How Google is creating AI solutions for healthcare

Artificial intelligence (AI) isn’t as new as some people might think. Many big tech companies had been working on advanced language models for years before ChatGPT became commonplace. One such company has not only developed impressive AI tools but has also set up a department dedicated to its impact in healthcare.

In this article, we’ll dive into that team, looking at its purpose, developments, and what we can expect from it in the future.

Google recognizes the need to conduct the best UX research to produce solutions for the healthcare industry. Meanwhile, Google AI healthcare tools also provide an immense business opportunity. The company even has its own Google Health team focused on creating new machine-learning tools and discovering opportunities to increase the availability and accuracy of healthcare globally.

What has the team produced?

The projects and technology coming from the team are significant. Deep learning for electronic health records (EHRs) is one example. It uses a broad set of predictions relevant to hospitalized patients and leverages de-identified EHRs to render predictions.

In addition to health records, the team is also helping to turn medical speech into actionable information. They have developed models for understanding medical speech and conversations to accurately recognize symptoms of speech disorders. The models used to do this are shared publicly and are available through Google Speech APIs. This affords the broader community and other tech companies the ability to build off of the same technology.

Applications in breast cancer research have been another focus. The team is working with researchers in mammography and pathology to develop meta language (ML) to help physicians screen for and diagnose breast cancer.

More broadly, the team’s research across all their developments is regularly published in medical publications, giving the community access to in-depth developments and encouraging collaboration that advances healthcare.

Who is the team made up of?

The team is scattered across multiple continents and nations, including North America, Europe, Asia, and Australia. It consists of computer scientists, research engineers, and research scientists. These experts team up to focus on algorithms and theories, machine intelligence, machine perception, natural language processing (NLP), and speech processing.

Each team has its own projects. For example, the Amsterdam team focuses more on computer vision and audio, producing work in a neural weather model. It’s also making developments in reinforcement learning (RL) with football players, which promise advancements in robotics and self-driving cars.

Another development team that is deeply invested in AI is Google DeepMind. The project was originally called DeepMind and started in 2010. In 2023, DeepMind joined forces with Google Research’s Brain team, and Google DeepMind came about as a combined effort.

Among their impressive catalog of current research efforts, one particular project is focused entirely on the future of healthcare, developing reliable AI tools for healthcare. Its goal is to develop a way to test for predictive AI’s relative accuracy to decide when the system should defer to a human clinician. If this is successful and the technology is self-aware enough to know when it needs human intervention, it could provide significant peace of mind—especially now, at the dawn of AI use in healthcare.

The results are looking promising, with the latest technology reducing the number of false positives by 25%! The dataset was a large, de-identified UK mammography, with results compared against commonly used clinical workflows.

For those who are still fearful of Google AI missing a positive diagnosis, fear not. The results also show the technology didn’t miss a single true positive in this scenario. As icing on the cake, the team published its technology on GitHub, opening up its immense promises to anyone hoping to leverage the same capabilities.

DeepMind and Google’s efforts are causing more and more ripples across the AI sector. They have even stolen the show from Google’s recent hardware releases.

In May 2023, DeepMind introduced its most advanced language model. Its capabilities are directed specifically at the medical space. Referred to as Med-PaLM 2, the model is designed to answer questions and summarize insights with certified medical competency. Likened to its Watson ancestor, Med-PaLM 2 is the first large language model to achieve “expert level” on the US Medical Licensing Exam.

But these impressive accomplishments don’t appear to be stopping anytime soon. In addition to all the projects already discussed, Med-PaLM 2 is set to add multimodal capabilities. This will allow the language model to analyze additional information outside text like x-rays, mammograms, and imaging. This exciting development promises to improve patient outcomes.

DeepMind has since opened Med-PaLM 2 up to a small cohort of Cloud customers to start gathering feedback and discover safe, beneficial use cases.

Google AI in healthcare is a unique endeavor for Google. Though the effort isn’t purely philanthropic, its developments are already proving themselves across a range of applications and providing otherwise unattainable benefits.

It might be best to preface these breakthroughs with some caution. Besides obvious HIPAA compliance concerns, ethical dilemmas are a worry—especially when it comes to implementing large-scale automation of healthcare work.

A study published by the Future Healthcare Journal shows that as valuable as AI is in healthcare, a number of obstacles will inhibit the speed of adoption.

With that said, the future of AI in healthcare is looking very promising.

Natural language processing

NLP is finally achieving the hopes and dreams of data scientists from as far back as the 1950s.

