Improving healthcare equity through artificial intelligence

Dr. Shakthi Kumar, DHA, MBA

Worldwide Head of Healthcare Solutions at Amazon

Improving healthcare equity through artificial intelligence

1 June 2022 | 12mins

Quick Takes

  • As artificial intelligence (AI) proves itself invaluable in the healthcare industry, leaders need to ask how it will increase patient access to treatment and improve outcomes

  • AI and machine learning (ML) models rely on representative datasets for self-learning that have inherent biases which need to be considered for more accurate results

  • Through partnerships and collaboration, healthcare leaders are using the technology and pushing the boundaries to improve healthcare equity

The exponential growth of artificial intelligence (AI) and its much-touted form of machine learning (ML) has proven invaluable in many areas of patient care – from detection of malignant, hard-to-detect tumors in Barrett’s esophagus to cardiac monitoring through wearables to predicting the severity of COVID-19 of critically-ill patients in their ICUs.1-3 With COVID-19 accelerating the adoption of AI, constrained health systems and biomedical organizations are exploring ways that AI can improve operational, clinical, and research networks to provide better healthcare access, treatments and outcomes. 

Most AI and ML enthusiasts will concede that while there is a huge promise and tremendous potential to deliver new solutions for possibly hundreds of much-needed patient interventions, there are still a lot of pitfalls to be considered. This is clearly evident in the lack of equity within healthcare systems, in areas such as ethics, data representation, and governance. These need to be sorted out as we scale AI across our world of healthcare.

Here, we discuss the growth of healthcare AI and what leadership and executives need to consider to improve equity in the health system.

The exponential growth of healthcare AI and ML

Within the healthcare industry, the adoption of AI and ML has skyrocketed in recent years with an exponential increase in the number of use cases and applications. It is estimated that the healthcare AI market will reach $USD 26.6 billion by 2025, growing at a CAGR of 41%.4 Furthermore, venture capital raised in health tech was $USD 14 billion in 2020, with 33 startups raking in $USD 100 million or more in late-series funding (C round or higher).5 

While there is this heightened level of market activity, the level of real-world adoption in medical care is still modest, and mainly emerging through non-clinical use cases for administrative operations such as health insurance billing, coding, revenue cycle management, and compliance. Therefore, there is an increased need for AI and ML supporters to push funding in all areas of patient and clinical care, including to improve healthcare equity.

The 4th Industrial Revolution and the growth of inequity

At each stage of every industrial revolution, we have applied scientific advancement through technological applications largely with the intent to benefit our world. While there have been significant advancements at every stage in medical technology, care models, and treatments, the design thinking of the technological applications has largely been utility-driven. How to offset and dispel any of the resulting socio-economic disparities and inequities that have cascaded and even aggregated across stages could perhaps best be described as an afterthought. 

We are presently pacing through the fourth iteration of the industrial revolution (4IR) with the deployment of cyber systems, physical networks, internet-of-things, and AI.6 For healthcare stakeholders, it is now more important than ever to frame the right questions and define solution parameters before we go running off to uncover the answers. We need to address some fundamentals before addressing use cases. 

Some important questions that need to be asked include:

  • How can AI and ML be used to help offset some of the disparities that have accumulated in our healthcare models?
  • Is there a possibility to help dispel some of these inequities through the novel application of AI and ML?
  • What are the parameters of normative governance and ethics code systems we need to adhere to in healthcare as we scale up this powerful technology?

With these foundational questions in hand, leaders need to take this opportunity to help address some of the basic systemic and structural deficiencies that pervade our healthcare systems. We now have a tremendous opportunity as part of 4IR to offset some of the global socio-economic disparities, exploitation, and inequity stemming from the prior industrial revolution stages, increasing medical access, treatments, and outcomes. 

Healthcare inequity – the evidence  

It is commonly acknowledged that the advent of modern medicine has resulted in remarkable strides in the diagnosis and treatment of disease. In the US, we have recorded increases in life expectancy and a decline in activities of daily living (ADL) disability prevalence that show longer lifespan and improved functional mobility. 

