Addressing the root causes behind AI’s slow adoption rates is the first step towards achieving sustainable development goals for Artificial Intelligence (AI) in healthcare.

by Michelle Tan Min Shuen

Somewhere in the United States, a middle-aged man is looking into the camera of his iPhone in a hospital waiting room. Using a light signal processing technology known as remote photoplethysmography (rPPG), his smartphone camera records light reflected by blood vessels beneath his skin, revealing critical vital signs such as heart rate and oxygen saturation levels with medical-grade accuracy. Through a mobile application, doctors are alerted with a set of his vital information in less than 45 seconds. AI-assisted mobile applications such as Docdot allow doctors to remotely find and monitor individuals at high risk of infection while limiting the risk of exposure to themselves and others, proving to be a crucial tool amidst the global pandemic.

While the speed and devastation of COVID-19 has exposed critical flaws in universal healthcare systems, it has indisputably driven the digital transformation of our healthcare landscape today. As more primitive healthcare systems around the world began to crack under pressure of the ravaging global pandemic, more have turned to digitisation to optimise healthcare facilities’ response mechanisms in spite of cost and manpower constraints. The result is a vastly more efficient healthcare landscape which features larger diagnostic capabilities, more innovative solutions and swifter operating systems.

From detecting brain bleeds in CT scans to recognizing patterns in coding to facilitate swift payment, AI’s applications have revolutionised key areas in healthcare, improving the detection, diagnosis, and treatment of illnesses worldwide. As more healthcare organizations continue to trailblaze the incorporation of smart technologies into patient care and business processes, the pace of AI usage in the field of healthcare is picking up, but still not quite to the extent we have envisioned.

The healthcare sector still trails other industries in incorporating smart technologies into routine medical practices. According to the AI in Healthcare Leadership Survey 2020, 42 percent of respondent healthcare facilities have yet to incorporate AI-based applications in clinical practice. Of the ones that do, AI use, namely speech recognition and computer vision, is concentrated in operations and customer service. In fact, in most hospitals, simple tasks such as appointment scheduling are still done manually, which siphons away valuable time from the already packed schedules of healthcare workers. If anything, this reflects the tremendous untapped potential of AI in the future of healthcare. AI methods are likely not to be adopted in routine medical practice beyond a limited number of niche applications unless these key challenges are addressed.

 

Lack of Digital and Human Infrastructure

AI integration in healthcare is most efficient when AI technologies are embedded into the workflows to support clinical decision making at the point of care. Automating workflow can not only optimise time and manpower to allow healthcare workers to focus more on providing patient-centric care, but also reduce manpower and operational costs.

However, it takes technology to grow technology. Counterintuitive as it may seem, this was a key point raised by McKinsey analysts Jacques Bughin and Nicolas van Zeebroeck, who have observed that potential adopters of AI technology “can’t flourish without a solid base of core and advanced digital technologies”. This is bad news for many existing medical institutions who still run on old operating systems. Many of these institutions trail behind in cloud and advanced analytics, using current programmes that are incompatible with AI requirements, preventing them from adding AI features to their existing systems.

Building an AI-ready workforce to build and subsequently complement digital advancements is essential. “Building an AI-ready workforce requires a wholesale change in the approach to training and how to acquire talent.” said Melissa Edwards, managing director and digital enablement at KPMG. This means that these companies need to start investing in hiring staff who possess the diverse expertise required for AI development and training staff in technology integration among many other technology-related skills. “Having people who understand how AI can solve big, complex problems is critical,” she added. Yet there is a crippling shortage of such workers. Knowledge in technological fields is rarely taught alongside traditional clinical sciences. “And so, through no fault of their own, today’s healthcare workforce is simply not yet equipped for the adoption of AI,” says Jorge Fernández García, director of innovation at EIT Health.

Digital and human infrastructure is the very cornerstone of the successful integration of AI technologies into the healthcare industry. Yet its construction entails costly investments which presents as a significant obstacle to the integration of AI. Compounding the issue, “implementing AI technology is an iterative process” which creates “a variety of organizational impacts that aren’t immediately quantifiable”, acknowledges Jane Kaye, a Healthcare Finance Consultant at HealthCare Finance Advisors. She notes this may further deter some healthcare organisations from investing in building digital and human infrastructure needed for the eventual incorporation of AI technology.

