“Doctor, what’s wrong with me?”
Sometimes a giveaway symptom – the intensely itchy, fluid-filled blisters of a chickenpox rash, for example – makes diagnosis relatively straightforward.
But many common symptoms – fatigue, abdominal pain, headaches, and more – are typical of a huge number of conditions. And patients may show combinations of vague symptoms that can be interpreted in different ways. In those situations, arriving at a diagnosis can be a challenge.
When clinicians have to piece together various observations, and perhaps test results too, to reach a diagnosis, the information stored in a patient’s health record is key. But finding the right words to describe vague symptoms can be tricky, and the language used can vary even more when the entries are made by different healthcare professionals over several appointments. These complexities add to the challenge of reaching a diagnosis.
Anoop Shah, THIS Institute fellow at University College London and consultant physician at University College London Hospitals, is developing tools to help both clinicians and researchers make the best use of the information stored in patient health records.
Extracting information from health records
Anoop’s interest in health records was first piqued during a summer job he held as a medical student. “I was asked to look at drug doses in patient records, and found that even numerical dosages were entered as free text. I worked out how to convert those text instructions into numbers,” he says.
Anoop explains that as most health records are now held electronically, details entered as free text can be translated ‘behind the scenes’ to defined codes, using natural language processing – which involves using computers to process text and extract meaning in a structured way. For example, it’s used in Google searches to interpret your query: if you enter “28 degrees Fahrenheit”, the search engine will interpret the query and calculate the conversion for you.
Recognising the exciting opportunity this approach presented, “I became enthused about the potential of using patient data on a large scale,” Anoop says.
For his THIS Institute fellowship, Anoop is generating evidence to show how natural language processing can improve recording of diagnoses. He aims to build an approach that will work for different diseases by developing methods to extract and organise the information in medical health records and converting data into a standardised format. Though it will be useful across many conditions, his fellowship focuses on the example of heart failure.
Using natural language processing to create structured data from the outset holds great potential. “If the system can process text as soon as it is typed or dictated, it could enable clinicians to enter structured information in a more intuitive way, and save time,” Anoop explains. Better clinical records could mean that healthcare professionals don’t have to enter data manually for other purposes, such as monitoring the quality of care, or research studies.
Helping to improve care
Anoop’s work will support clinicians to achieve accurate diagnoses and help them to respond promptly should new information emerge. “Diagnoses aren’t necessarily fixed,” explains Anoop. “For example, if a patient with a suspected stroke turns out to have something completely different when they are assessed in hospital, there is no easy way to record that the understanding of the diagnosis has changed. That needs to be addressed.”
Getting better at data might also help with tailoring care more precisely for patients. For example, recent work reveals that breast cancer is at least eleven genetically distinct diseases. “If we’re going to provide more specific treatments, we need to be able to record medical conditions more precisely,” says Anoop.
Having better structured records should also help with sharing information among healthcare settings. When patients need to go to a different hospital or clinic for a specialist treatment, or perhaps move to a different part of the country, transferring the data isn’t always straightforward. That makes it harder to optimise a patient’s care – and it doesn’t help the patient experience. “Patients find errors in their records. And they have to keep repeating their story each time they see a new clinician,” says Anoop.
The structured information Anoop is building will enhance rather than replace current approaches. “Free text will always be needed because it’s the most authentic way of recording patient stories and information about how the condition has affected them,” he says. But natural language processing could help to understand these stories in new ways.
Watch Anoop’s talk given during our THIS Space event below:
.png)