However, over the past few decades, we notice increasing chronic disease scores and physiological status measures (markers for inflammation and glucose levels) that show a reverse trend, and these results are further skewed by geography, socio-economic differences, and race. For instance, developed countries like the US and European nations report higher increases in life expectancy and decreases in ADL disability prevalence than low-and-middle-income countries (LMIC).11,12 

In healthcare access and outcomes indicators, we also find stark inequities. For instance, the proportion of remaining life spent in good health is higher for men than women, for caucasian people compared to racial and ethnic minorities (in the US), and for the most educated versus the least educated.13,14 

Many of the thought frameworks such as the hypothesis of “dynamic equilibrium,” the theory of “persistent inequality” and the theory of “cumulative disadvantage” lead us to the thinking that some of these disparities and inequities in access to healthcare will continue and widen in advanced age without systemic interventions. Cohort-based studies also support the same thinking making healthcare inequity a real and vexing issue for public health, policymaking, societies, and humans at large.11

The same disparities and inequity also can be observed in the field of clinical research. For instance, African American women show a 41% higher mortality from breast cancer compared to white women but only represent 2-6% of the clinical trial participants.15-17 

African American men have a 76% higher incidence rate and 2.2 times higher death rate from prostate cancer compared to white men, but less than 5% of the participants enrolled in these studies are black.18,19 Additionally, it has been reported that 35% of black adults are hesitant to get the COVID-19 vaccine compared to 27% of the general U.S. public.20 

The contributing factors for prevailing inequities are a complex web of socio-economic and cultural factors, lack of education, awareness, and in some cases, pure apathy. Furthermore, while ethics, governance, and trial design, like randomized clinical trials, are structured to avoid these biases and have greater diversity and representation in the study population, the results show much progress still needs to be made. 

With such pervasiveness of inequity, the unfortunate reality is that our present healthcare ecosystem and its infrastructure (health systems, medical research, delivery models, therapies, caregivers, access, outcomes, information systems), and most importantly the “data” all inadvertently carry and propagate some form this inherent bias and skewness.

Why is this important to AI and ML? 

Challenges with current AI and ML models

Current AI and ML models heavily rely on representative datasets for self-learning and having an inherent bias in these sources causes inconsistent and skewed results. There may also be potential biases that have been introduced by the human creators of these AI and ML algorithms, and this compounding effect results in a muddled mess. 

Furthermore, any decisions and scores predicted by these models are ultimately consumed by our prevailing delivery models and social-economic infrastructure that have the potential to further aggravate the bias resulting in lopsided and prejudiced outcomes. 

Of course, health inequity has always been and will be a multifactorial outcome and not everything can be solely attributed to technologies, like AI and ML. We will obviously need other forms of interventions that create a more widely distributed health system infrastructure with access to all as well as greater diversity within healthcare and research professionals, social development programs, improved living conditions, better education, and participation from all stakeholders. 

How removing bias in AI and ML systems will improve healthcare equity

As our world of healthcare expands and evolves, its reliance on digital technologies, including AI and ML will continue to grow. The data and associated set of AI and ML algorithms that are employed will drive many of the decisions undertaken as part of patient care and care administration. 

By improving the completeness of data (to be more inclusive and diverse) and designing AI and ML algorithms that are free from bias, we can more effectively impact how we deliver care, how it is accessed, and how we measure outcomes. This can become a big improvement over prevailing models where inherent biases drive many characteristics of healthcare and perpetuate health inequity. 

Thus, more inclusive data sets and better designed AI and ML algorithms can help offset such imbalances. In addition, AI and ML also have the potential to drive personalized care decisions based on each individual characteristics (social, clinical, economic) if the training data sets include all of these nodal variances, and algorithms are then built to more fairly account for treatment decisions based on such differences. 

