With late 2019 marking the 20th anniversary of the landmark report on medical errors “To Err is Human,” now is time for a renewed focus on novel ways to improve patient safety.
The report launched the modern patient safety movement by shedding some much-needed light on the prevalence of medical errors and preventable deaths in the U.S., spawning many improvements to patient safety over the subsequent two decades.
But before the healthcare industry gets too self-congratulatory, we could use a quick reality check. Patient safety remains a persistent global issue that exacts a huge human cost, as well as a financial one, as a recent report from the World Health Organization (WHO) starkly illustrates.
While it is estimated that there is a one in 3 million risk of dying while travelling by airplane, the risk of patient death while receiving healthcare due to a preventable medical accident is estimated to be one in 300, according to the WHO. Disturbingly, industries with a perceived higher risk, such as aviation and nuclear, have a much better safety record than healthcare, the WHO states.
The report, called “Patient Safety Fact File,” is filled with troubling statistics that reveal the extent to which we must improve patient safety. For example, as many as one in 10 patients is harmed while receiving hospital care in high-income countries, with nearly 50% of these instances considered preventable.
Further, as many as four out of 10 patients are harmed while obtaining care in primary and outpatient settings, with up to 80% of the harm considered to have been preventable. The most detrimental errors are related to diagnosis, prescription, and the use of medicines, according to the WHO.
What Natural Language Processing can do for patient safety
Faced with as vexing and persistent an issue as patient safety, many in the healthcare industry have turned to technology to address these problems. Among the most promising is natural language processing (NLP), which holds great potential for increasing patient safety by enabling healthcare organizations (HCOs) to uncover valuable information hidden among troves of unstructured data.
NLP is the process of examining large collections of documents to discover new information or help answer specific research questions. NLP enables computers to “read” text, simulating the ability of humans to understand a natural language and enabling the analysis of unlimited amounts of text-based data in a consistent, unbiased manner.
NLP holds particular relevance for the healthcare industry, which is awash in unstructured data. While the industry has made great progress in digitizing data over the last decade or so, about 80% of medical data remains unstructured and untapped, much of it trapped within the clinical notes sections of electronic health records (EHRs) systems. Without NLP, accessing this data is a time-consuming process that requires highly paid clinicians to manually comb through patient records to find specific information.
One of the most valuable use cases for unstructured patient data is the detection of future risk. For example, unstructured radiology reports may include mentions of pulmonary nodules, indicating the possibility of early lung disease.
Accurate prediction of future risk has become more important to healthcare organizations as they enter into more value-based contracting arrangements, which require them to measure, track, and report on their quality activities. Under value-based agreements, undetected patient risk may lead to future financial losses if the patient must undergo costly treatments and procedures that could have been prevented with better predictive analytics. Value-based care has also highlighted the importance of addressing social determinants of health (SDoH), such as social and economic factors, lifestyle choices, and living conditions, which frequently exert substantial influence on health outcomes.
How an academic medical center is using NLP to uncover social determinants of health
While nearly all HCOs would like to know more about the SDoH their patients are facing, this information is often inaccessible for clinical decision-making because it is trapped in clinical notes. In many cases, SDoH issues remain unknown to clinicians until those social factors have already begun to exact a toll on patients’ health.
To overcome this challenge, researchers from the Medical University of South Carolina (MUSC) conducted a study to investigate the effectiveness of NLP in identifying cases of social isolation from clinical narratives for patients with prostate cancer. Prostate cancer patients sometimes experience social isolation because of treatment-related side effects such as incontinence. Often this information is not recorded as coded data, but rather collected from patient self-reports or documented in clinical narratives.
Researchers analyzed records of prostate cancer patients, conducting searches on 24 terms relevant to social isolation, such as “lack of social support,” “lonely,” “social isolation,” “no friends,” and “loneliness.” Results showed that the approach demonstrated an impressive 90% precision and 97% recall, leading researchers to conclude that, “NLP algorithms demonstrate a highly accurate approach to identify social isolation.”
Healthcare’s new data problem
Two decades ago, around the time of the release of “To Err is Human,” one of healthcare’s biggest problems was a lack of digital data. Now, as a result of the widespread transition to EHRs, the bigger problem is that much of this recently digitized data remains unstructured, limiting its accessibility and usefulness in clinical decision-making and potentially compromising patient safety.
NLP offers a way to overcome the challenge of unstructured data, helping clinicians improve safety and reduce patient risk by uncovering what previously was essentially unknown information that was hiding in plain sight.