Using Machine Learning for Healthcare Resource Allocation in COVID-19: Opportunities and Challenges for LMICs

One of the greatest challenges for any health system facing COVID-19 is the equitable and efficient allocation of scarce resources. Even well-resourced health systems, such as that of Italy, struggled to quickly allocate vital healthcare resources like beds and staff to areas where they were most needed. As vaccines and therapeutics emerge, the efficient allocation of resources will become increasingly important. Although high-income settings were worst affected by the pandemic initially, countries such as India, Russia, and Mexico now lead the world in new daily cases.

Artificial intelligence (AI) has been widely applied in COVID-19, for early detection of disease, monitoring patients, contact tracing, as well as the development of drugs and vaccines. Machine learning (ML) is a subset of artificial intelligence (AI) that provides systems the ability to automatically learn and improve from experience, making predictions or decisions without being explicitly programmed to do so. Given the explosive nature of COVID-19 and the phenomenon of asymptomatic transmission, digital systems that use real-time data, improve over time, and provide adaptive feedback, will be vital to mitigate the impact of COVID-19. With health systems stretched, and concerns about COVID-19 vaccine nationalism, machine learning presents a valuable opportunity to help guide decisions on the allocation of scarce resources like hospital beds, staff, and vaccines, in low- and middle-income countries (LMICs).

Opportunities

Rationing hospital beds

Machine learning has been utilized in high-income countries to help better ration resources for COVID-19. In the UK, a team at the Cambridge Centre for AI in medicine developed a system called Cambridge Adjutorium, which uses a state-of-the-art ML framework to accurately predict the rate of mortality, ICU admission, and the need for ventilation in hospital patients with COVID-19. The system, even though trained using very small datasets (CHESS data from Public Health England), had a high accuracy rate, ranging from 77 percent to 87 percent. Another ML-based model using electronic health records in the US analyzes patients’ data and assigns them a score based on how sick they are and how likely they are to need escalated care. Fu-Yuan Cheng et al. similarly developed a ML-based risk prioritization tool for COVID-19 patients used for identifying patients with an increased need for ICU transfer.

Although such tools have not yet been widely used in LMICs, they have huge potential. India has approximately one ICU bed per 13,684 people compared to the US with one ICU bed per 3,398 people. Many LMIC health systems are dominated by large private sectors, further limiting the availability of affordable and accessible hospital resources. Since the pandemic began, private sector engagement has proliferated, and is actively encouraged by the WHO, to help mobilize and coordinate resources in often-fragmented systems. The South African government, for example, has agreed to pay a daily fee of up to 16,000 rand ($950) for COVID-19 patients who are treated in ICU beds in private hospitals. ML-based risk stratification tools, through reducing the unnecessary use of private beds, may help to preserve public sector finances at a time of significant global economic downturn.

Workforce Planning

ML can also be used to guide workforce planning. Health professionals are integral to the response to COVID-19. There are typically fewer health professionals in urban poor and rural areas, and this problem is particularly stark in LMICs, raising concerns about the spread of COVID-19 in rural areas. In Zambia there are 20 times more physicians in urban than in rural areas, despite having a predominantly rural population. Cities have been the focus of a number of large COVID-19 outbreaks until now. But with individuals unable to continue working in urban areas and moving away from cities, many experts are worried about COVID-19 emerging in largely isolated rural areas that lack the qualified health professionals they need.

ML-based risk monitoring systems can analyze the electronic health records of patients in a hospital, or cases identified through COVID-19 testing, to predict future clinical and public health staffing requirements. Identifying the geographical areas in which staff shortages are most likely to develop will be valuable for policymakers given that outbreaks are now widely distributed across many countries and are rapidly changing. Many countries have reacted by scaling-up their existing medical and public health workforce. In India, with more than a quarter of India’s 736 districts having no district-level epidemiologists, authorities began a rapid search for epidemiologists in April. ML can be linked with adaptive online assessments to help predict how well suited people are to the job they are applying for. This may be particularly useful in an epidemic, where the rapid need for expanding workforce capacity risks the hiring of poorly qualified candidates.

