Self-organizing Maps Using Acoustic Features for Prediction of State Change in Bipolar Disorder

Acoustic Feature Selection with Fuzzy Clustering, Self Organizing Maps and Psychiatric Assessments

“…Due to all these uncertainties, in this research we incorporate fuzzy clustering for alternative subsets of acoustic features. This paper is a continuation of our previous works [4]. However, now we aggregate data to a single phone call, whereas in citech26ref4 the aggregation has been done to one day.…”

“…The current state of the art lacks a clear indication which of the acoustic features are the best predictors of BD phase. In [4], the authors use the following 12 parameters: spectral slope in the ranges 0-500 Hz and 500-1500 Hz, energies in bands 0-650 Hz and 1000-4000 Hz, alpha ratio, ratio of the energy in band 50-100 Hz to the energy in band 1000-5000 Hz, spectral roll-off point the frequency below which 25% of the spectrum energy is concentrated, harmonicity of the spectrum, maximal position of the FFT spectrum, Hammarberg index, entropy of the spectrum, modulated loudness (RASTA), and zero-crossing rate. Different subset of parameters was selected with filter feature selection by [3], who use the following parameters: kurtosis energy, mean second and mean third MFCC, mean fourth delta MFCC, max ZCR and mean HNR, std and range F0.…”

“…Important feature of the self-organizing maps is the preservation of neighborhood between the clusters in the two-dimensional space. Similarly, as in [4], we apply rectangular map topology of dimensions 3 × 1, which is intended to identify the two most distant affective states mania and depression and euthymia in between. The fourth affective state, mixed, is in its nature a combination of both, manic and depressive symptoms, and as confirmed with our preliminary experiments, it is more adequately represented as a mixture of depression and mania within the 3 × 1 Kohonen network, than as an additional dimension of a map, e.g., 4 × 1.…”

“…In majority of the related work, see e., [2,3], the problem is stated as a supervised learning task. Recently, to alleviate the problems of uncertainty about patients state and limited data, Kamińska et al [4] applied unsupervised learning technique (self-organizing maps) to find groups (clusters) in acoustic data for each patient without taking into account the psychiatric assessments. As a consequence, the whole dataset was used for learning rather than constraining it only to a few days before and after the visit to the psychiatrist.…”

“…Then, the relation between the learned clusters and the labels from psychiatric assessment was investigated. Kamińska et al [4] have noted that the degree of agreement between the results of unsupervised learning (clusters learned on acoustic data) and the results of the psychiatric assessments is related to the type of the BD phases recognized during the psychiatric evaluation.…”