Search

Fever is a common symptom of many diseases. Fever temporal patterns can be different depending on the specific pathology. Differentiation of diseases based on multiple mathematical features and visual observations has been recently studied in the scientific literature. However, the classification of diseases using a single mathematical feature has not been tried yet. The aim of the present study is to assess the feasibility of classifying diseases based on fever patterns using a single mathematical feature, specifically an entropy measure, Sample Entropy. This was an observational study. Analysis was carried out using 103 patients, 24 hour continuous tympanic temperature data. Sample Entropy feature was extracted from temperature data of patients. Grouping of diseases (infectious, tuberculosis, non-tuberculosis, and dengue fever) was made based on physicians diagnosis and laboratory findings. The quantitative results confirm the feasibility of the approach proposed, with an overall classification accuracy close to 70%, and the capability of finding significant differences for all the classes studied. %An abstract is a brief of the paper; the abstract should not contain references, the text of the abstract section should be in 12 point normal Times New Roman.

Despite its widely demonstrated usefulness, there is still room for improvement in the basic Permutation Entropy (PE) algorithm, as several subsequent studies have proposed in the recent years. For example, some improved PE variants try to address possible PE weaknesses, such as its only focus on ordinal information, and not on amplitude, or the possible detrimental impact of equal values in subsequences due to motif ambiguity. Other evolved PE methods try to reduce the influence of input parameters. A good representative of this last point is the Bubble Entropy (BE) method. BE is based on sorting relations instead of ordinal patterns, and its promising capabilities have not been extensively assessed yet. The objective of the present study was to comparatively assess the classification performance of this new method, and study and exploit the possible synergies between PE and BE. The claimed superior performance of BE over PE was first evaluated by conducting a series of time series classification tests over a varied and diverse experimental set. The results of this assessment apparently suggested that there is a complementary relationship between PE and BE, instead of a superior/inferior relationship. A second set of experiments using PE and BE simultaneously as the input features of a clustering algorithm, demonstrated that with a proper algorithm configuration, classification accuracy and robustness can benefit from both measures.