Considerations for healthcare leaders

As we develop our AI and ML models for the myriad of healthcare applications, we have an opportunity to be conscientious of not inducing any additional bias and explore opportunities to address both epistemic and normative concerns present in our systems and decision-making.

Given that lot of the representative and training data sets in healthcare used by AI and ML models already have a certain element of inherent bias, the design and applications, especially in unsupervised and deep learning models, may experience a level of apophenic effect and misapplication.22  Hence, some areas that executives can improve upon include:

1 Understanding the social situations that drive decision making – build applicable models irrespective of the social situations under which it is being delivered. For instance, care commonly administered in hospital systems for the general population vs. privileged care in private hospitals
2 Applying many decision frameworks across multiple demographic sets and social determinants of health (SDoH) – accounts for differences in patient population and need to include social determinants as a criteria for care. For example: how to deliver care to under-represented and poorer sections of our society while considering economic and social constraints of access to such care
3 Making direct use of the human decision – making models with the required contextual augmentation in machine models – incorporate decision-making factors of how clinicians and care providers may use the AI and ML models for care purposes. For instance, if an AI/ML model computes a cardiovascular risk score for patients in minority populations, incorporate the right mix of causal variables that provide adequate context for such decision making
4 Utilizing representative datasets – utilize broad and wide-ranging sets of data to train AI and ML models. For instance, obtain data sets for the minority and under-represented patient populations from several sources (other than ones available in the environment where it is developed or being deployed) to enhance the richness and sensitivity of the AI and ML models to differences in each population
5 Ensuring ethical governance and review – employ robust governance and vigilance to monitor how AI and ML models are being developed and deployed to eliminate the introduction of additional bias and propagation of prevailing ones. For instance, implement systematic surveillance and vigilance on how AI and ML algorithms are designed for various types of data (and its subjects), how it is deployed, and how they perform in real-world conditions 

Many health systems serving demographics that have a larger proportion of underrepresented populations can act as rich sources for creating diverse and well-represented datasets. Obviously, the process of carefully curating this data with required medical annotations without compromising patient privacy is time-consuming and expensive, but it needs to be done. 

Additionally, there are opportunities to leverage AI and ML to extend the realm of research datasets through carefully modeled digital mirroring as an option. But more research needs to occur to ensure these human-modeled digital datasets do not induce and propagate further unconscious bias as part of the design process.

Change in action through collaboration

While our ability to construct widely representative demographic data to train AI and ML models remains a challenge, the industry is starting to come together to help address this issue through partnerships and data sharing. 

For instance, the National Institutes of Health (NIH) has programs to address health inequity challenges and recently funded the formation of a consortium aimed at advancing health equity and researcher diversity.21 The World Health Organization (WHO) has several initiatives to promote health and reduce health inequities by addressing the social determinants of health.22 

There are also many health systems that have specific programs that focus on improving access to all communities through health clinics and educational campaigns. The Mayo Clinic platform for collaboration is an example of how conscientious health systems can drive healthcare equity and improve care for underrepresented segments of our population by using widely represented data and more inclusive AI models.23,24 Likewise, many life sciences organizations actively participate and have dedicated initiatives designed to improve access as reflected in the access to medicine index (ATMI) that measures their contributions.25

It is also commendable that many healthcare organizations that act as powerful bastions of precious annotated medical data are willing to open them for the larger benefit. We are now moving in the right direction to address these inequities through a slew of public and private initiatives. 

For instance, Stanford University Center for AI in Medicine & Imaging (AIMI) in collaboration with Microsoft’s AI for Health program is creating an open forum to both share its own privacy-protected medical data and also build a repository of aggregated datasets from other sources.27 

Ethics and governance of AI and ML models, especially in healthcare, are critical. Powerful geopolitical blocs and the industry have been making improvements in this area. 

