AI Now Has The Ability to Smell And Can Detect Illness Through Human Breath

AI has been shown mastery in some of human’s essential senses such as sight, as seen with driverless cars, and sound, as with Apple’s Siri. But the sense of smell is not something well developed with humans and could gain a great benefit from AI. Researchers from the University of Edinburgh and Loughborugh University in the UK have neural network based AI that can analyze human breath to identify various illnesses.
 
Up until now, the detection of compounds in gases have been performed by specialized machines called the gas-chromatography mass-spectrometers (GC-MS). These machines are used to detect a specific type of molecules called volatile organic compounds (VOCs). The GC-MS analyzes the air using a thin tube called a capillary column that is lined with stationary phase, which is a type of material that separates the compounds in the air. Different compounds interact with the stationary phase in different ways which creates a unique pattern; this is what is used to identify each individual molecules in the air.
 
These unique patterns are mapped on a graph and reveal various peak sequences. Even the smallest of spikes are essential for analyzation. Hundreds of compounds can exist in the human breath, including those that are present when a person is sick.
 
The analysis of these patterns are conducted by experts and due to the sheer size of the data they are dealing with, end up being a very time consuming process. This is where AI comes in. Using the latest developments in neural networks, the team trained the AI to recognize patterns in aldehydes, a compound that has association with stress conditions and ailments.
 
The air is first measured through the GC-MS, producing a two-dimensional data matrix. As the compounds were first identified by their chemists, the data was fed into the deep learning platforms. Researcher Angelika Skarysz reported, “Using NVIDIA Tesla GPUs and the cuDNN-accelerated Keras, and TensorFlow deep learning frameworks, the team trained their neural network on data from participants with different types of cancer receiving radiotherapy.”
 
The team noted that the network performed better when GC-MS data was reorganized through one-dimensional filters and fed through three channels of high, medium and low intensity signals. In order to increase the efficiency of the system, the research team has utilized data augmentation over 100 times. This is recorded to be the first time the analysis of raw GC-MS data of ion patterns and compound detection has been performed by machine learning.
 
What took hours of analysis by human hands, took mere minutes by the neural networks with considerable increase in reliability. The system has even discovered errors by the human experts in labeling, demonstrating better performance averages than its human counterparts. The added benefit with neural networks are their ability to improve overtime as it analyzes more and more data. This also makes this system unrestricted to any substance and can be trained for minuscule amounts of volatile compounds.

Original Article: https://theconversation.com/ai-is-acquiring-a-sense-of-smell-that-can-detect-illnesses-in-human-breath-97627