Identifying Near-Term Use Cases for AI in the Healthcare Setting

With competition heating up in a rapidly changing, consumer-driven environment, healthcare organizations are increasingly investigating how to integrate artificial intelligence into their clinical workflows and administrative processes.

Improving efficiency, reducing provider burdens, and creating enjoyable consumer experiences are all on the agency for health systems as they strive to maintain profitability and maybe even gain an edge over competitors in the region.

Artificial intelligence has the potential to help. While experts continue to caution that AI is no magic bullet, emerging mathematical strategies such as deep learning and neural networks do have the ability to bring major positive changes to the everyday experiences of providers, patients, and administrative staff.

Healthcare is finally reaching the point where AI strategies can be applied to the industry, said Dr. Katherine Andriole, Director of Research Strategy and Operations at MGH & BWH Center for Clinical Data Science and Associate Professor of Radiology at Harvard Medical School.

“We now have the data storage and compute infrastructure we need, including the graphics processing units (GPUs), that can support really rapid analytics at scale,” she told

“And we’re finally ready, from a cultural aspect, to go truly digital. We have the EHR data; we have the lab data; we have digital images that are absolutely critical for training and validating AI models. That’s all relatively new for an industry that has been…well, let’s say reluctant to embrace change.”

Part of that reluctance is due to a very understandable hesitancy to invest in unproven technologies or tools that do not have a crystal-clear use case.

In 2017, a HIMSS Analytics survey found that close to a quarter of providers could not see solid use cases for AI in the care environment, while an additional 19 percent said the business proposition for machine learning was difficult to understand.

Even more problematic for organizations considering AI tools is the response from early adopters. Half of AI users responding to the poll said that the technology was not fully developed. The same number stated that they struggled to solve concrete business problems with the tools they purchased.

The industry has changed significantly since then – two years ago is practically ancient history in AI terms – but many organizations still find it a challenge to understand how AI will fit into their workflows.

“There is enormous hype around artificial intelligence, and we do need to be cautious when discussing what is possible now and what will be possible in the near future,” Andriole acknowledged.

“But there is also a lot of low-hanging fruit that can be picked right now. The use cases on the clinical side are generally the more exciting, but we can do so much to automate administrative processes and really reduce the inefficiencies that we have right now.”


At the Center for Clinical Data Science, Andriole focuses on radiology – one of the earliest places where AI has taken hold.

“Radiology is definitely leading the way when it comes to AI,” she said. “Pathology is there, too, to some extent. That’s because imaging analytics is one of the ‘easier’ things to do with artificial intelligence. Computer vision is very advanced, and it can close a lot of gaps.”

For example, Partners HealthCare has developed a suite of tools that can support providers as they care for stroke patients, Andriole said.

“Typically, if a provider suspects a stroke, they will order a CT scan to determine where the stroke is and what type of stroke it is,” she explained. “A hemorrhagic stroke and an ischemic stroke require two very different treatments, and time is absolutely critical when deciding what to do.”

“AI can very quickly determine the characteristics it sees on the image and help guide clinicians towards the right care for that individual. When you’ve got 50 or 100 patients in the waiting room of the emergency department, an alert that says, ‘hey, this patient has signs of stroke on their CT, you better get to them first’ is a really valuable asset.”

Artificial intelligence also has applications for predicting future events, such as the development of sepsis or the likelihood of a patient having seizures, Andriole continued.

“We need to make sure that things aren’t falling through the cracks, and using AI can create a safety net for providers so that they get alerted to issues quickly – often long before the human eye can detect a problem,” she said.

“That doesn’t just speed things up, it also reduces costs and helps organizations perform better on measures such as length of stay or readmissions rates. It can improve outcomes, and that’s the goal of any health IT tool.”

In the near future, Andriole also sees AI being used to decision-making around transitions of care or decisions in the ICU, as well as helping providers connect the dots for diagnosing patients.

“Diagnostics is a very manual and very subjective process right now,” she observed. “If you’re a physician with a group of paper charts, or even a group of electronic charts, it’s up to you to put the patterns together and remember every little detail so that you can decide on how to treat each individual.”

“Well, now we this huge treasure trove of digital data, and faster ways to analyze it so that the computer is pulling out the patterns and presenting them to the clinician in an organized, coherent way. I think every provider could benefit from that, because it can support much more informed care.”