COVID-19 vaccine

Another major issue that countries are likely to face in the near future is ensuring adequate uptake of a novel COVID-19 vaccine. It is possible that a COVID-19 vaccine will require more than one dose, or boosters to maintain immunity over a long period of time. This will be a huge challenge in LMICs. In 2018, it was estimated that the average dropout rate for the second dosage of BCG vaccine was 34.6 percent in low-income countries. Traditional factors responsible for poor vaccination uptake and completion will be exacerbated by the fact that a novel COVID-19 vaccine will have a limited safety and efficacy profile. ML can be used to identify and target people who are less likely to actively seek or complete vaccination. Subhash Chandir et al. used predictive analysis to identify children who were at a higher risk of missing routine immunization appointments. They used variables like gender, place of residence, vaccine, language, timeliness of the vaccination, vaccinator, date of birth, and age to predict how likely it was that a child would miss their follow-up vaccination. The system had a relatively high accuracy of 79.1 percent. Approaches like this can be utilized to improve future COVID-19 vaccination programmes, given the wealth of data that will likely be available for a vaccine in such high demand. When aiming for herd immunity in LMICs, it will also be important to focus public health efforts on areas where vaccination is needed most. A group of researchers at the Facebook Boston office used ML-powered maps to filter out 97 percent of uninhabited terrains in Malawi. They then used the remaining 3 percent of inhabited terrains to focus their vaccination efforts. With 66 percent of India’s population living in rural areas that can be hard to locate and access, ML-based maps can be valuable in LMICs to help better allocate vaccines to vulnerable populations that may otherwise be missed.

Challenges

Infrastructure

Applying ML in these ways will be a huge challenge for LMICs, with a number of costs arising from investing in expensive technology that relies on the availability of basic infrastructure. It has been estimated that less than 30 percent of the health facilities on the continent of Africa have reliable electricity, and there is a very low level of internet penetration across the continent (39 percent). In many low-income countries, it will be necessary to remove these barriers before digital tools can be used to aid resource allocation in the fight against COVID-19.

Data gaps

COVID-19 has highlighted many data gaps in LMICs, meaning a priority for COVID-19 control must be generating better information using real-world data to inform modelling forecasts for evidenced-based policy. But data on individual health is both expensive and time consuming to collect. There is generally a low level of digitization of the available health data in LMICs because medical records are often handwritten in local languages. Due of the unavailability of datasets, the few ML systems developed in LMICs often have to rely on the publicly available datasets from the US and Europe. Since ML systems are only as good as the data they are trained on, this approach could be damaging.

India provides a great example of how data can be locally generated and used to improve systems. Following a nationwide rollout in 2011, the Aadhaar programme has registered over 1.2 billion individuals on its biometric database. This data can be linked with mobile phones, bank accounts, insurance policies, income tax, pensions, and welfare programmes. Mobilizing existing digital infrastructure, such as Aadhaar, to support the development of ML systems in healthcare may be a realistic next step to aid the COVID-19 response in some LMICs.

AI and health workforce

A survey of local professionals in Pakistan revealed that a lack of trained professionals was the most commonly cited barrier to using AI in healthcare. Using ML to aid resource allocation decisions in COVID-19 will require professionals who are trained and prepared to fully embrace digital tools. Effective, sustainable support systems and training, as well as clarity on roles and accountability for decision-making and supervisory structures, are essential. The cost of embedding ML and other digital tools into LMIC health systems initially may therefore be substantial. In northern Ghana it was estimated that it would cost $1,060 per health worker to implement a computer-assisted clinical decision support system for antenatal and delivery care alone.

Improving supportive infrastructure, the digitization and linkage of health records and public health intelligence, as well as training and engaging professionals, will be difficult. But with large development actors and LMIC governments making substantial investments into the COVID-19 response, alongside an increased focus on digital health at an international level, the pandemic may serve as a critical juncture, where some of the traditional challenges associated with integrating technology into LMIC health systems are finally overcome.

Conclusion

Low-and middle-income countries must make the most of scarce resources if they are to mount a successful response to COVID-19. Given the possibility of asymptomatic transmission and rapid local spread, as well as the continued surge in cases in LMICs, ML tools present a valuable opportunity through using real-time data to inform decisions on the future allocation of scarce healthcare resources. Effectively embedding technology into health systems and decision-making processes has traditionally been challenging in LMICs. But there are many recent examples of progress with a number of LMICs already building the infrastructure needed to implement national digital health strategies, prior to the arrival of COVID-19. The pandemic provides a chance for policymakers in LMICs to double down, build on this progress and vitally aid the public health response to COVID-19.