无忧传媒

About 4.5 million patients聽聽in the United States. Transfusion side effects, including respiratory and circulatory reactions, are often hard to specifically聽. Scientists at the U.S. Food and Drug Administration (FDA) in the Office of the Commissioner and Center for Veterinary Medicine wanted to research methods that could be used to proactively monitor electronic health records (EHR) to discover new safety concerns by reviewing historical EHR data to identify previously聽.听

无忧传媒 worked with the scientists to examine more than 20,000 past transfusion cases identified in the聽. Applying advanced natural language processing methods to unstructured clinicians鈥� notes within EHRs, our team of data scientists helped them quickly glean valuable information about actual and possible transfusion reactions, paving the way to a much faster, more聽direct way of discovering side effects, improving clinical care, and saving lives.

The Challenge

An FDA core mission is to protect patients from harm by ensuring that medical products such as blood and聽blood products intended for transfusions are safe and effective. Safety monitoring traditionally involves聽collecting and reviewing reports鈥攕hort summaries of events that healthcare providers and patients submit to FDA. While these episodic reporting systems provide the core of critical adverse event information, , including social media posts, scientific literature, and EHRs.听

EHRs typically include both coded information and written clinicians鈥� notes about patients, including both observations and conclusions. Because EHR notes generally contain much more information than the coded data, they can capture critical information about side effects that are known, suspected, or not recognized as associated with blood transfusions and other therapies. Our problem is how to find these side effects amid the large volume of raw, unstructured text data in clinical notes.

The Approach

Using natural language processing, machine learning, and visualization, FDA鈥檚 and聽无忧传媒鈥檚 data scientists developed new ways to clean and organize free-text聽medical record data, preparing the unstructured notes for input into sophisticated聽algorithms. This 鈥渢ext mining鈥� unlocked the information-rich portions of the EHRs,聽enabling scientists to gather actionable insights. Inspired by the advanced text聽mining methods used to verify the true authorship of literary works attributed to聽William Shakespeare, the use case was called the 鈥淪hakespeare EHRs鈥� project.听

Processing the enormous amount of EHRs data required a high level of computing聽power. 无忧传媒 provided a cloud-based infrastructure and analytic tools to house聽and manipulate the data. In addition, as part of the text mining process, the team identified blocks of exactly-duplicated patient notes, which clinicians often copied聽and pasted from previous entries. Our solution, named 鈥渂loatectomy,鈥� filtered out聽the duplicate sections, which, if left untouched, can reduce the overall quality of聽machine learning. The team then applied machine learning methods to find the聽10,000 words or phrases (other than those that simply say the patient was transfused)聽that best distinguished notes about admissions that involved blood transfusion聽from notes about admissions that did not involve transfusion.

The Solution

With usable data in hand, our data scientists then identified terms and keywords that commonly occurred in the notes for the same admission and used them to聽describe topics. Those, in turn, were used to identify cases where patients experienced聽actual and possible transfusion reactions. In a matter of hours, our process聽resulted in an initial automated assessment of more than 20,000 transfusion cases,聽finding cases of transfusion reactions in the historical data immediately, far more聽than the 10 cases per year nationwide that were reported over the same years as the聽clinical data.听

In addition, using an open source search and analytics tool, we built an interactive聽data visualization platform. With this digital dashboard, epidemiologists could quickly search specific terms, view our analyses, and investigate specific EHRs with聽intuitive graphics and point-and-click filtering right from their laptops.听

Another important aspect of the solution is that it is translatable to other types of adverse events. By designing a solution that does not rely on specific interpretations聽of word meaning, but rather relying on statistical properties of clinical terms,聽 we聽believe this approach will scale well to other applications. In testing this hypothesis,聽we鈥檙e helping FDA scientists look at other applications from population health to individual care. For example, our data scientists are continuing to validate this聽advanced text mining framework by looking at historical EHRs and attempting to聽retrospectively uncover adverse events that were identified in the past. The team plans to publish the work in open scientific journals and the code in an open repository. Further work will need to be done to find out how adaptable the methods聽are for other EHRs data and formats.听

In the future with further refinements, the framework could be implemented into聽near-real-time operations. With a faster, scalable, and more comprehensive review of聽EHRs text notes, FDA could be more proactive鈥攔ecalling dangerous products, alerting hospital systems before outbreaks spread, and saving lives.