently conducted a study where they utilized an AI-based virtual screening method that uses “deep docking” to only screen 1% of ligands, resulting in a significant reduction in the size of the 1B-ligand library by 100-fold¹¹. In a follow-up study, they applied the same method to target a SARS-CoV-2 target, namely main protease (M^pro) and screened a ligand library of more than 1.3B ligands. Their AI-based approach successfully identified a novel series of compounds that were later confirmed as M^pro inhibitors with IC₅₀ values in the low micromolar range¹².

Machine learning has also been crucial in drug repurposing research, particularly since the emergence of COVID-19. Drug repurposing involves identifying and investigating novel uses of drugs or drug combinations based on their existing pharmacological properties¹³. Compared to the traditional drug discovery strategy, drug repurposing is more efficient and cost-effective, and it has additional advantages in combating pandemics. Several drugs, such as chloroquine, phosphoate, and remdesivir, have been evaluated for their therapeutic effect on COVID-19. More interestingly, scientists worldwide have utilized machine learning-based approaches to identify effective drugs. For example, Zeng et al.¹⁴. developed a deep-learning methodology called CoV-KGE to identify repurposable drugs for COVID-19 by building a comprehensive knowledge graph that connects drugs, diseases, proteins/genes, pathways, and expression from 24 million PubMed publications. Using a combined approach of the network-based and deep-learning framework along with the use of Amazon’s AWS computing resources, the study successfully identified 41 drugs of potential anti-COVID-19 activity (e.g. dexamethasone, indomethacin, niclosamide, and toremifene) that were validated by experimental data and preliminary findings of ongoing clinical trials. Hence, the use of machine learning in drug repurposing can accelerate the drug discovery process and improve human health.

Therefore, machine learning and AI are emerging technologies that will aid in drug repurposing and scanning larger portions of the vast chemical space with a quicker pace, lower cost, higher efficiency, and fewer false positives. Hence, developing machine learning methods can accelerate drug discovery and improve human health in ways that were previously impossible.

ML Challenges

The road to implementing ML in healthcare is filled with technical, practical, and security challenges. Taking into consideration that ML requires high-quality data¹⁵, patient confidentiality and privacy must be ensured. Therefore, the use of advanced de-identification techniques which remove personally identifiable information from patient data, using encryption protocols during data transmission, and adding extra noise to patient data using privacy-preserving techniques, in order to ensure data safety is deemed necessary Furthermore, health systems should ensure transparency, accountability, and routine data security audits while ML is active. Given that ML algorithms are complex and may be challenging for clinical teams to use¹⁶, training programs on the integration of health information technologies into clinical processes and communication channels between the clinical teams and ML researchers should be a focus in future health plans. In order to produce accurate predictions, ML models require vast amounts of data that need to be unified in format, type, and label code to ensure consistency, which represents another challenge for its implementation. Additionally, the accuracy of ML outcomes is significantly impacted by inaccurate or irrelevant data, which could result in misleading recommendations¹⁷. During a pandemic, data need to be collected and processed quickly. This highlights the need for intelligent devices that can offer real-time data. Finally, creating ML models that work across different populations is not feasible due to a lack of standardization in data formats and varying data protection regulations across countries¹⁶. One potential solution is to drive global efforts toward standardizing data formats and privacy laws, creating clear guidelines for data ownership and intellectual property rights, as well as offering organizations incentives to share their data through research grants or tax breaks. By promoting data sharing, we can support the development of more precise and efficient ML models, leading to significant progress in healthcare response to pandemics.

Conclusion

By adopting ML in epidemiology, clinical practice, and drug discovery and repurposing, health systems’ responses to pandemics have a greater chance of being significantly improved. Specifically, ML has the potential to predict the most efficient and cost-effective strategies to prevent the spread of the pandemic. ML can also inform policies and decisions about the proper allocation of resources, ultimately leading to saved resources, reduced workload, improved mental health for healthcare workers, and better patient outcomes. Additionally, ML can accelerate drug virtual screening and prioritize drug candidates.

Transparency

Author contributions

AZA, AJ, and MG have contributed equally to the study design development, data extraction, manuscript drafting and reviewing.

Reviewer disclosures

Peer reviewers on this manuscript have no relevant financial or other relationships to disclose.

Acknowledgements

None reported.

Declaration of funding

The paper was not funded.

Declaration of financial/other relationships

The authors have no relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript. This includes employment, consultancies, honoraria, stock ownership or options, expert testimony, grants or patents received or pending, or royalties.