Say 'Ah' for two seconds, AI can detect aspiration risk

Sometimes, food or drink can accidentally go down the wrong pipe, triggering sudden coughing that can leave the person embarrassed and those around them concerned.

Some people, however, are more prone to such episodes than others.

Seoul National University Bundang Hospital said Thursday that it has developed a method to identify people at risk of aspiration by having them say “Ah” for two seconds after eating.

A research team led by Ryu Ju-seok, a professor of rehabilitation medicine at the hospital, developed an artificial intelligence-based program that can distinguish people who are at higher risk of aspiration.

Aspiration occurs when food, liquid or saliva enters the airway instead of the esophagus. While coughing can clear the material in many cases, repeated aspiration can allow food particles or pathogens to remain in the lungs, potentially causing complications such as pneumonia.

According to the research team, aspiration is more likely to occur when the larynx weakens and fails to adequately protect the airway during swallowing, particularly among older people and those with neurological disorders.

Previous screening methods have relied on X-ray-based examinations to observe swallowing. The new method instead detects changes in vocal cord vibrations that can occur when food remains around the vocal cords or when laryngeal function is weakened.

The team collected voice samples from October 2021 to February 2023 and used the data to develop and refine the AI model. It then evaluated the model’s ability to identify people at risk of aspiration.

The model achieved a sensitivity of 81.78 percent in detecting people at risk of aspiration and a specificity of 80.02 percent in correctly identifying those without the risk, according to the hospital.

“Aspiration occurs during the everyday process of swallowing food or saliva, so it is important to be able to continuously monitor the risk in daily life,” Ryu said.

“We plan to conduct follow-up studies, including validating the performance of the AI model using data from multiple institutions, to develop it to a level where it can be applied in clinical settings.”