Data sorting for electronic noses.

When a chemical spill releases an unknown mixture of environmental contaminants, quick and accurate identification of toxic components is an essential part of any response to such an emergency. One promising approach to identifying troublesome vapors is the combination of a hand-held chemical-sensor system--an electronic nose--and a novel classification scheme modeled on the way people visually group nearby objects into clusters.

Physicist Gordon Osbourn and his coworkers at the Sandia National Laboratories in Albuquerque, N.M., studied how people decide whether various patterns of points drawn on a sheet of paper or displayed on a computer screen belong together in clusters. The researchers discovered that study participants acted as if they were superimposing an imaginary two-lobed, or dumbbell-shaped, template over each pair of points. If other points intruded into any part of this region, a person would conclude that the original two points do not belong together.

Why that particular strategy works in people remains a mystery, Osbourn remarks. Nonetheless, a computer program applying the same strategy allows researchers to identify clusters in data sets and to determine whether a new data point belongs to any of the clusters already in a database. The Sandia team calls the resulting data classification scheme VERI, for visual-empirical regions of influence, and have applied for a patent.

Using a dumbbell template modeled on the shape discovered in experiments with people, the VERI computer program applies appropriately sized templates to all pairs of points and then links any two points that meet the criterion for belonging to a cluster. In two and three dimensions, the resulting patterns closely resemble the clusters that people would identify by eye among the data points. The computer program has the advantage that it also works with more complicated, higher-dimensional data--something the human eye can't do.

A chemical-detection system, for example, may have six sensors, each of which gives a reading for a given concentration of a particular substance. Those six pieces of data can be thought of as the coordinates of a point in six-dimensional space. Changing the chemical's concentration changes the sensor readings. Hence, each set of sensor readings for a different concentration defines a new point in six-dimensional space. Together, those points form a pattern, which can be detected using the VERI technique. Other chemicals produce different patterns.

One can apply the dumbbell test to a new point, representing freshly collected sensor readings at a chemical spill site, along with all points already in the database. It's then possible to identify the substance as one detected earlier or show there is no match. The latter capability is important for avoiding false alarms under unusual conditions out in the field, Osbourn says.

The researchers are incorporating the VERI data classification scheme into a lab-on-a-chip chemical-sensor system designed by Sandia for battlefield use. The same technique may also prove useful for identifying atoms in scanning tunneling microscopy, interpreting remote-sensing data from satellites, and differentiating tissues of the body in magnetic resonance images.