Can AI accurately identify patients with respiratory symptoms?

Some doctors may also be susceptible to time pressure or patient demands which incentivize unnecessary antibiotic prescriptions, resulting in increased drug resistance. 

With this in mind, a team of researchers at the University of Iceland trained a machine learning model (MLM) to analyze the symptoms of patients with respiratory infections and classify them to help determine when patients need to go to see the doctor in person. 

Here’s how that works.

Respiratory Symptom Triage Model

Machine learning (ML) is a branch of artificial intelligence (AI) that uses large datasets to train algorithms to perform tasks or solve problems without requiring explicit programming for each specific scenario. In effect, the algorithms learns how to approach new scenarios and data, based on the data it has previously received.

A training model is an algorithm that is designed to learn from data and make predictions or decisions based on what it has learned. 

During the training phase, the model analyzes the input data, identifies patterns, and adjusts its internal parameters to minimize the difference between its predictions and the correct outputs given in the training dataset. This adjustment process is often referred to as "learning," and it is what allows the model to generalize its knowledge from the training data to make accurate predictions on new, unseen data.

In medicine, triage is a process by which medical professionals pre-evaluate the symptoms of patients to determine which ones should receive treatment first, based on the severity of their condition. The goal is to allocate resources and attention to the most critical or time-sensitive situations in the most efficient manner.

Researchers have combined the two processes to develop the Respiratory Symptom Triage Model (RSTM) — a machine learning model specifically trained to aid in triaging patients with respiratory symptoms and determine whether they need further consultation or treatment.

First, the RSTM was trained to differentiate between upper and lower respiratory tract infections by assigning a value between 0 and 1, where 1 indicated a higher chance of a lower respiratory infection.

For training data, the model used 17,177 medical records of patients with respiratory symptoms who had been diagnosed with acute bronchitis, bacterial pneumonia, influenza, chronic obstructive pulmonary disease (COPD), asthma, or the common cold; although only the latter was considered an upper respiratory tract infection only for the purpose of the study.

After that, the researchers applied cross-validation and intrinsic validation methods to assess the RSTM’s performance, and fine-tuned the training dataset with new parameters. They also reduced the number of medical records to 2,000, which fed the model with about 26,971 annotations. 

After this, the RSTM divided 1,915 patients into 10 different risk groups for lower respiratory tract infections: 1-5 were low-risk groups and 6-10 were high-risk groups. 

In each group, the researchers checked the average levels of c-reactive protein (high levels of c-reactive protein in the blood suggest inflammation); the distribution of medical diagnosis codes; the number of patients who went back to their primary care doctor or to the emergency department for re-evaluation within one week; the number of patients who had been prescribed antibiotics; how many patients were referred for a chest X-ray and whether these turned up signs of pneumonia or other unexpected findings.

The classification made by the RSTM was found to be relatively accurate. No patients in the low-risk group had symptoms of pneumonia or had chest X-rays showing signs of pneumonia. The low-risk group contained only those with colds and acute bronchitis, while the algorithm accurately placed those diagnosed with pneumonia and COPD in the high-risk group.

Broader implications

In spite of this success, the RSTM and similar systems may not be appropriate for broad use. After all, because the system learns from medical records, it is important that the training data is accurate, error-free and unbiased.  

The researchers anticipated that the RSTM could fall into selection bias when using medical records containing fewer than eight symptoms. They also warned about availability bias: certain data points —such as the chest X-rays and CRP values— might be overrepresented or underrepresented in the datasets used to train the RSTM.

But they also suggest that the quality of training data can be improved in future trials, for example, by adding annotations from other physicians on the same patient.

If additional data is added this could make the RSTM reliable enough for clinical use. In the future, systems like this could even be used as an online triage tool to reduce unnecessary visits to the doctor’s office. It could also lower overall medical costs and risks by requiring fewer chest X-rays and reducing the prescription of unneeded antibiotics.

There are already AI algorithms that can analyze vast amounts of medically relevant data (such as medical history, symptoms, lab results, and imaging data) to calculate the risks of certain diseases, diagnose conditions, and even recommend personalized treatments based on each patient’s unique characteristics. 

AI has also been used in medical imaging analysis, as it can review medical images, such as X-rays, MRIs, and CT scans, to detect even the most subtle abnormalities and identify potential diseases more quickly and more accurately than human physicians. 

The risks and benefits of further use of AI in medical diagnosis are similar to those of the RSTM, namely reduced medical costs and faster diagnosis, set against the risk of bias and unreliability and the loss of human skills and connection.

However, the proper use of AI in diagnosis, alongside human medical professionals could lead to earlier detection of disease, lower costs, and better patient outcomes across a wide range of medical specialties.