Google has developed a system for diagnosing diseases based on cough sounds With the help of AI

Researchers from Google Research, in collaboration with colleagues from the Center for Infectious Disease Research in Zambia, have developed a machine learning system aimed at diagnosing lung diseases based on cough sounds. In their study, they utilized YouTube videos to train the system.

The Google team named their new system Health Acoustic Representations (HeAR) and initiated its development after healthcare workers reported that they could often discern the presence of COVID-19 during the pandemic by the sound of a cough. Other researchers are also engaged in similar efforts, hoping to create systems capable of detecting a wide range of diseases from cough sounds.

Google adopted a different approach to disease diagnosis compared to other teams. Instead of training an AI system using labeled records identifying specific diseases, they employed an approach very similar to the one used to create large language models (LLMs) like ChatGPT.

Their system converted numerous YouTube recordings of human sounds, including normal breathing, gasping, or coughing, into spectrograms. The team then masked out specific parts of each recording and tasked artificial intelligence with predicting the missing segment, akin to how large models learn to predict the next word in a sentence. The result was a foundational model that researchers suggest can be customized for various applications.

In their study, the researchers trained the model to detect tuberculosis or COVID-19. They then evaluated HeAR’s accuracy using a standard scale compared to random guesses. They found that it achieved a score of 0.739 on one dataset and 0.645 on another for detecting COVID-19, and an average score of 0.739 for tuberculosis, outperforming results obtained from other systems.

It’s important to note that news materials cannot replace a doctor’s diagnosis. Before making any decisions, it’s advisable to consult a specialist.