Artificial Intelligence (AI) tools are set to change the way medicine is practiced. In his book Deep Medicine, Eric Topol argues that AI can change medicine for the better, if implemented in a way that focuses on improving the doctor-patient connection. Here we look at 7 ways in which AI can improve the patient experience.
A number of algorithms show promise when it comes to medical diagnoses.1 Generally, these are Machine Learning (ML) algorithms, such as neural networks or clustering algorithms, where a computer is trained on a very large data set. The data might be X-rays, pre-labeled with ‘disease present’ or ‘disease not present’ as judged by an expert. With each pass, the computer adjusts the parameters of its algorithm. Once the training is complete, the algorithm is validated on a portion of the dataset that was held back during training.
As Topol points out, such algorithms work best within a narrowly defined domain. For example, looking at one type of X-Ray for one type of disease. I like to think of Machine Learning as an alternative to statistical methods. Where a statistical approach wants to aggregate and average, machine learning wants to discover details that make something stand out.
Humans will always outperform Machine Learning at tasks that require more general knowledge. You wouldn’t want to replace your doctor with a machine. However, you also wouldn’t want your doctor to miss something that a machine might have flagged. AI diagnosis is about giving medical professionals additional tools.
Not only will providers have these tools, but when the machine is incorrect, a doctor will be able to provide feedback that can be used to improve its accuracy. The algorithm can benefit from the experience of the entire medical community.
Most people see the future value of an Electronic Medical Record (EMR), but currently doctors spend over twice as much time entering data than they do interacting with their patients. Natural Language Processing (NLP) is the field of AI expected to help solve this.2
NLP takes spoken or written language and attempts to extrapolate meaning in a way that can provide useful information to a computer system. Older attempts at NLP used rule-based methods, but modern NLP algorithms use Machine Learning.
In Deep Medicine, Topol explains that while good efforts are being made, there are still some drawbacks, such as capturing nonverbal clues, or the need for providers to review the record before approving it. After all, the machine doesn’t understand the medical record, it is only assigning a probability that a particular line of text relates to, say, a diagnosis code (which it also doesn’t know the meaning of).
Still, for all the hype, small steps are being made. CloudMedX®3 uses natural language processing and machine learning in real time to prompt doctors to fill in details that can not only help make the medical record complete, but also reduce the friction of data entry.
Most nursing units schedule nurses manually. Optimizing the schedule can be difficult task. Over-staffing is expensive, and under-staffing leads to safety issues and burnout. Often hospitals will compensate by supplementing their staff with agency nurses, which is even more expensive.
An algorithmic system could optimize the schedule, not only scheduling the right nurses at the right times, but also accounting for nurse's preferences in a fair and transparent way. The result is increased safety, at less cost, and a manager with more free time.
An algorithmic system could optimize the schedule, not only scheduling the right nurses at the right times, but also accounting for nurse’s preferences in a fair and transparent way.
The way this works is fairly simple by today’s standards, using algorithms that take advantage of the computers willingness to exhaustively try every option and then choose the answer that fits most closely to the given parameters. This sort of algorithm wouldn’t be considered AI today, but in the 1990’s it was indiscernible from magic. If you want to sprinkle some future dust on it, combine the scheduling with historic trends, moon phases and a real-time feed from ER and the bed management system, and we will be able to predict scheduling needs further in advance.4
Most healthcare organizations implement a patient satisfaction survey to measure patient satisfaction. A numerical score is useful, but what if we had a more nuanced measurement to describe a patient’s feelings. Instead of “strongly agree” to “strongly disagree”, you could measure gratitude, anger, contentment and frustration.
Sentiment analysis uses Natural Language Processing and Machine Learning to assign an emotion to unstructured text, such as comments from surveys. Currently this is cutting edge in marketing.5 Both Amazon AWS and IBM Watson provide sentiment analysis as tools that can be accessed by other systems over the internet.
One obvious application is to review patient survey comments, rapidly highlighting issues that need to be addressed. From a broader view, aggregating and measuring how this data changes over time offers a way to see the result of process improvements. Additionally, research shows that sentiment analysis can be used to detect suicidal thoughts6, which supports the goal of patient care.
You might not want a robot in place of a nurse, but what if a robot was able to do some of the repetitive and time-consuming work, freeing up the nurse to spend more time with their patients?
Diligent Robotics has designed a robot called Moxi that can be used to deliver supplies and samples, and it has turned out to be quite popular.7 The robot is designed not to be intimidating, and goes about its daily tasks quietly and efficiently.
If you’ve used a Roomba, heard about self-driving cars, or seen what a group of middle-schoolers can do in a robotics competition, you know that the robotic assistants are definitely coming.
Medication errors are a significant risk to patient safety. A robot in the pharmacy might not be quite as anthropomorphic as Moxi, but it is certainly willing to work continually and with extreme accuracy.
UC San Francisco has implemented a robotic pharmacy that dispenses medicines and uses robots to deliver them to nurses’ stations with claims of 100% accuracy.8 The machine receives prescriptions, fills them and delivers them by means of autonomous robot “tugs”.
Not only does this improve patient safety, it changes the role of the pharmacist. Less time filling orders means more time counseling patients and consulting with doctors about patient care.
Financial interactions are a large part of the patient experience. Improving this is an excellent opportunity to improve the patient experience.9 Patients want transparency and simplicity, and in some cases, need flexibility. Determining the cost of a procedure might be straightforward for an algorithm with some historical data to analyze, but once you add in government regulations and insurance contracts, even R2-D2 would give up.
So, where do we start? In the financial area, insurance companies are beginning to use AI to look for evidence of fraud.10 It might be a stretch to say that this is about patient satisfaction, but fraud has been linked to patient harm. Fraud can impact patients financially, medically and can leave lingering problems from identity misuse.11
The paradox here is that a machine’s ability to tame a complex system might encourage systems to make more complex rules, which is another reason we should always advocate for better patient experience in the healthcare system.
Currently AI technology is in the middle of a “hype cycle” — a period of over-inflated claims which will eventually give way to a “trough of disillusionment.”12 Hopefully, this article leaves you with a realistic picture of the tools available and the actual, incremental changes that can be made. Whether or not those changes improve the doctor-patient relationship, and in turn, the patient experience, is up to us.
References:
1. https://medicalfuturist.com/top-ai-algorithms-healthcare
2. https://hitinfrastructure.com/news/artificial-intelligence-uses-ehrs-as-smart-analytics-tools
3. https://healthitanalytics.com/features/what-is-the-role-of-natural-language-processing-in-healthcare
4. https://www.beckershospitalreview.com/healthcare-information-technology/3-things-you-need-to-know-about-nurse-scheduling-software.html
5. https://towardsdatascience.com/five-practical-use-cases-of-customer-sentiment-analysis-for-nps-a3167ac2caaa
6. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6111391/
7. https://www.fastcompany.com/90372204/a-hospital-introduced-a-robot-to-help-nurses-they-didnt-expect-it-to-be-so-popular
8. https://www.kqed.org/futureofyou/153628/when-a-robot-counts-out-your-pills-what-will-your-pharmacist-do
9. https://revcycleintelligence.com/news/aligning-the-healthcare-revenue-cycle-with-patient-experience
10. https://scholarlycommons.law.northwestern.edu/cgi/viewcontent.cgi?article=7231&context=jclc
11. https://www.forbes.com/sites/insights-intelai/2019/02/11/how-ai-can-battle-a-beastmedical-insurance-fraud/
12. https://en.wikipedia.org/wiki/Hype_cycle