Going Beyond Natural Language Processing with Analytics at Medal

How would you handle a new patient who arrives with a batch of folders containing 4,000 pages of medical records? This is the kind of information overload faced regularly by health clinics, hospitals, and rehab facilities–unless they’re in even the worse position of getting no records at all. Scads of healthcare startups are tackling this information problem. This month I talked to a company claiming a particularly sophisticated solution to ingesting and quickly analyzing patient records to produce high-value insights.

Lonnie Rae decided she could help more people by pursuing a leave of absence from medical school to found Medal, a company that uses analytics to funnel the huge jet of information into a useful summary. Like many such solutions, Medal runs records through a natural language processing (NLP) tool that is finely tuned to recognize medical terminology, including abbreviations. But that is just the start of the analytics.

The next stage of Medal’s processing is the disambiguation of words and ideas. For example, the word Huntington’s could refer to a diagnosis, but it could be part of a postal address instead. Abbreviations such as AFP mean different things in different contexts (Alpha-fetoprotein versus Atypical Face Pain). So Medal looks at the context to establish the real meaning of a word or phrase. It can determine whether a medicine is being taken currently, was taken in the past, is being newly prescribed, or triggers an allergic reaction. Medal can not only recognize a phone number but can determine whether it belongs to the patient, the caregiver, a clinic, etc.

I asked Lonnie what technical means Medal uses to pull information from the wide range of sources they support: they work with databases of many types and formats such as the CCD, along with free text in flat files or faxes. Lonnie answered that instead of trying to use fancy column matching and other ETL techniques popular in the database industry, Medal just dumps all data as free text into its processor, which looks at 300 different elements of the context for each word. I found this to be an impressive affirmation of the power of modern AI.

Medal also provides innovation in patient matching, because records may come in without clear identification. Techniques for patient matching are fairly standard by now because it is another big neck pain for the health care industry (thanks, I feel compelled to complain, to the lack of standardization and lack of patient control over records). Patient matching usually involves comparing up to 17 data fields, such as name and birthdate, weighing how closely each match, and weighing the likelihood that it would change. Medal is improving patients matches based on HIPAA identifiers, applying its highly accurate system to the task of patient identification.

A third task sometimes required is de-identification, because sometimes records are shared with clinicians or support staff who shouldn’t be told whose records they’re looking at. De-identification, like patient matching, is complex but benefits from a lot of research and knowledge of best practices.

The final task is the most impressive. Medal determines the most important information the clinician needs to see right away and provides it as a summary. This includes simple demographic information such as name and address but also diagnoses that stand out as requiring urgent attention. Lonnie cited the example of nausea: it could be a normal side effect of cancer treatment or a sudden and life-threatening condition for an emergency room patient. Medal does its own triage and suggests to physicians who has serious conditions. Lonnie calls Medal a “unified, contextual” service.

How successful is all this analysis? Lonnie says that within two days of receiving and processing a data set, Medal can produce output with 99% accuracy. In their pilots, clinical and administrative staff achieved 20-fold increases in productivity on the tasks that Medal assists with. For example, preparing and reviewing clinical case presentations with colleagues might be reduced from over 40 minutes to under five minutes. I look forward to seeing Medal’s results when it has been out in the field for a while–and to see what they do to make the service even more precise and accurate.

About the author

Andy Oram

Andy Oram

Andy Oram writes and edits documents about many aspects of computing, ranging in size from blog postings to full-length books. Topics cover a wide range of computer technologies: data science and machine learning, programming languages, Web performance, Internet of Things, databases, free and open source software, and more. My editorial output at O'Reilly Media included the first books ever published commercially in the United States on Linux, the 2001 title Peer-to-Peer (frequently cited in connection with those technologies), and the 2007 title Beautiful Code. He is a regular correspondent on health IT and health policy for HealthcareScene.com. He also contributes to other publications about policy issues related to the Internet and about trends affecting technical innovation and its effects on society. Print publications where his work has appeared include The Economist, Communications of the ACM, Copyright World, the Journal of Information Technology & Politics, Vanguardia Dossier, and Internet Law and Business.