Edzard S. Gelsema

Recognition of daily life motor activity classes using an artificial neural network

OBJECTIVE To investigate a possible role of artificial neural networks for the automated recognition and classification of daily life activities (eg, sitting, lying, standing, walking, etc) in an attempt to reduce the cost of manual recognition and classification. METHODS Data from sessions of about 10 hours of continuous recording of eight ambulatory patients were used to train and evaluate eight probabilistic neural networks, each of which is configured for one subject. To provide the reference data for building the training set, the instrumented subject follows a 15- to 30-minute protocol consisting of several daily life activities. To properly evaluate the networks, the remaining manually labeled data of each subject were compared with the output of each trained network. RESULTS The average recognition rate of the trained neural networks was equal to 95% good classification of all presented cases of the daily life activity. Automatic misclassification of 5% resulted from certain activities being too short or the occurrence of activities that were not included in the training set. CONCLUSION The preliminary results of the trained neural networks have indicated that the probabilistic neural network is a potentially useful tool for the recognition of daily life motor activities.

Abductive reasoning in Bayesian belief networks using a genetic algorithm

A set of computational experiments is described in which genetic algorithms are used for abductive reasoning in Bayesian belief networks. It is shown that good solutions and explanations are consistently found with high probabilities. The efficiency of genetic sampling w.r.t. random sampling is shown to increase with increasing complexity of the search space and with increasing complexity of the search goal.

Quantitative electron spectroscopic imaging in bio-medicine: Evaluation and application

Electron spectroscopic imaging (ESI) with the energy‐filtering transmission electron microscope enables the investigation of chemical elements in ultrathin biological sections. An analysis technique has been developed to calculate elemental maps and quantitative distributions from ESI sequences. Extensive experience has been obtained with a practical implementation of this technique. A procedure for more robust element detection has been investigated and optimized. With the use of Fe‐loaded Chelex beads, the measurement system has been evaluated with respect to the linearity of the element concentration scale, the reproducibility of the measurements and the visual usage of image results. In liver specimens of a patient with an iron storage disease the detectability of iron was tested and we tried to characterize iron‐containing components.  The concentration measurement scale is approximately linear up to a relative section thickness of ≈ 0.5. Monitoring of this parameter is therefore considered to be important. The reproducibility was measured in an experiment with Fe‐Chelex. The iron concentration differed by 6.4% between two serial measurements. Element distributions are in many applications interpreted visually. For this purpose the frequently used net‐intensity distributions are regarded as unsuitable. For the quantification and visual interpretation of concentration differences mass thickness correction has to be performed. By contrast, for the detection of elements the signal‐to‐noise ratio is the appropriate criterion.  Application of ESI analysis demonstrated the quantitative chemical capabilities of this technique in the investigation of iron storage diseases. Based on an assumed ferritin iron loading in vivo, different iron components can be discerned in liver parenchymal cells of an iron‐overloaded patient.

Quantitative electron spectroscopic imaging in bio‐medicine: Methods for image acquisition, correction and analysis

Many questions about the metabolism of specific elements in the human body might be answered if elemental concentrations could be measured in situ in cells. With electron energy‐loss spectroscopic imaging (ESI), concentrations can potentially be determined with high spatial resolution. The theory of the quantification procedure has already been derived. Many practical instrument‐related problems, however, have to be solved. In the current research an energy‐filtering TEM is used and the image‐acquisition chain is examined in detail. Quantification requires images to be recorded over a large dynamic range. To solve this problem, the use of optical attenuation filters has been introduced. The use of the combination of a scintillator screen and a TV‐camera as a detection system has consequences for the processing of the data. Corrections for the camera photometric sensitivity and, to some extent, for shading are necessary. Further consequences of such a detection system for the correction of the element a‐specific spectral background and element detection are discussed. The derived methodology is tested in several ways and finally applied for the quantitative analysis of iron in liver parenchymal cells of a porphyria cutanea tarda patient.

