Utilizing natural language processing to assess healthcare quality
About 10 years ago in the US, there was a huge effort to measure healthcare quality; to understand how well a system or an institution is providing care and services. Many metrics and interventions were developed as a result to assess efficiency and target areas for improvement. In spite so, a lot of the valuable healthcare information are still hidden in what known as the “unstructured text”. Physicians take notes about a patient’s progress and these hand or type-written data are not machine interpretable.
To automate a manual assessment process
“If a physician says you have a family history of breast cancer, how do we use that piece of information to help our analytics to do a better job? Likewise, if you are about to receive a surgery, how can we ascertain your rate of infection?” Dr. Brian Thomas Bucher, Pediatric Surgeon, Division of Pediatric General Surgery and Assistant Professor of Surgery, University of Utah School of Medicine said in a recent interview with AIMed.
Some institutions hire human to manually go through and process every type-written document in the electronic health records (EHR) systems. The advantage of this method is these trained experts guaranteed the generation of a high quality and reliable clinical dataset. The downside being the method is not scalable because a human can only handle so much data within a given period of time. Thus, many hospitals have turned towards random sampling.
“In my institution, we have two people, who will randomly sample and review about one-fifth of all the surgeries that had taken place. That doesn’t seem congruent with what we can do in year 2020”. That was what motivated Dr. Bucher to utilize artificial intelligence (AI), particularly, natural language processing (NLP) to help improve the process and in turn, measure healthcare quality.
Interoperability as the biggest barrier to AI deployment
“Regardless of the surgery, the most common complications a patient is likely to have after which include infection, blood clot, urinal tract infection or pneumonia. So, the projects I was involved in deployed NLP and a variety of other AI algorithms to help identify these events from EHRs”. Although Dr. Bucher’s research interest has a significant meaning in the realm of surgery, deploying AI and NLP in actual operations is still faced with huge barriers.
“I think the barriers that I experienced aren’t different from what others are also experiencing. The biggest one is interoperability, followed by how data is being stored and labeled”. Because EHRs do not communicate with one another and most of the time, they are offered by different vendors, so an AI system developed for a hospital does not actually mean it can automatically be used in a separate hospital. Dr. Bucher believes EHRs’ interoperability is probably the biggest barrier to any type of AI deployment.
While there are solutions, such as FHIR (Fast Health Interoperability Resources) which gives others a glimpse of how their EHR data may look like without venturing into the backends but it is a slow process, because developing FHIR resources is time-consuming and labor intensive. “I think healthcare technology needs to be user-friendly and integrated into the workflow. For example, our smartphone has done a remarkable job. Most smartphones are intuitive and with design aspects that are currently found in EHRs. That can be frustrating,” Dr. Bucher cites.
AI is already in healthcare; it’s just not as advanced as we think
Dr. Bucher said while AI is a buzz word, clinical decision support has already existed in the first generation of EHRs. “If you have a penicillin allergy and a physician orders penicillin for you, the system should alert the physician about it. This feature has been in the EHRs for decades and it’s a form of AI, right? This is an external intelligence that is informing the clinicians to form a better clinical judgement. I think AI already has a place in medicine, it’s just not as advanced as we think”.
“So, I think an important question to address when we deploy a technology is how to integrate it into the workflow and get people to trust the results. A lot of our effort has been developing what we called rules-based NLP, which as compared to statistical-based AI, is more explainable and transparent. For example, if I found a fact that you are likely to have an infection after a surgery, I am able to pull up a document, the phrase that triggers the algorithm which says this is the case”.
Despite the challenges, Dr. Bucher thought AI is an interesting area to venture into, especially for a surgeon like himself. “I can probably count in my hand, the number of surgeons who are actively involved in AI research at the moment. Surgeons, unfortunately, can be rather reluctant to develop and explore new technologies. Their mentality tends to be ‘if we have done this 1000 times, why would we change?’ So, this has created a lot of room for improvement but I think this will change”.
*The article was originally published on AIMed Blog on 19 August 2020.