The Organisation for Economic Co-operation and Development (OECD) passed its AI and ML guidelines in 2019 and followed up with an ethics framework in June 2021.28 Furthermore, the U.S. Food and Drug Administration (FDA), Health Canada, and the United Kingdom’s Medicines and Healthcare products Regulatory Agency (MHRA) recently strengthened their guidelines by issuing ten principles for AI and ML models in medical devices, which sets a promising path for future development.29

Improve healthcare equity through AI and ML is plausible

The hope is bright and the opportunity to leverage this powerful platform of AI and ML to improve healthcare equity is real. However, everyone from researchers, academia, ventures and startups, healthcare institutions, and technology organizations needs to come together to purposefully build in the healthcare equity agenda in addition to their commercial and market growth plans. Not as a checkbox that is ticked off to satisfy supplementary regulatory reporting, but as a conscious choice.

To quote Stephen Tobolowsky’s The Dangerous Animals Club, “Any endeavor has unintended consequences. Any ill-conceived endeavor has more”

If we are not careful in how we conceive, build and utilize AI and ML models, we may end up exacerbating the situation into a real ugly one. Let us hope that as we plow through this stage of cyber-revolution or 4IR, we are paying attention to the full spectrum of fundamentals. Not just the “ever-important” financial ones, but also those that can make a tremendous impact on the construct of our societies and really change how healthcare is delivered to the communities we strive to serve.

To a world of better healthcare equity – access, treatments, and outcomes!

Dr. Shakthi Kumar, DHA, MBA has held several leadership roles as CEO, EVP, VP & COO, and co-founder for several healthcare and life sciences organizations and startups. He currently serves as Worldwide Head, Healthcare Solutions for Amazon/AWS where he works with an amazing team of global healthcare experts and practitioners all striving to make a difference in the lives of patients, providers, and customers. Dr. Kumar also has served as an adjunct professor for MD/MPH students on Global Medicine and Health Policy track. He holds a Doctorate in Health Administration with inter-clinical disciplinary focus and post-doctoral training, an MBA, and majors in neural networks and computer engineering.

