Kati Iltanen

Evaluation and classification of otoneurological data with new data analysis methods based on machine learning

We improved the classification ability of multilayer perceptron networks by constructing a set of networks of as many as output classes and investigated the influence of different input variables on the classification. We have developed methods named scattering, spectrum and response analysis to express the classification complexity, especially the overlap of output classes, to disentangle the relation between the input variables and output classes of perceptron neural networks, and to establish the importance of input variables. The methods were tested by exploring complicated otoneurological data. In contrast to the variable selection problem, our methods characterize the importance of variables for classification and also describe the importance of the different values of each variable for output (disease) classes. When complex data is distributed in a biased manner between disease classes, we improved classification accuracy by developing a network set called NetSet, which increased average sensitivity and positive predictive value for at least 10% up to 85% and 83% respectively, compared to our earlier neural network classifications with the same data, which clarified class distribution effects and supported our comprehension of the significance of input.

Genetic Algorithm Based Approach in Attribute Weighting for a Medical Data Set

Copyright

On computation of calcium cycling anomalies in cardiomyocytes data.

Induced pluripotent stem cell (iPSC) lines derived from skin fibroblasts of patients suffering from cardiac disorders were differentiated to cardiomyocytes and used to generate a data set of Ca2+ transients of 136 recordings. The objective was to separate normal signals for later medical research from abnormal signals. We constructed a signal analysis procedure to detect peaks representing calcium cycling in signals and another procedure to classify them into either normal or abnormal peaks. Using machine learning methods we classified signals into normal or abnormal signals on the basis of peak findings in them. We compared classification results obtained to those made visually by an expert biotechnologist who assessed the signals independent of the computer method. Classification accuracies of around 85% indicated high congruence between two modes denoting the high capability and usefulness of computer based processing for the present data.

Classification of iPSC colony images using hierarchical strategies with support vector machines

In this preliminary research we examine the suitability of hierarchical strategies of multi-class support vector machines for classification of induced pluripotent stem cell (iPSC) colony images. The iPSC technology gives incredible possibilities for safe and patient specific drug therapy without any ethical problems. However, growing of iPSCs is a sensitive process and abnormalities may occur during the growing process. These abnormalities need to be recognized and the problem returns to image classification. We have a collection of 80 iPSC colony images where each one of the images is prelabeled by an expert to class bad, good or semigood. We use intensity histograms as features for classification and we evaluate histograms from the whole image and the colony area only having two datasets. We perform two feature reduction procedures for both datasets. In classification we examine how different hierarchical constructions effect the classification. We perform thorough evaluation and the best accuracy was around 54% obtained with the linear kernel function. Between different hierarchical structures, in many cases there are no significant changes in results. As a result, intensity histograms are a good baseline for the classification of iPSC colony images but more sophisticated feature extraction and reduction methods together with other classification methods need to be researched in future.

Signal analysis and classification methods for the calcium transient data of stem cell-derived cardiomyocytes

Calcium cycling is crucial in the excitation-contraction coupling of cardiomyocytes, and therefore has a key role in cardiac functionality. Cardiac disorders and different drugs alter the calcium transients of cardiomyocytes and can cause serious dysfunction of the heart. New insights into this biochemical phenomena can be achieved by studying and analyzing calcium transients. Calcium transients of spontaneously beating human induced pluripotent stem cell-derived cardiomyocytes were recorded for a data set of 280 signals. Our objective was to develop and program procedures: (1) to automatically detect cycling peaks from signals and to classify the peaks of signals as either normal or abnormal, and (2) on the basis of the preceding peak detection results, to classify the entire signals into either a normal class or an abnormal class. We obtained a classification accuracy of approximately 80% compared to class decisions made separately by an experienced researcher, which is promising for the further development of an automatic classification approach. Automated classification software would be beneficial in the future for analyzing cardiomyocyte functionality on a large scale when screening for the adverse cardiac effects of new potential compounds, and also in future clinical applications.

Machine learning approach to an otoneurological classification problem

In this paper we applied altogether 13 classification methods to otoneurological disease classification. The main point was to use Half-Against-Half (HAH) architecture in classification. HAH structure was used with Support Vector Machines (SVMs), k-Nearest Neighbour (k-NN) method and Naïve Bayes (NB) methods. Furthermore, Multinomial Logistic Regression (MNLR) was tested for the dataset. HAH-SVM with the linear kernel achieved clearly the best accuracy being 76.9% which was a good result with the dataset tested. From the other classification methods HAH-k-NN with cityblock metric, HAH-NB and MNLR methods achieved above 60% accuracy. Around 77% accuracy is a good result compared to previous researches with the same dataset.

Attribute weighting with Scatter and instance-based learning methods evaluated with otoneurological data

Treating all attributes as equally important during classification can have a negative effect on the classification results. An attribute weighting is needed to grade the relevancy and usefulness of the attributes. Machine learning methods were utilised in weighting the attributes. The machine learnt weighting schemes, weights defined by the application area experts and the weights set to 1 were tested on otoneurological data with the nearest pattern method of the decision support system ONE and the attribute weighted k-nearest neighbour method using one-vs-all (OVA) classifiers. The effects of attribute weighting on the classification performance were examined. The results showed that the extent of the effect the attribute weights had on the classification results depended on the classification method used. The weights computed with the Scatter method improved the total classification accuracy compared with the weights 1 and the expert-defined weights with ONE and the attribute weighted 5-nearest neighbour OVA methods.