Could Natural Language Processing Help Close Gaps in Infection Prevention and Patient Safety Events?

Natural language processing (NLP) is an increasingly ubiquitous form of artificial intelligence (AI). Best described as where computer science meets linguistics, it uses computational linguistics and machine learning to analyze human language. 

NLP powers virtual assistants like Alexa and Siri, predictive text for emails, spelling and grammar checkers, and sentiment analysis in reviews. In addition to consumer use cases, NLP is used in medicine to identify risk factors, estimate risk, or predict events of disease development or readmissions across cardiovascular, endocrine, metabolic, and neurological diseases. And it’s put to work in patient engagement—with NLP-powered chatbots supporting pharmacy interactions, COVID-19 management, and primary care triage.

Does NLP have a potential role in supporting infection prevention? Let’s explore the possibilities.

Strengthening Linkages Between Risk Management and Infection Prevention

As a quick refresher, the Patient Safety and Quality Improvement Act of 2005 created a framework for establishing patient safety organizations (PSOs) to collect and analyze confidential information from healthcare providers. PSOs—including ECRI and the Institute for Safe Medication Practices PSO—gather safety event data that healthcare providers voluntarily report. PSOs then use this information to improve patient safety and healthcare quality by identifying risks and hazards associated with patient care.

Even so, there can be gaps in how data is collected, shared, and used within a healthcare organization. For example, risk management departments may be responsible for reporting on adverse events, near misses, and other errors. Although such reports could yield valuable insights for infection prevention teams, this information is often underutilized in supporting infection control activities.

One of the reasons for this gap is that information is often collected as unstructured data. Compared to structured data, free-text event descriptions are difficult to rapidly mine. They also demand hours of time for manual analysis and can be subject to reviewers’ personal biases.

And yet, unstructured narratives offer a treasure trove of insights for helping teams understand which interrelated factors affect infection control, and why. That’s where NLP—more specifically, centralized systems powered by NLP and machine learning—could help automate and improve identifying infection control breaches and surveillance of hospital-acquired infections (HAIs). Through this process, we can see that breaches are related to use of personal protective equipment (PPE), that there may be sterility issues with sterilized instruments, or that there is a trend in equipment malfunctions.

Piloting What’s Possible

ECRI’s own experiments with NLP modeling affirm the potential value of pursuing these capabilities. They also underscore the investment of time and expertise it takes to realize that value. 
In our experiments, we applied predictive models and then had to manually review results to further improve model performance. Even before that, we found that the modeling itself was quite time intensive. In fact, the ECRI dataset—representing just one year of PSO-reported data—required at least eight hours to run on a remote server to create results. Running a model against the full universe of more than 5.5 million reported events would demand an extensive amount of computational time and resources.

Four Key Considerations

Should your healthcare organization invest in NLP and other forms of AI to help extend and “industrialize” infection prevention efforts? As you evaluate the options, here are some obstacles to overcome and complications to consider: 

1. Skills and expertise. The first major consideration is the computer science, coding, and troubleshooting expertise needed to build these models. Ensure that you can hire or otherwise engage such professionals—and that they stay tightly aligned with your own experts in risk management and infection prevention.
2. Data quality and reliability. The effectiveness of any AI-powered system correlates with the quality of the data that feeds it. To be well trained, your NLP models must have datasets—data that is consistent, highly structured, and clean. Supervised machine learning also requires enough labeled data. This pre-processing is imperative for any effective data analysis to occur.
3. Time requirements. ECRI’s experiments demonstrated that even with access to the right expertise and high-quality data, the process of training an NLP model is time consuming and labor intensive. It involves manual validation of thousands of event reports—a significant undertaking requiring a collaborative, iterative process between data scientists and infection prevention experts. Account for this time and effort when planning your initiative.
4. Responsible AI. Although computer models are objective, they are only as good as the data they are trained on. Biases in training data will influence the predictive outcomes of NLP models. Examples may include not having datasets that are generalizable to the population of interest, as well as missing data, misclassified data, and even measurement errors. Consider, for example, that not all healthcare systems use electronic health records (EHRs), and those that do may have greater resources and/or serve different populations. Thus, patients from vulnerable populations, those with low socioeconomic status, those with psychological issues and/or housing insecurity, and even immigrants can be underrepresented in training data. These biases can carry through to predictive outcomes generated by a model—perpetuating harmful stereotypes and/or reinforcing social inequalities.

Closing Thoughts

As NLP and other forms of AI continue to advance, organizations have growing opportunities to develop robust, scalable solutions for fully mining large volumes of narrative-rich patient safety event reports.

Keep exploring opportunities to collaborate across your invention prevention, risk management/patient safety, and data science functions. With the right commitment of resources and a thoughtful approach to responsible AI, these solutions can be a powerful tool for driving systemic change to promote patient safety.

Learn how ECRI can help you enhance safety and quality with our infection prevention and control expertise.