In addition to speech recognition, the technology can also be used to analyze unstructured clinical notes, transcribe patient interactions, prepare reports, and conduct conversational AI.

Rule-based expert systems

If you’re familiar with tools like IFTTT or Zapier, you’re probably also familiar with the if-then statements that rule-based expert systems are built from. AI takes these a step further in healthcare with what’s coined “clinical decision support,” a common tool that supports physicians to make the most accurate and up-to-date diagnoses.

The current systems run into errors if over 1,000 rules are set. Thankfully, AI developments will help replace these tools with high-capacity options that handle even more complexity.

Physical robots

Over 200,000 industrial robots are installed globally each year. In the US alone, 2,000 surgical robots will soon be equipped with AI, providing surgeons with previously inconceivable capabilities.

With all critical decisions still made by human doctors, AI-equipped robots can help with precise and minimally invasive incisions and stitching wounds, among other things.

Robotic process automation

Contrary to what the name suggests, robotic process automation doesn’t necessarily involve physical robots. More focused on the administration side of healthcare, robotic process automation can help update patient records, assist in prior authorization, and handle patient records or billing tasks.

This AI technology, combined with image recognition, can also be used to extract data from, for example, faxed images and input it into transactional systems.

Implications for the healthcare workforce

AI technology is a threat to the healthcare workforce. However, it faces many obstacles that will slow the timeline. As such, the field stands to benefit before experiencing job losses.

For example, interpreting images isn’t a radiologist’s only responsibility. Although AI can take on this particular task, radiologists can provide other value through alternative roles. What’s more, AI algorithms analyzing medical images can help humans identify patterns and anomalies much faster and help avoid errors. Fortunately, AI in healthcare appears more likely to provide support before it threatens careers.

AI in healthcare currently poses certain risks to be assessed as it starts its journey in the field.

Data breaches

Large data sources tend to attract sinister actors. As Google develops AI tools for healthcare, keeping sensitive patient data protected from cybercriminals will become paramount. Such data will be vulnerable across the entire AI data journey.

Privacy attacks

Membership inference, property inference attacks, and reconstruction are different types of attacks AI algorithms could fall victim to. These attacks target data about individuals. What’s even more concerning is that this kind of identity theft is expected to increase, especially around individual data included in AI training sets.

Data input poisoning and model extraction

Though data input poisoning and model extraction sound complex, they are simply two different ways to manipulate datasets within an AI algorithm.

Data input poisoning is when an adversary introduces bad or faulty data into the training set. On the other hand, model extraction is when the bad actor takes out or extracts enough information about the AI algorithm itself. An attacker typically carries out these attempts to create a similar model that either competes with or substitutes for its origin.

AI for malicious purposes

One of the biggest threats to a tool as powerful as AI is adversaries weaponizing the technology itself.

For example, AI has already been known to spread falsehoods through propaganda and faulty content to wide, impressionable audiences. ChatGPT, for example, can generate phishing emails just as convincingly as any human.

UX designers are critical to this space. In fact, when it comes to AI in healthcare, most advancements simply wouldn’t work in practice without UX designers and researchers.

A study released by Google themselves showed that new Google AI for healthcare necessitates careful UX studies to ensure practical implementation. This is especially true outside of the lab; patients and physicians in real-life situations are the least able to incorporate this technology into everyday practices.

This issue came about with an AI tool for ocular use. The tool was developed to save time reviewing patients’ retinal images. The AI acts as a digital ophthalmologist, reviewing retinal images to scan for diabetic retinopathy. It achieved over 90% accuracy, but proved impractical. Disagreements between the system and the clinician led to frustration. Thanks to UX intervention, the system was taught to refer any questionable images back to ophthalmologists.

Amending AI systems to better handle human–machine conflict will be the job of UX teams. This vital input will make all the difference between AI in healthcare flourishing or remaining impractical.

The future looks bright for AI in healthcare and for UX designers interested in the AI advancements happening at Google and beyond. Google has a team called AIUX specifically so that UX designers and researchers can be involved in their AI developments.

How Google is creating AI solutions within healthcare is as impressive as it is inspiring. However, they will need to follow UX best practices to help keep medical workers, physicians, and patients safe and empowered by the technology rather than frustrated and threatened.

Google will also need to take special care to ensure hackers and other adversaries don’t use the developments for malicious intentions.

Thankfully, Google’s AI endeavors all appear to be heading in a positive direction. Faster test results, streamlined health administration, and even countless advancements in medical science are on the horizon.