Quantitative analysis of electron energy-loss spectra from ultrathin-sectioned biological material: II. The application of Bio-standards for quantitative analysis

Electron energy‐loss spectroscopy (EELS) has been used to determine elemental concentrations in biological specimens, consisting of ultrathin‐sectioned cells and tissues. Chelex100‐based Ca‐ and Fe Bio‐standards are used for elemental quantification to establish iron and calcium concentrations. These Bio‐standards, as well as the biological materials, are treated in a standard EM procedure such that ‘known’ and ‘unknown’ sites are located in one cross‐section.

A computer program for constructing multivariate reference models

From a statistical point of view the simultaneous interpretation of multiple variables should be performed with a multivariate reference model rather than with multiple univariate reference intervals. A computer program for constructing and testing multivariate reference models is described. The use of the computer program is illustrated with a data set of total serum calcium concentrations and serum albumin concentrations from 222 2nd year medical students. Using a single univariate reference interval for total serum calcium, 17 students were classified as having an abnormal calcemic status while using a bivariate reference model for total serum calcium and serum albumin, 13 of these 17 students had in fact normal total serum calcium concentrations, taking into account their serum albumin concentrations.

Design and representation of multivariate patient-based reference regions for arterial pH, Pco2 and base excess values☆

OBJECTIVES To determine and compare the shape and location of three data sets of arterial pH, PCO2, and BE values from intensive care patients in a new acid-base chart for the purpose of deriving multivariate reference regions. DESIGN AND METHODS The new chart is constructed by applying a statistical technique called principal component analysis (PCA). Three different data sets, each comprised of 1500 arterial pH, PCO2, and BE values, were subjected to PCA. The 3 data sets were collected in a respiratory intensive care unit (ICU) of a University Hospital, in a general ICU of a District Hospital, and in a neonatal ICU of a Childrens Hospital. RESULTS The outlines of the resulting charts are similar for all 3 data sets, but the representations of the three distributions in the new chart are highly dissimilar, both in shape and in location. CONCLUSIONS PCA can be used to derive a patient-based reference region for arterial pH, PCO2, and BE values. Furthermore, the new chart may be useful for the graphical monitoring of acid-base data because distances between consecutive observations are faithfully represented.

An efficient method for calculating the least-squares background fit in electron energy-loss spectroscopy

In quantitative electron energy‐loss spectroscopy, an estimate has to be made of the unspecific spectral background to obtain the element‐specific contribution of an ionization edge. For energy losses above c. 100 eV, such a background estimate is commonly performed by calculating the least–squares fit of a power‐law model to a part of the pre‐edge region of the spectrum and by extrapolating this model to the region beyond the edge. Existing methods are cumbersome in terms of computation time and stability, or yield biased estimates. Computation efficiency becomes important when processing spectral‐image sequences. In this paper a new, computationally efficient numerical method is presented, which appears to be extremely robust in practice.

Quantification procedures for electron energy-loss spectroscopy and imaging : the use of correspondence analysis for element determination

A systematic investigation of the application of correspondence analysis to the determination of element responses in electron spectroscopic imaging is presented. The formalism of correspondence analysis has been applied to a set of artificially generated images, in which the effects of non‐homogeneous specimen thickness and noise were incorporated. The simulation experiments indicate that the results may be used for both qualitative (element distribution) and quantitative (element concentration) purposes.

New developments and applications in quantitative electron spectroscopic imaging of iron in human liver biopsies

Reliable iron concentration data can be obtained by quantitative analyses of image sequences, acquired by electron spectroscopic imaging. A number of requirements are formulated for the successful application of this recently developed in situ quantitative type of analysis. A demonstration of the procedures is given. By application of the technique it is established that there are no significant differences in the average iron loading of structures analysed in liver parenchymal cells of a patient with an iron storage disease, before and after phlebotomy. This supports the hypothesis that the process of iron unloading is an organelle specific process. Measurement of the binary morphology, represented by the area and contour ratio of the iron containing objects revealed no information about differences between the objects. This finding contradicts the visual suggestion that ferritin clusters are more irregularly shaped than the other iron objects. Also, no differences could be found in this sense between the situations before and after phlebotomy. With respect to the density appearance, objects that have an inhomogeneous iron loading averagely contain more iron. This observation does correspond well with the visual impression of the increasingly irregular appearance of more well-loaded structures.