Artificial intelligence in healthcare: How machines can improve patient experience and outcomes

Share with your network

Technology has become an integral part of our lives – including medicine. Now, healthcare industries are harnessing novel innovations, including artificial intelligence (AI), to improve many aspects of care. AI has been applied in patient diagnosis and monitoring, treatment protocol development, radiology, and drug development. While some of this might seem like science fiction, medical professionals use it every day to better the lives of their patients.

How AI improves healthcare

AI allows us not only to analyze data, but also to find their subtle and complex patterns. Machine learning algorithms, specifically, are responsible for these advancements. Engineers have developed software designed to look at a dataset, find relationships between a bunch of variables, and then make mathematical models we can use to predict behaviors in the future or analyze images, patient records, and other data sources.

What does that mean in the real world? With machine learning, computer scientists can make systems that can analyze images and detect diseases better than humans. They can make programs to predict the notoriously volatile process of how proteins fold, or even make drugs that are currently in human trials.

How machine learning systems work

Machine learning systems are complex algorithms. Algorithms are methods of “treating” data, a process in which a computer receives a data input, recognizes the data, and sorts it. Algorithms create a formulaic process for healthcare professionals to evaluate patient symptomology and decide on the best course of treatment. While they cannot replace human decision-making or medical expertise, algorithms can help guide doctors and nurses through a logical thought process. An ideal algorithm is structured to help prevent flawed decisions that harm patients.

AI graphic

Taking available input concerning a patient’s age, weight, risk factors, and symptoms, algorithms can offer the probability of whether a patient has a given condition. Online services like WebMD allow people to enter symptoms and quickly obtain a preliminary diagnosis, an example of a basic algorithm. At a much higher level, algorithms can monitor hospital patients and predict when a patient is at high risk of deteriorating or going into cardiac arrest. Such algorithms are complex versions of “decision trees,” in which the algorithm’s judgment is based on answers to many yes/no questions. This method minimizes the potential for bias by using a formulaic and straightforward process to diagnose medical conditions.

The future is coming: Specific algorithms and their applications

Specific algorithms can take large amounts of data and look for statistical relationships between them. Machine learning systems can make new models by using what are called learning algorithms. There are many types, but the simplest ones involve supervised learning. Supervised learning occurs when “training data” is given to learning algorithms, and the data identifies both the input (labeled) and output (answer key) so the algorithm can be “trained” to distinguish between good and bad results. Do this a bunch of times, and eventually, the algorithm learns what data points lead to the correct output. In other words, the machine learns. Over time, these algorithms get very accurate, and we can use them for specific applications.

Researchers at U.C. San Francisco created an algorithm that scans echocardiogram images and looks for heart issues. The algorithm achieved an accuracy rate over 10% higher in detecting heart issues from these images than human doctors. Algorithms can even predict human behavior in some cases, with a Vanderbilt University Medical Center algorithm being able to predict with 84% accuracy whether a patient admitted for self-harm or suicide attempts would try again within two weeks. These can be helpful predictive tools, allowing doctors to tailor their treatments to best serve patients.

Emory has also developed several algorithms, including one deployed during the 2009 H1N1 pandemic that prevented unnecessary patient visits. Patients experiencing flu symptoms were asked to input their age and answer a few questions about their symptoms. The algorithm could then determine their risk of developing complications and recommend whether they seek medical attention or stay home. Emory Healthcare also deployed a similar tool during the COVID-19 pandemic. Those exhibiting COVID symptoms could visit, which assessed their risk of serious illness based on several questions. The website has helped Emory manage emergency room capacity by allowing doctors to recommend recovery from home for low-risk patients.

Aside from those algorithms, Emory has also developed algorithms that have impacted other areas of healthcare. One algorithm allows for more precise measurements during eye scans, which can greatly improve the accuracy of such scans in picking up signs of Alzheimer’s and dementia. Another created a uniform protocol system for blood transfusions, allowing doctors to track and monitor adverse reactions. Finally, a third improves machine learning itself, strengthening neural networks that give algorithms greater accuracy and learning ability.

How is liability determined?

If a healthcare provider misdiagnoses or mistreats a patient, it’s easy to establish the party at fault. If a machine does the same, the issue becomes much more complex. There are no laws specifically about medical algorithms, but manufacturers can be held liable under general product liability law. The Consumer Protection Act of 1987 allows plaintiffs injured by defective products to sue. Under this act, patients can recover damages if they prove an algorithm is defective. However, significant legal grey areas still exist. Creators of medical algorithms may be shielded from liability if their algorithms go through a Food and Drug Administration (FDA) approval process. Under the concept of preemption, the Supreme Court has ruled that in some, but not all cases, manufacturers of medical products cannot be held liable by state courts if their product received FDA approval.

The US Supreme CourtMore specific regulations by the FDA about algorithms may be forthcoming. Currently, most medical devices require premarket approval from the FDA. But since these regulations were implemented before the development of machine learning techniques, algorithms fall into a grey area and do not generally require such approval. In September 2019, the FDA published a draft guidance regulatory framework for algorithms, which would create an approval process for the software. Since machine-learning algorithms constantly change and adapt, FDA approval is unnecessary for every modification. Instead, the manufacturer would provide a “predetermined change control plan” to the FDA, containing the algorithm’s underlying methodology and anticipated changes.

It’s clear that artificial intelligence will play an increasingly large role and will soon become a critical component in analyzing healthcare data in the future in ways both known and unknown to us now. Just like human judgment, algorithms are not infallible. But when used correctly, they can make our healthcare system safer and more efficient. New applications of machine learning systems are rapidly developing as engineers aim to create new machine learning algorithms that can analyze and find complex data. Combined with the expertise of healthcare professionals, these algorithm-based technologies can potentially improve public health on an even larger scale in the future. Look for algorithms to take on new roles and an even more active role in patient care in the future as artificial intelligence continues to advance.

– Angela Chan