Why and how we must accelerate AI’s impact on global health

Before GPS, drivers relied on the humble printed map. GPS changed everything. The technology wasn’t perfect — it was ignorant of new roads, temporary closures and traffic conditions — but we adopted it because it beat the status quo. The occasional glitch wasn’t the worst thing in the world and, while the public might not have fully understood satellite-based radio navigation, they didn’t need to.

Self-driving technology, however, faces an uphill climb given its complexity, life-or-death stakes and the need to work everywhere and in all conditions. AI in healthcare resembles self-driving vehicles in that the data and assumptions used to build algorithms are obscure, the consequences of a clumsy algorithm are dire and, to transform healthcare, solutions must be deployable at scale.

Three seismic shifts are driving AI innovation and AI’s impact on global health. Firstly, there is the data deluge. The doubling time for medical knowledge in 1950 was 50 years. In 2020 it was 73 days. In our survey of 1,000 doctors in the US, Europe and Asia, the vast majority said they’re overwhelmed by the volume of patient data. Without technology, humans can’t keep up.

Secondly, there are novel problems. COVID-19 exacerbated existing healthcare issues, worsening an already severe doctor shortage and testing the solvency of many hospitals, among other issues. Technology must enable healthcare providers to do more with less or else patients will suffer. Thirdly, there is a technological renaissance. Look at ChatGPT or what deep learning can do to unlock the mysteries of the universe in search of cures.

We interviewed around 50 senior executives from across biopharma, health tech, government, academia and non-governmental organizations to explore today’s landscape for AI applications, barriers to their adoption and ways to accelerate their widespread use. While 98% of executives report either having an AI strategy in place or planning one, up to 92% of AI projects fail to deliver the desired value. According to our interviews, people overestimate AI in the short term while underestimating it in the long term.

The most promising use cases are in data-driven diagnosis and risk stratification. That’s because AI is cheap, fast and scalable. For example, the most common form of pancreatic cancer has a five-year survival rate of less than 10%; with earlier detection, it’s 50%. Meanwhile, researchers at Cedars Sinai have developed an AI imaging tool that can detect the disease up to three years earlier than a traditional diagnosis — with 86% accuracy. It’s impactful because the cumulative benefit of all those early diagnoses outweighs the occasional false positive.

Interestingly, nearly three in five physicians believe AI technology is most useful for solving chronic diseases, which account for 75% of all US healthcare spending. Apollo Hospitals, one of the largest hospital systems in India, has developed an algorithm for the early detection of cardiac disease that outperforms the benchmark Framingham risk score. Algorithms are now being used for early diagnosis of diabetes, asthma and even depression.

ZS evaluated 400 use cases to gauge the state of AI in healthcare today. While the immediate patient impact is highest in early diagnosis and risk stratification, AI is also being deployed to develop new treatments and improve health system operations. In the realm of adding health system value, BlueDot and Metabiota, for example, use surveillance platforms spanning health, mobility and climate patterns to improve health system decision-making through their stunningly accurate predictions about pandemic spread. And, in drug development, companies such as H1 and Medable are advancing health equity and accelerating speed to cures by using AI to create more diverse and efficient clinical trials.

These innovations represent the pinnacle of human ingenuity, but we mustn't build the Ferrari of healthcare while neglecting the road. We must speed adoption and ensure everyone benefits through:

We need usable, representative data. Without it, bad data will beget bad algorithms. Thankfully, opportunities exist to solve the most pressing data obstacles. Federated learning, for example, can train analytical models to solve biased data sets and allay privacy concerns, while data consortia can help eliminate variations in data security, patient privacy and interoperability across borders.

The best AI feels like a natural add-on to innate human capabilities, yet it so often fails this test. No wonder nearly half of US doctors say they don’t trust it. ZS has identified three principles for closing the AI trust gap: It must be deployed (1) responsibly, with care to eliminating algorithmic bias; (2) transparently, with regard to what it can and cannot do; and (3) competently, such that its benefits outweigh the consequences of an occasional 'miss' such as ordering superfluous lab work.

By creating efficiency at scale, AI can liberate doctors from repetitive tasks in favour of ones humans can do better than AI. But to gain acceptance, AI must integrate seamlessly into existing workflows — either by making tasks easier or less tedious or by doing things humans aren’t already doing. And, AI innovation can’t stay locked inside a few high-income countries. Cross-border, multilateral partnerships must be in place to export them to low- and middle-income countries, where data infrastructure is nascent but fewer entrenched players stand in the way.

One pharma executive put it memorably: In healthcare, we have more pilots than a commercial airline. Indeed, everyone we interviewed agreed we must be much more intentional about which innovations we pursue. Forget clever AI solutions for niche problems. Let’s industrialize the process for creating high-impact AI innovations that can be deployed quickly at scale.

Mayo Clinic launched its 'AI factory' in 2021 to validate and scale AI-backed healthcare innovation through access to its expertise and de-identified data sets. Andreessen Horowitz has partnered with rural US hospitals to test early-stage innovations, including AI. These examples share a recognition that AI holds the key to solving our most intractable human problems — if, like GPS, we make its adoption so easy and its value so axiomatic that it’s unavoidable.

Unlocking this power requires sustained public-private, international cooperation between traditionally unaligned (even misaligned) players. Our future depends on it.