Published on
In this case study, we evaluate the implementation of Sequential Possibilistic Gaussian Mixture Models (SPGMM) for accurately modeling changes in feature streams antecedent to known health events, thereby providing predictive relevance for clinical use, including identifying the preprocessing requirements for streams prior to algorithm input.
SPGMM is a change detection algorithm developed for use in online data stream processing applications where feature vectors are introduced sequentially as inputs for iterative clustering. SPGMM is comprised of two components: the Sequential Possibilistic One-Means (SP1M) algorithm for generating the initial clusters (Gaussian mixtures), combined with the Gaussian Mixture Model (GMM), which defines how the clusters (mixture of probability densities), are approximately oriented within feature space.
By measuring feature stream changes over time and identifying when those changes are meaningful, such as when the underlying trajectory has significantly changed in direction or surpassed a distance threshold to existing clusters introducing a new cluster, SPGMM may be a more accurate method of modeling health state change; Particularly when detecting changes that occur as low amplitude events across multiple features, which go undetected by our current method of detecting points falling +/- 3 standard deviations within a windowed univariate distribution for each feature.