 

Inaccessibility of Medical Data

The healthcare landscape lacks a robust data infrastructure due to the scarce availability of relevant data and the fragmentation of patient data across multiple data silos.

As with any other form of AI, machine learning algorithms used in the sphere of healthcare must be trained using extensive amounts of input data. However, the ownership and sharing of patient medical records are classified as highly sensitive information and thus strictly regulated, which stymies the collection of relevant input data required by deep learning applications and other AI tools. This makes it trickier to get the sheer volume of data necessary to train, design and develop safe, effective deep learning algorithms which can be used for various clinical applications.

Moreover, healthcare data are often siloed in a multitude of medical imaging archival systems, pathology systems, electronic health records (EHRs), and insurance databases, making it difficult for doctors to locate and consolidate a patient’s data. Though there have been large increases in hospitals’ adoption of EHRs, which are a digital compilation of patient information gathered into one database, health data interoperability still remains a challenge for clinicians worldwide as patient data resides across multiple EHRs. The fragmentation of healthcare data can make it difficult even for the most successful of AI algorithms to establish a patient’s full healthcare profile and subsequently make informed recommendations.

Enabling interoperability between datasets and technology is vital to the efficient operation of an effective digital service. The technology infrastructure should allow separate data silos to interact securely, maintaining data fidelity while ensuring patient confidentiality. This would provide the large, integrated databases required for AI algorithms to accurately approximate the correlation between diseases and patient outcomes. As reiterated by Janet King, a senior director of marketing insights at HIMSS Media, “Extending advancements in interoperability to the broader healthcare ecosystem will be critical to progressing digital health initiatives and enabling new ways to serve patients.”

 

The Black Box Problem

One critical obstacle which prevents AI adoption rates from picking up lies in the immense difficulty of validating the outputs from AI systems. Machine learning algorithms based on complex neural networks can examine a variety of patient records to make predictions and recommendations about patient care. While these algorithms predict outcomes with extreme precision, they offer little to no comprehensible explanation to the logic behind its decisions. Much like a black box, the inner workings of these computational models have a complexity that rivals the human brain themselves, and are rendered opaque to human understanding.

Citing an example, a software was used on a neural network that was trained to recognise photographs of horses. The neural network picked up on the copyright signal at the bottom right of the input horse pictures instead of the unique biometrics of a horse, and used the former as a criterion to identify horses. As Sheldon Fernandez, CEO of DarwinAI rightfully points out, these algorithms might just produce “the right answers for the wrong reasons”, to the oblivion of its creators.

This is highly unsettling in medical applications, where approaches not only need to reap accurate results, but also be reliable, explainable and transparent. The lack of understanding on how neural networks arrive at their conclusions may result in incorrect predictions based on unsound relations that go unnoticed by doctors and patients alike. Without knowledge of the reasoning behind the algorithm’s predictions, doctors also lack a basis to oppose AI-generated predictions should they have a difference of opinions about its accuracy, which can lead to disastrous consequences. Understandably, doctors are averse to using AI systems without understanding their logic, even if these systems could statistically deliver better diagnoses than doctors themselves.

Not only are explanations needed to validate AI-generated results, they are also needed in order to continually improve the algorithms and software behind it. The lack of explainability in such algorithms hinder the development and refinement of these AI systems to more accurately process information and predict recommendations.

Increasing the explainability of AI algorithms are necessary for healthcare practitioners to understand, appropriately trust, and subsequently incorporate such technologies into their routine medical practice.

 

Moving Forward

“Our saviour here is going to be our technology.” This was former US Food and Drug Administration (FDA) Commissioner Scott Gottlieb’s comment regarding the United States’ response to the COVID-19 pandemic.

During a critical period in which saving minutes can mean saving lives, the FDA saw the need to improve the tedious and lengthy legislative process required for the testing and approval of AI-assisted medical devices, which were constantly being adapted to improve their performances. In October 2020, they proposed a new regulatory framework on AI and machine learning (ML) technologies in medical devices, and have since stepped forward to acquire more public feedback to further refine the framework to better suit patients’ needs. In doing so, they have subtly underscored the crucial role of AI-enabled digital health technologies in the future of healthcare, empowering more health providers to go beyond traditional approaches, embrace change and adopt such technologies into their routine medical practices.

