Technology

Technology & Services: KDE™ Tutorial: Overview

Overview | Applications / Classification | How it Works | Supporting Software | Publications

The Knowledge Discovery Engine®
A Powerful, General, Classification Algorithm

Applications

Correlogic’s own research focuses primarily on the application of the KDE to the identification of biological patterns that classify diseased from non-diseased states and the deployment of those patterns for use in commercial disease diagnostics. Correlogic’s lead products are cancer diagnostic tests. We have effectively applied the KDE to classification of data derived from:

  • mass spectrometry [4 – 13]
  • hyphenated chromatography techniques including GC-MS, LC-MS and GC-FAIMS [2,3]
  • NMR
  • gene expression data
  • protein array data

However, the KDE potentially addresses any signal streams -- such as those generated by financial markets, processing facilities, geographic surveys, or weather stations, among others.

How the KDE Classifies Data Sets

The KDE outputs a collection of patterns, or “models”, represented by an N-dimensional cluster map. The value of N is the number of features that define the pattern. For example, a 3-feature pattern (N = 3) results in a cluster map in 3 dimensions. Each cluster is defined by its center, the centroid, and a decision boundary of defined size, much the same way as any given circle is defined by its center and a specific radius (Figure 1). In addition, each cluster is assigned a specific phenotype, or classification state (e.g., diseased or not-diseased). To ensure that all data is scaled appropriately, each data point is normalized prior to constructing the cluster map.

Once a model of N features is created, unknown samples can be scored. The KDE extracts the appropriate N features from the unknown’s data stream, normalizes their intensities, and plots their coordinates in the N-dimensional map. The sample is classified as having the phenotype of the cluster it falls within, or to which it is closest.

Overview | Applications / Classification | How it Works | Supporting Software | Publications

Back to top