Inspiration
Wait times are extremely long in emergency departments with a large part due to incoming patients being unaware if there issue meets the severity to enter the emergency department priority queue also known as triage. Current methods of triaging involve having a triage nurse assess patients as a first come first serve basis, which can result in higher priority patients in transit to miss out on available hospital resources.
What it does
Comed's solution is twofold, patients can provide their medical history and symptoms prior to arrival to the emergency department and prepare the triage nurse to assess their position in the queue or provide course of action if the patient's concerns could otherwise be met at a clinic. Furthermore, we use cohere’s large language model (LLM) to enhance the triage process by providing suggested triage levels and once a patient is seen by a physicians, has a ‘Physicians View’ that suggests a diagnosis and a treatment plan.
Features
- Patient View: Can fill out their current medical emergency information online.
- Nurse View: Can access a patient's full chart, adjust the suggested triage level, and send a patient directly to a physician or add them to the priority queue.
- Physician View: Can access a patient's full chart, medical history, symptoms, interact with a chat that provides diagnosis and treatment (doctor can ask more questions to the bot/patient), and review the most recent medical literature.
How we built it
Using the cohere large language model (LLM), patients' input data is securely processed and provided to various cohere trained models as a prompt. To intake a patient's data and triage them appropriately, we trained Cohere's classifier model with thousands of emergency room data points, processing the patient data according to the Canadian Triage and Acuity Scale (CTAS).
Level No. | Description |
Level 1 | Immediate: life threatening. |
Level 2 | Emergency: could become life threatening. |
Level 3 | Urgent: not life threatening. |
Level 4 | Semi-urgent: not life threatening. |
Level 5 | Non-urgent: needs treatment when time permits. |
Example Training Data Set Visualization
The result is a suggested triage level and confidence value for the patient. Next, we utilized Cohere's chat model to determine a diagnosis, treatment plan, and relevant medical research for the physician to view and act as a co-pilot. The doctor can then review the results, engage in a back-and-forth conversation with the model, search for new information, and preview relevant medical research.
Cohere Chat API:
Lastly, we utilized Cohere's rerank model API (
rerank-english-v2.0) to provide relevant medical literature and research data, ranking them accordingly. This feature would allow physicians to keep up to date with current literature easily to help with research and diagnosis. The endpoint accepts a query and a list of texts, generating an ordered array with each text assigned a relevance score. const request = require('request'); const options = { method: 'POST', url: 'https://api.cohere.ai/v1/rerank', headers: {accept: 'application/json', 'content-type': 'application/json'}, body: {return_documents: false, max_chunks_per_doc: 10}, json: true }; request(options, function (error, response, body) { if (error) throw new Error(error); console.log(body); });
Accomplishments that we're proud of
We are proud to have created a minimum viable product showcasing the potential of how such a system could contribute to our healthcare system.
What's next for comed
Comed aims to partner with hospitals to continuously refine and test the market, assessing if this approach can truly address many of the pain points in the emergency room. This product is still untested, but with a more robust data model, we aim to improve the classification, chat, and rerank models.
This project was built at UofTHacks 11 within 36-hrs. Checkout our project link below:
