5

Are there standard rules or references for deciding whether a patient has a medical emergency, based on the symptoms the patient report? Are those standard rules international?

Details

I am in Germany and no health expert, but a data scientist creating an artificial intelligence that should detect a medical emergency based on patients' symptom reports.

For instance, is the following an emergency (taken from https://huggingface.co/datasets/BI55/MedText)?

A 24-year-old woman with schizophrenia presents with acute restlessness, fidgeting, and a feeling of being unable to sit still. She started taking aripiprazole two weeks ago.

DaveFar
  • 151
  • 3
  • 1
    I am pretty sure that you refer to what generally is called "red flags" in medical professions. There should be a great variety of signs and tests to be found under that term. – Philip Klöcking Oct 25 '23 at 13:01
  • Thanks a lot, @PhilipKlöcking, I found some very helpful documents (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6060920/, https://www.gponline.com/education/medical-red-flags, https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4953318/pdf/251.pdf) searching for "red flags". – DaveFar Oct 25 '23 at 15:08
  • Though those documents (see previous comment) are quite helpful, they are all but complete. I still don't know whether "A 24-year-old woman with schizophrenia presents with acute restlessness, fidgeting, and a feeling of being unable to sit still. She started taking aripiprazole two weeks ago." contains any red flags. – DaveFar Oct 25 '23 at 15:14
  • I guess the German term is "Alarmzeichen", for which I found another interesting but incomplete document: https://www.biham.unibe.ch/e67732/e335450/files335455/Gujer_AnhangzuRedFlags_Dissertation2013.pdf – DaveFar Oct 25 '23 at 15:17
  • 6
    I don't really think this is a good use of the Stack Exchange Q&A model, but I'm open to alternative opinions by others. Basically, if you want to do this sort of thing, you need to collaborate with medical experts and plan to pay them for their time. You shouldn't expect to substitute a Q&A resource. – Bryan Krause Oct 25 '23 at 15:45
  • 9
    It sounds like what you want is annotated data for training a large language model. The traditional way to get that is to pay physicians a lot of money to annotate them for you. Alternatively, if you had someone who was really good with an electronic health record's data model, you might be able to extract the clinical text of triage calls and then determine the ultimate outcome (ie, did they show up at the emergency department). Either way, you need some experts. – Ian Campbell Oct 25 '23 at 16:03
  • 2
    Wait, so the idea is a clinician evaluates a patient, types it in, and then the AI says "ah, emergency!" One would think the clinician already would have noticed .... – Azor Ahai -him- Oct 25 '23 at 21:54
  • A lot of this will depend on what the actual use case for the trained AI is. If you intend to do something like supplement emergency dispatchers, then yes, you can get some very consistent rules, but the training data you link is borderline useless, and you may be better off using an AI as means to direct an expert system than trying to do it all with ML. If you intend to be a replacement for a fully trained clinician though, no, you’re not likely to find consistent rules that are actually useful for this (because most of the consistent rules rely on information you probably can’t have). – Austin Hemmelgarn Oct 26 '23 at 02:24
  • 1
    The use of „black box“ AI systems in healthcare is also notoriously controversial. What is already being used and where improvements are definitely possible are simple and comprehensible algorithms for triage / dispatch settings. Such decisions are usually undertaken primarily by specialised medical personal but not necessarily clinicians, and here one could argue for a standardized system that does not hold all the unconscious biases humans have. – Narusan Oct 26 '23 at 09:10
  • 1
    @AzorAhai-him- Or maybe an operator of the national emergency telephone number answers the phone, the victim on the phone describes their situation, the operator types it in (they are already supposed to type it in anyway, AI or no AI); and then an AI can immediately suggest relevant possible medical emergencies. Actually it doesn't even have to be a very sophisticated AI. Note that most emergency phone operators are not clinicians, and in fact in many places they're not even medical professionals (especially if the police and medical emergency numbers are the same) – Stef Dec 06 '23 at 18:27
  • 1
    @AzorAhai-him- Or the phone operator sends a first-aid worker to the scene; the first-aid worker takes whatever observations they can, including respiratory and cardiac rythms etc; no clinician is on the scene, but an AI assistant could immediately suggest possible medical emergencies (especially if the AI has access to the victim's history as a medical patient) – Stef Dec 06 '23 at 18:30

1 Answers1

7

Taking the question as written, no there are not standard rules that universally apply.