References

  1. Pearson D. (2021). Article available from https://aiin.healthcare/topics/diagnostics/barretts-esophagus-ais-radar?utm_source=newsletter&utm_medium=ai_weekly [Accessed December 2021]
  2. Duncker D et al. (2021). Sensors, 21(7), 2539. Article available from https://www.mdpi.com/1424-8220/21/7/2539 [Accessed December 2021]
  3. Miller K. (2021). Article available from https://hai.stanford.edu/news/ai-predicting-covid-19-prognosis [Accessed December 2021]
  4. UnivDatos Market Insights. (2021). Information available from https://www.marketwatch.com/press-release/artificial-intelligence-ai-in-healthcare-market-to-reach-us-265-billion-by-2025-globally-cagr-41-2021-10-21 [Accessed December 2021] 
  5. Micca P et al. (2021). Deloitte Insights. Article available from https://www2.deloitte.com/us/en/insights/industry/health-care/health-tech-private-equity-venture-capital.html [Accessed December 2021] 
  6. Schwab K. (2017). The Fourth Industrial Revolution. Book available from https://www.amazon.com/Fourth-Industrial-Revolution-Klaus-Schwab/dp/1524758868 [Accessed December 2021]
  7. Seto E et al. (2012). Int J Med Inform, Aug;81(8):556-65. Article available from https://doi.org/10.1016/j.ijmedinf.2012.03.001 [Accessed December 2021]
  8. Forgani R. (2020). Neuroimaging Clinics of North America, 30, Issue 4. Available from https://www.sciencedirect.com/journal/neuroimaging-clinics-of-north-america/vol/30/issue/4 [Accessed December 2021]
  9. Kenner B et al. (2021). Pancreas, Mar 1;50(3):251-279. Article available from https://journals.lww.com/pancreasjournal/Fulltext/2021/03000/Artificial_Intelligence_and_Early_Detection_of.1.aspx [Accessed December 2021]
  10. Andras I et al. (2020). World J Urol Oct, 38(10):2359-2366. Article available from https://link.springer.com/article/10.1007%2Fs00345-019-03037-6 [Accessed December 2021]
  11. Beltrán-Sánchez H et al. (2015). Cold Spring Harb Perspect Med, Nov; 5(11): a025957. Article available from http://perspectivesinmedicine.cshlp.org/content/5/11/a025957 [Accessed December 2021]
  12. World Health Organization (WHO). (2019). Article available from  https://www.who.int/news/item/04-04-2019-uneven-access-to-health-services-drives-life-expectancy-gaps-who [Accessed December 2021]
  13. Crimmins E et al. (2010). Demography, 47, S41–S64. Article available from https://read.dukeupress.edu/demography/article/47/Suppl%201/S41/169942/Biodemography-New-approaches-to-understanding [Accessed December 2021]
  14. Solé-Auró A et al. (2015). Popul Res Policy Rev, Jun; 34(1): 1–18. Article available from https://link.springer.com/article/10.1007%2Fs11113-014-9337-6 [Accessed December 2021]
  15. Richardson LC et al. (2016). MMWR Morb Mortal Wkly Rep, Oct 14;65(40):1093-1098. Article available from https://www.cdc.gov/mmwr/volumes/65/wr/mm6540a1.htm [Accessed December 2021]
  16. Fernandes LL et al. (2018). J Clin Oncol, 36, no. 15_suppl. Article available from https://ascopubs.org/doi/abs/10.1200/JCO.2018.36.15_suppl.e18671 [Accessed December 2021]
  17. Chen C, Wong R. (2018). Article available from https://www.facingourrisk.org/XRAY/racial-diversity-in-cancer-drug-clinical-trials [Accessed December 2021]
  18. American Cancer Society. Cancer Facts and Figures for African Americans 2019-2021. Article available from https://www.cancer.org/content/dam/cancer-org/research/cancer-facts-and-statistics/cancer-facts-and-figures-for-african-americans/cancer-facts-and-figures-for-african-americans-2019-2021.pdf [Accessed December 2021]
  19. Rencsok EM et al. (2020). Cancer Epidemiol Biomarkers Prev, 29(7):1374–1380. Article available from https://cebp.aacrjournals.org/content/29/7/1374 [Accessed December 2021]
  20. Hamel L et al. (2020). Kaiser Family Foundation. Article available from https://www.kff.org/coronavirus-covid-19/report/kff-covid-19-vaccine-monitor-december-2020/ [Accessed December 2021]
  1. National Institutes of Health. (2021). Article available from https://datascience.nih.gov/news/nih-funds-new-consortium-aimed-at-advancing-health-equity-and-researcher-diversity [Accessed December 2021]
  2. World Health Organization. (2017). Article available from https://www.euro.who.int/__data/assets/pdf_file/0009/345798/HEN52.pdf [Accessed December 2021]
  3. Anastasijevic D. (2020). Article available from https://newsnetwork.mayoclinic.org/discussion/mayo-clinic-completes-deidentification-of-expansive-medical-dataset/ [Accessed December 2021]
  4. Anastasijevic D. (2020). Article available from https://newsnetwork.mayoclinic.org/discussion/mayo-clinic-launches-its-first-platform-initiative/ [Accessed December 2021]
  5. Access to Medicine Foundation. (2021). Article available from https://accesstomedicinefoundation.org/access-to-medicine-index [Accessed December 2021]
  6. Tsamados A et al. (2021) AI & Soc. Article available from https://link.springer.com/article/10.1007/s00146-021-01154-8 [Accessed December 2021]
  7. Stanford University Center for Artificial Intelligence in Medicine & Imaging. (2021). Article available from https://aimi.stanford.edu/news/stanford-aims-global-data-access-ai-research [Accessed December 2021]
  8. Organisation for Economic Co-operation and Development (OECD). Website available at https://www.oecd.org/digital/artificial-intelligence/ [Accessed December 2021]
  9. U.S. Food & Drug Administration. (2021). Article available from https://www.fda.gov/medical-devices/software-medical-device-samd/good-machine-learning-practice-medical-device-development-guiding-principles [Accessed December 2021]