As a doctor or nurse have you ever scrolled through a patient’s records desperately trying to recall the last time they came in with a cough, complained of feeling tired or fell at home?
You know the record is there somewhere or maybe there is just a voice in the back of your head saying: “There’s something more to this.” Maybe you just gave up looking, risked frustrating the patient instead and asked them to recount their history again, just because it was quicker.
Wouldn’t life be a lot easier if the patient record did the checking for you? Wouldn’t it be great if, as soon as you typed “cough” or “fall” or “migraine”, the electronic health record presented all the previous records for respiratory, neurological or orthopaedic consultations?
Tagging Free Text Notes
We’ve been listening to the primary care GPs that use Vision’s patient records system and we know finding what you need in the patient record quickly is a perennial issue.
But not for much longer.
Dr Jonathan Behr, Vision’s former Chief Medical Director, has kicked off a ground-breaking research study with the help of Dr Ben Brown, a GP using Vision and academic in Manchester University’s Centre for Health Informatics, recruiting 50 general practitioners and John Wakefield, a machine learning expert and founder of RoboSoup.
Using Vision to access the free text field in the patient record the team has started “tagging” 100,000 consultation notes. Each note is being given a primary tag to denote the principal reason for the consultation. These have been divided into 27 specialities ranging from allergic reaction and endocrine to renal and urological. Then each consultation note is also being given multiple secondary tags from the same list of 27 specialities.
This library of “tagged” consultation histories will then be used to teach the algorithm created by RoboSoup how to tag free text consultation reports in the Vision system automatically. The end result will be a system that will understand that words like 'cough', 'chest' and 'wheezing' all relate to the tag respiratory. Therefore, whenever a doctor starts typing in those words to the patient record, the system will automatically find similar episodes in the patient record.
Vision and the Medical Research Council’s Proximity to Discovery Fund have jointly funded the project. The money is going towards paying for the estimated eight hours it will take every GP to tag 2000 reports.
First Steps towards AI in General Practice
The main aim for the research is the summarising of patient information for GPs but long term the new approach could:
- Reduce workload
- Save time during consultations
- Highlight patterns of consultations and symptoms
- Flag potentially serious conditions
Dr Brown explained:
“This could proactively alert GPs to hidden patterns or recurring symptoms which may indicate something more serious. The primary use case is that by tagging the free text excerpts from the history when a clinician is typing in a new field, it will bring up previously related consultations. Ultimately, we want this innovation to help GPs work more effectively and efficiently to provide good quality patient care through more usable electronic health records.”
Patient Data to Help Other Patients
We expect the consultation note tagging research to be the first of several artificial intelligence initiatives.
The next obvious step on from tagging reports would be tagging the words in the free text with SNOMED CT codes.
What is even more exciting is the prospect that free text in GP records may one day allow Vision to look more deeply at the relationship between a patient and their data. What does that mean? It means GPs may be able to predict what a patient’s health pathway is likely to be.
Science fiction? Google published a paper in Nature Partner Journals in May demonstrating that patient data could be used to accurately predict patients’ admissions and the risk of death. Using raw data from the patient electronic health records, Google developed an artificial intelligence neural network capable of predicting the course of a patient’s disease and the risk of death during a patient’s hospital stay.
Engineers trained the network on 216,000 records, referring to more than 114,000 patients, who had been hospitalised for at least one day at either the University of California, San Francisco, or the University of Chicago Hospital. The network was 95% accurate in predicting a patient’s risk of dying while in hospital.
In time, we want to do a similar analysis and look at specific patients and patients similar to them and then advise doctors what treatments benefited the similar patient.
“What gets me excited,” Dr Brown explained to me, “is the prospect of actionable information. I want to see if patients could have better outcomes if prescribed treatments tailored to people like them. Most clinical decision support is rules-based, not probabilistic and isn’t actionable – they just say a patient hasn’t had this or that treatment."
“A better system would help nurses and doctors make better decisions in collaboration with patients."
“Electronic Health Records provide reminders, for instance, that a patient’s diabetes is uncontrolled. But they don’t say specifically how to address that issue. In the future, they could suggest treatment changes because patients with similar characteristics have benefited from those medicines or other intervention. And this could be based on real-world evidence obtained from electronic health records, rather than randomised controlled trials whose findings often don’t translate to the type of patients we see in primary care."
This tagging trial is just the first step to creating a platform to much more exciting things and what is great from my perspective is that it shows Vision is willing to invest in the future.
Found this interesting?
Watch the 4 minute video with Dr Behr from the HSJ website, he describes the benefits of adopting a population level healthcare approach to improve healthcare at scale.
Find out more about population health
Vision's population health solution is Outcomes Manager. A growing number of CCGs, health boards and public health teams are using it for:
- NHS health checks
- Identifying patients with undiagnosed long-term conditions
- Managing care for specific diseases