The back office can benefit, too, Andriole stressed, by leveraging AI to speed up workflows and streamline mundane tasks. It might not be as exciting as diagnostics or personalized medicine, but there are significant cost savings to be had from using machine learning to close administrative gaps.

“There are so many opportunities in the administrative realm,” she said. “Scheduling, for example, is something that computers are great at doing. If you have a no-show, it isn’t just an inconvenience. It’s a cost issue, because now you have a very, very expensive MRI scanner sitting empty, and you’re effectively throwing money down the drain.”

“If you can use AI to predict which patients might not show up, then you’ve got a much better chance of ensuring that your machines are always contributing to the bottom line – and that your staff isn’t just twiddling their thumbs, either. It’s a winning situation for patients, staff, and the health system as a whole.”

Billing and coding are other near-term use cases for artificial intelligence.

“Incorrect billing can cost millions of dollars, and in some cases, it can even be viewed as fraud,” Andriole said. “Using natural language processingto match documentation with standardized billing codes is a prime use case for AI.”

“Health systems should already be familiar with natural language processing, since it powers the voice recognition and dictation tools used by so many physicians. It’s not a new technology, but it is becoming more and more refined to get better at extracting meaning from unstructured documentation.”

In addition to recouping more accurate revenue, automating the documentation and coding processes could save time and stress for physicians who currently spend too much of their day on administrative processes.

“Burnout is a real problem, and AI has very strong potential to relieve some of the burdens that affect providers every day,” she said. “The important thing to remember is that AI is a supplement to humans, not a replacement. It can cut down on the messy, annoying parts of doing a job, but it won’t take people’s jobs from them.”


Healthcare organizations that aren’t sure how to bring AI into their environments should start by asking front-line staff about their pain points.

“The life of a provider is very busy and hectic, so you need to understand what problems they are having and how those issues impact patient care,” Andriole said. “Unless these technologies are seamlessly integrated into the real-world workflow, they simply won’t be used.”

“Collecting feedback from your physicians, nurses, and administrative staff is definitely the first step, even before you start looking at vendors.”

Once an organization has identified a clear problem that AI can help to solve, they can start exploring the market.

“There is such a broad spectrum of applications out there, and there are hundreds of vendors offering tools with AI baked into them,” she said. “You need to understand a little bit about how each model operates, what kind of data it has been trained on, and how it can be adapted to your patient population or the quality of the data assets you have now.”

Andriole suggests asking vendors for a chance to import the organization’s data and test the tool against the specific types of information it will be expected to analyze.

“Most AI tools that are being demonstrated to prospective customers have been trained on a specific patient population that doesn’t always represent the same patients that you will be treating,” she said.

“So may look like it’s working really well during that sales meeting, but you need to try it on your own data to see if it is truly capable of meeting your specific needs.”

“We don’t do any of this without a clinical champion on the team,” Andriole added. “We need to understand if, when, and how the tool is going to be used, and a clinician is really the only one who can give you an honest, clear answer.”

If the tool receives approval from the end-users, it will be easier to convince executive sponsors that the purchase will be worthwhile.

Organizations should set clear measures of success for a pilot program using AI. Evaluating key factors such as changes in completion time, accuracy of diagnoses, patient outcomes, or revenue accrued will help to identify areas for improvement and define the return on investment.

“Does the tool make your providers more efficient? Does it improve detection of something that was troublesome in the past? Does it help a physician in training understand something better, or bring a non-specialist closer to the level of a specialist? These are important things to monitor if you’re implementing AI in the clinical environment,” Andriole said.

“Medicine is very, very complex, and some of these performance indicators are going to be difficult to measure. But if you make sure that you’re involving clinicians and collecting their feedback, you will have a good idea of whether it’s worth expanding implementation or taking a different path in the future.”

Artificial intelligence will continue to evolve into a game-changing technology for healthcare, Andriole concluded, and organizations should be prepared to dive in sooner rather than later.

“It’s not magic, certainly, but it is something that can drastically change the way we work,” she said. “I find that very exciting. I do believe that we will be able to bring AI into the care setting in a way that produces a huge net benefit for patients and for providers, and it’s going to happen very quickly.”

Originally published in Health IT Analytics by Jennifer Bresnick on March 22, 2019