Robotics, chatbots, artificial intelligence, and machine learning algorithms are the new norm in healthcare. These innovations and other factors are the basis of a so-called “new medicine,” that is predictive, proactive, and data-based. Technological advancement, and the related shift from products to services/solutions, democratizing access, explosion of medical data—this all puts pressure on medical businesses to transform.
PwC predicted that the market for artificial intelligence in healthcare will expand rapidly:
People usually ignore minor ailments and don’t visit the doctor until it’s an emergency. Unfortunately, the emergency room is the most expensive and time-consuming way to treat people, especially the uninsured. Medical facilities perform triage, meaning they evaluate the degree of emergency to prioritize the most urgent or time-sensitive treatments. Traditionally, doctors have relied on clinical judgment to identify high-priority patients needing intensive care. However, the practice of medicine has no common and standardized criteria or framework to make these decisions.
Triage’s major, unpredictable challenge is the human factor, which is prone to doubts and errors. Over-triaging, when a doctor doubts their own evaluation and recommends over-treatment, is not unusual. This results in people being sent to unnecessary, expensive, and time-consuming intensive care treatment.
“In those cases, the patient may be unnecessarily exposed to multidrug-resistant bacteria and have an increased overall length of stay. On the other hand, under-triaging means a patient that should have been in the ICU is sent to a recovery or step-down unit, and the opportunity for quick rescue of a deteriorating condition is delayed because monitoring is not as intense.”
— Marcovalerio Melis, MD, FACS, an Associate Professor of Surgery, New York University Langone Hospital System, New York City, and co-author of the pilot study
Under- and over-triaging can lead to irreversible, negative consequences. Medical facilities can avoid both by using artificial intelligence to triage patients. AI models are shown to be effective in predicting risk of complications, possibility of cardiac arrest, and likelihood of CT (computer tomography) scans identifying medical problems.
“The only way to know if you need to see a doctor, is to see a doctor. This maxim can and should change. Patients are now going to Google all the time. They prefer an imperfect answer now than a perfect answer in a week, and need to know whether they should be worried.”
— Josep Carbó, Partner at Mediktor and Vice-President at Barcelona Health Hub
Triage machine learning has proven to be an effective tool. According to a research paper by a team at John Hopkins University, ML-based e-triage improves risk assessment and categorization of patients; predictive analytics adds accuracy to the triage decision-making process. When working with data, statistical analysis quality is critical—redundancies and duplication must be removed.
Classification Metrics: Building a Triage Algorithm
Suppose we are building an algorithm to prioritize a large flow of patients. First, we have to distinguish admitted patients (who will stay in a medical facility overnight and consume various resources) from temporary patients (will leave in a day). At first, this classification might sound too simple given the wide range of situations and diseases, but the two groups are quite definite and distinct, and there is no need for a more granular approach.
Let’s take data from a triage study conducted at two hospitals:
From even a brief review, it is obvious that the data is “dirty,” full of confusing and irrelevant tags, features, etc. We must clean the data and eliminate features which will not contribute to the process. Since we have nulls, we must impute median volumes, non-uniform features must be classified into broader categories, and categorical features must be given dummy variables.
Having done that, our clean data looks like this:
In this context, machine learning algorithms can help computers learn complex and non-linear interactions between variables. To do so, we optimize the error between predicted and real outcomes.
For building a classification model, we may use the following algorithms:
Building a Data-driven AI Model
When it comes to mass casualty incidents, the survival rate of patients is determined by a huge number of factors; time of triage and the quantity of medical personnel are among the most important. As hospitals cannot rapidly adjust the quantity of personnel, it becomes critically important to classify patients quickly. Triage can be sped up with AI models of survival prediction.
Al models can be data-driven or model-driven:
- Data-driven AI is used to build a system for detecting the right answer based on previously seen examples of question-answer pairs. This approach requires a huge dataset with perfect labeling.
- Model-driven AI, also known as symbolic AI, is used to capture knowledge and make decisions. Model-driven AI can process images, deconstructing them into colors, lines, and shapes, which are then compared to explicit rules.
A model-driven approach requires an immense number of rules, as well as exceptions to these rules. Naturally, it is hard to capture them all and this sometimes leads to model-driven AI being dismissed.
“When working on healthcare for our clients, it is our major goal to deliver user-centric, accessible and plain solutions. Innovative tech and algorithms can have a positive impact on all stakeholders, bringing fresh approaches and a new level of accuracy. Our AI-based triage use cases show we can unlock value for healthcare companies and patient journeys.”
— Vlad Medvedovsky at Proxet, a custom software development solutions company.
Healthcare experts are creating and benefitting from AI innovations. For example, doctors at Stanford are currently experimenting with a new ML-based system. In contrast to existing ML applications, such as habitual chatbots, this new system will assist medical staff in cases of high inflows of patients. This system will identify high-risk patients, and help clinicians and nurses reach consensus when there are differing opinions on health interpretation. As Stanford physician Ron Li said in an interview with Fortune, the software will function as a “shared mental model.”
Another project, NinesAI, can detect intracranial hemorrhaging and flag CT scans for urgent investigation by clinicians. NinesAI recently received clearance from the FDA and is now available for busy radiology departments. Infermedica, a good triage example, closed a $10 million funding round to support R&D and boost symptom-checking features.
NinesAI and Infermedica improve healthcare service delivery by complementing the work of doctors and nurses. Developers, data scientists, and project managers have joined AI/ML efforts at Proxet because they are positive that AI triage tools will change the medical landscape.
Dive deep into the technicalities of embedding models, vector databases, and optimization strategies to revolutionize information retrieval
Explore the correlation between latency and token generation in Large Language Models. Learn how prompt size impacts response time.