AI solutions fundamentally make healthcare systems more sustainable as they are vastly more efficient and can help to compensate for a looming shortage of healthcare workers. That being said, it is worth noting that simply adding AI applications to healthcare systems without doing anything to mitigate the fundamental flaws in underdeveloped healthcare systems is likely to result in unsustainable change. Addressing the aforementioned challenges is a critical stepping stone to achieving the sustainable development goals for AI in the healthcare landscape, as we have set out to achieve. The machines can heal us, but only if we work towards building a conducive environment to enable them to. [APBN]

---

References

1. Abadicio, M. (2020, September 08). AI in the Hospital Setting – Challenges and Trends. Retrieved from https://emerj.com/ai-sector-overviews/ai-in-the-hospital-setting/
2. AI-assisted COVID-19 diagnosis app launched. (2020, May 05). Retrieved from https://www.med-technews.com/news/ai-assisted-covid-19-diagnosis-app-launched/
3. ARTIFICIAL INTELLIGENCE THE NEXT DIGITAL FRONTIER? (2017, June). Retrieved from https://www.mckinsey.com/~/media/McKinsey/Industries/Advanced%20Electronics/Our%20Insights/How%20artificial%20intelligence% 20can%20deliver%20real%20value%20to%20companies/MGI-Artificial-Intelligence-Discussion-paper.ashx
4. Bhateja, V., Satapathy, S. C., & Satori, H. (2020). Embedded Systems and Artificial Intelligence Proceedings of ESAI 2019, Fez, Morocco. Singapore: Springer Singapore.
5. Bresnick, J. (2018, September 17). Arguing the Pros and Cons of Artificial Intelligence in Healthcare. Retrieved from https://healthitanalytics.com/news/arguing-the-pros-and-cons-of-artificial-intelligence-in-healthcare
6. Chen, M., & Decary, M. (2019). Artificial intelligence in healthcare: An essential guide for health leaders – Mei Chen, Michel Decary, 2020. Retrieved from https://journals.sagepub.com/doi/full/10.1177/0840470419873123
7. Evans, J. (2020, April 09). The integration of AI into healthcare. Retrieved from https://www.innovationnewsnetwork.com/the-integration-of-ai-into-healthcare/4785/
8. FDA proposes new regulatory framework on artificial intelligence, machine learning technologies. (2020, October 16). Retrieved from https://www.news-medical.net/news/20201016/FDA-proposed-regulatory-framework-for-modifications-to-ALML-based-software-as-a-medical-device.aspx
9. McKendrick, J. (2018, July 31). Most companies lack the infrastructure for artificial intelligence. Retrieved November 15, 2020, from https://www.zdnet.com/article/artificial-intelligence-requires-an-intelligent-it-strategy/
10. Reddy, M. (2020, October 21). Digital Transformation in Healthcare in 2020: 7 Key Trends. Retrieved from https://www.digitalauthority.me/resources/state-of-digital-transformation-healthcare/
11. Sheridan, P. (2018, November 27). From Fragmented to Coordinated: The Big Data Challenge. Retrieved from https://www.healthcareittoday.com/2018/11/27/from-fragmented-to-coordinated-the-big-data-challenge/
12. Tkatchuk, R. (2020, April 01). Key Challenges That Healthcare AI Needs to Overcome in 2020. Retrieved from https://dataconomy.com/2020/04/key-challenges-that-healthcare-ai-needs-to-overcome-in-2020/
13. West, H. (2020, March 30). Healthcare and Data Science: How EHR and AI Can Go Hand-in-Hand. Retrieved from https://www.technologynetworks.com/informatics/articles/healthcare-and-data-science-how-ehr-and-ai-can-go-hand-in-hand-332840
14. Wittbold, K. (2020, April 03). How Hospitals Are Using AI to Battle Covid-19. Retrieved from https://hbr.org/2020/04/how-hospitals-are-using-ai-to-battle-covid-19