Actually determining whether someone has a medical emergency is based on assessment by a clinician, which will typically include taking a medical history, usually some physical examination, and often supplemented by other investigations to reach a diagnosis. The diagnosis will determine whether the presentation is a medical emergency or not.

Taking a step back in the process, there are many 'red flags' which if noted within the symptoms should prompt investigations for particular diagnoses. These can be as broad as 'chest pain' (which would prompt for specific further details in the history), or 'sudden onset, severe chest pain, tearing in nature, which radiates to the back' which would prompt investigations for aortic dissection.

Prior to clinical assessment, there are triage systems that look to risk assess a patient to give the likelihood that they are presenting with a medical emergency, and which can be used to prioritise patients when resources are insufficient to see everyone immediately. There are a range of these which are used in difference settings, and are not universal or international.

In the UK, ambulance services use AMPDS (also used in the USA) or Pathways. Most hospitals use the Manchester Triage System. Many others exist, some on a national basis, or for mass casualty incidents.

The risk assessment/triage tools currently rely on seeking responses to specific criteria; the information might be included within free text, but without all of the required information they could not be utilised. The actual diagnostic process needs an interaction with the patient; red flags might be elicited at an early stage within that process, but relying on a limited statement about the condition would not reliably ensure the necessary information to identify an emergency would be present.

Michael
  • 689
  • 1
  • 12
  • 2
    Very good answer! But I feel like you could highlight a bit more that in the end it is the Triage systems that @DaveFar is interested in. The AI model he wants to train needs to beat conventional triage systems regarding both mortality as well as efficacy / resource allocation. In the end a very relevant question is the use case of the model: Is it to recommend users to go to a GP vs call an ambulance? Is it to assist dispatch operators? Is it triage patients in a clinical setting, or a mass casualty incident? You could structure your answer (which lists all relevant systems) a bit – Narusan Oct 25 '23 at 20:52
  • 3
    [cont'd] regarding use case. And then I think a differentiation needs to be made between the diagnostic process and the triage for prioritisation of first contact to medical professionals. At this point in time, it is still unreasonable to expect any model to be able to reliable and correctly order follow-up examinations and diagnose patients, but triage should become very algorithmic and standardised, and there are already triage algorithms and softwares on the market. – Narusan Oct 25 '23 at 20:56
  • Thanks a lot for this helpful answer, @Michael (+1). The AI model is actually a phone bot for medical practices, for appointment booking, prescriptions and the like. But if the patient wants an appointment for some symptoms containing red flags, the phone bot should not offer a regular appointment but detect an emergency. I guess this depends on the practice's specialty, but is a simpler task than a general triage system. Right? – DaveFar Oct 30 '23 at 09:00
  • @DaveFar - Re: "... but is a simpler task then a general triage system..." um, no. In my not-so-humble opinion, the system should be able to pick up ANY red flags, since the nature of the potential emergency can't be known. The system should be able to detect as wide a range of red flags as possible to ensure that emergencies [and potential emergencies] are captured and routed appropriately. For example, a Cardiology practice bot should be able to detect red flags that are due to diabetes, pulmonary issues, and gastro-intestinal issues as these may actually be cardiac or cardiac related. – Sharpenologist Nov 01 '23 at 15:54
  • @Sharpenologist: Good point, thanks. Are medical assistants in small practices sufficiently competent to do this? – DaveFar Nov 06 '23 at 10:32
  • @Michael: Thanks again for your good answer. You mentioned ambulant services and hospitals. Could you please also mention how small practices are detecting red flags? Then your answer would also cover the case most relevant for me, and I could accept your answer :) – DaveFar Nov 06 '23 at 10:35
  • 1
    @DaveFar - I can't speak for all practices, and I'm only familiar with how things work in the USA ... but ... that being said, yes they are. At the very least they are trained to know what they don't know - so if a situation is outside of their usual realm, they will hand over to a Physicians Assistant or a Certified Nurse Practioner [don't know if these exist outside of the USA, but they are highly trained and can fully assess nearly anything]. – Sharpenologist Nov 06 '23 at 13:19