When a patient is brought to the A&E (Emergency room) by ambulance, the patient is quickly transferred to a waiting bed where a team of doctors and nurses begin treatment, stabilizing the patient, ordering tests, and starting treatments. It’s a familiar scene from television dramas (and, unfortunately, to any parent).
Less visible is the rapid transfer of patient information from the ambulance to the hospital: the history, vital signs, and procedures during the initial encounter on-scene with the medics. We’ve gotten much better at transmitting demographic and device data ahead of the patient’s arrival. But data alone is not enough.
Watch closely, and you’ll see the medic and physician have a brief conversation about the patient during the handoff. They are sharing a structured story, explaining the incident’s circumstances, the treatment rationale, the observations that give meaning to the charted measurements.
Long ago, I led a research project to capture and interpret these narratives: we combined a voice recorder, transcription software, and medical language parser to try to extract the “facts of the case” from the narrative. The technology was too crude for more than a proof of concept demo and a patent, but I’ve always had a soft spot for the idea.
Historically, machine transcription and translation based on linguistic rule-based approaches has always been fragile and unreliable. I remember reading to Dragon for hours, training it with prepared texts and hoping it would get smarter. We took a license to MedLEE, an expert system that extracted clinical concepts from medical texts.
Today, transcription and translation software uses statistical approaches that sidestep the whole question of structure and meaning. Given a snippet of spoken or written text, computers simply search vast libraries of paired ‘voice-text’ and ‘native-foreign’ documents to find a match.
As the libraries grow, so does the accuracy of the process. Like any ‘big data’ project, the process needs well-curated content to get started. But as it grows it can ‘learn’ and improve itself through simple removal of inconsistencies (cleaning our errors) and recognition of similarities (colloquialisms and metaphors).
The growth of medical information systems means that vast libraries of ‘medical narrative-patient chart’ pairs are available. Although held behind patient privacy walls, 
de-identified data should be useful in developing better systems that mimic the human process for extracting the ‘facts of the case’ from the stories that medics tell physicians in those first moments at the doorway to the hospital.
It may add an important dimension to understanding the simple ‘flow sheet’, the chart of vital signs data and medications over time. Medical stories can explaining why things happened alongside the simple ‘what’ .
