On classification capability of neural networks: a case study with otoneurological data.

We investigated the capability of multilayer perceptron neural networks and Kohonen neural networks to recognize difficult otoneurological diseases from each other. We found that they are efficient methods, but the distribution of a learning set should be rather uniform. Also it is important that the number of learning cases is sufficient. If the two mentioned conditions are satisfied, these neural networks are similarly efficient as some other machine learning methods. The conditions are known in the theory of neural networks [1,2], but not often taken seriously in practice. Both networks functioned as well, excluding the case with several input variables, where the Kohonen neural networks surpassed the perceptron.

Classification of patients on the basis of otoneurological data by using Kohonen networks.

Machine learning methods such as neural networks, decision trees and genetic algorithms can be useful to aid in the classification of patients. We tested Kohonen artificial neural networks, which are known to be effective for classification tasks. Our sample included patients with six different diseases. The Kohonen network algorithm recognized the four largest groups reliably, but the two smallest groups were too small for the method. Neural networks seem to be promising for the computer-aided classification of otoneurological patients provided that the number of patients used is sufficiently large.

Evaluating training data suitability for decision tree induction.

Decision tree induction, as well as other inductive learning methods, requires training data of high quality to be able to generate accurate and reliable classification models. Example cases should form a representative sample from the application area, and the attributes used to describe example cases should be relevant and adequate for the classification task to be solved. In this paper, measures of the strength of association and an entropy-based approach have been used to assess the quality of the training data. Studied classification tasks related to three otological data sets: a conscript data set, a vertigo data set, and a postoperative nausea and vomiting data set. The paper suggests that the studied approaches give some guidelines about the quality of the training data, but other approaches are also needed to guide training data building.

Using decision tree induction to model oculomotor data.

Decision tree induction is a machine learning method used to generate classification models from data sets. Numerous decision trees were constructed to examine relationships between oculomotor test parameters and lesion sites in a data set containing cases with operated cerebello-pontine angle tumour, operated hemangioblastoma, infarction of cerebello-brainstem and Ménière's disease, and control subjects. The aim was to find useful parameter combinations with discriminatory power. Decision trees constructed using both pursuit eye movements and saccadic eye movements yielded the best classification results. This is reasonable: oculomotor test results vary according to the site of the lesion and so the performance ability of subjects has to be taken into account in the classification. The decision tree program was able to generate classification models from the oculomotor data set. Generated decision trees were intelligible and can be utilized in physicians' research work.

Experiences of otoneurological expert system for vertigo.

We have developed an OtoNeurological Expert system (ONE) to aid the diagnostics of vertigo, to assist teaching and to implement the database for research. The database contains detailed information on the patient history, signs and test results necessary for the diagnostic work with vertiginous patients. The pattern recognition method was used in the reasoning process. Questions regarding symptoms, signs and test results are weighted and scored for each disease, and the most likely disease is recognized from the defined disease profiles. Uncertainties in reasoning, caused by missing information, were solved with a method resembling fuzzy logic. We have also applied adaptive computer applications, such as genetic algorithms and decision trees, in the reasoning process. In the validation the expert system ONE proved to be a sound decision maker, by solving 65% of the cases correctly, while the physicians' mean was 69%. To improve the expert system ONE further, a follow-up should be implemented for the patients, to ease the diagnostic work of some difficult diseases. The six diseases were detected with high accuracy also with adaptive learning methods and discriminant analysis. An expert system is a practical tool in otoneurology. We aim to construct a hybrid program for the reasoning, where the best reasoning method for each disease is used.

Application of artificial intelligence in audiology.

In this paper, machine learning methods based on artificial intelligence theory are applied to the computer-aided decision making of some otoneurological diseases, for example Ménière's disease. Three methods explored are decision trees, genetic algorithms and neural networks. By using such a machine learning method, the decision-making program is trained with a representative training set of cases and tested with another set. The machine learning methods are useful also for our otoneurological expert system, One, which is based on a pattern recognition approach. The methods are able to differentiate most of the cases tested between the six diseases included, provided that a sufficiently large training set is available.

Production of diagnostic rules from a neurotologic database with decision trees.

A decision tree is an artificial intelligence program that is adaptive and is closely related to a neural network, but can handle missing or nondecisive data in decision-making. Data on patients with Meniere's disease, vestibular schwannoma, traumatic vertigo, sudden deafness, benign paroxysmal positional vertigo, and vestibular neuritis were retrieved from the database of the otoneurologic expert system ONE for the development and testing of the accuracy of decision trees in the diagnostic workup. Decision trees were constructed separately for each disease. The accuracies of the best decision trees were 94%, 95%, 99%, 99%, 100%, and 100% for the respective diseases. The most important questions concerned the presence of vertigo, hearing loss, and tinnitus; duration of vertigo; frequency of vertigo attacks; severity of rotational vertigo; onset and type of hearing loss; and occurrence of head injury in relation to the timing of onset of vertigo. Meniere's disease was the most difficult to classify correctly. The validity and structure of the decision trees are easily comprehended and can be used outside the expert system.

Comparison of genetic algorithms and other classification methods in the diagnosis of female urinary incontinence.

Galactica, a newly developed machine-learning system that utilizes a genetic algorithm for learning, was compared with discriminant analysis, logistic regression, k-means cluster analysis, a C4.5 decision-tree generator and a random bit climber hill-climbing algorithm. The methods were evaluated in the diagnosis of female urinary incontinence in terms of prediction accuracy of classifiers, on the basis of patient data. The best methods were discriminant analysis, logistic regression, C4.5 and Galactica. Practically no statistically significant differences existed between the prediction accuracy of these classification methods. We consider that machine-learning systems C4.5 and Galactica are preferable for automatic construction of medical decision aids, because they can cope with missing data values directly and can present a classifier in a comprehensible form. Galactica performed nearly as well as C4.5. The results are in agreement with the results of earlier research, indicating that genetic algorithms are a competitive method for constructing classifiers from medical data.

Modelling oculomotor data with decision tree induction.

We studied relationships between oculomotor test results and the site of lesion in the data set containing patient cases with operated cerebello-pontine angle tumour, operated hemangioblastoma, infarction of cerebello-brainstem or Meniere's disease and control subjects. Several classification models were generated by decision tree induction to find parameter combinations which are efficient in classification. The studied data were characterised by missing values and scarcity of example cases. This study suggested that decision tree induction can be a useful method even in problematic real world classification tasks. Models generated by decision tree induction and evaluated by an expert physician may give information that is beneficial for research.

Decision tree induction in the diagnosis of otoneurological diseases.

Expert systems have been applied in medicine as diagnostic aids and education tools. The construction of a knowledge base for an expert system may be a difficult task; to automate this task several machine learning methods have been developed. These methods can be also used in the refinement of knowledge bases for removing inconsistencies and redundancies, and for simplifying decision rules. In this study, decision tree induction was employed to acquire diagnostic knowledge for otoneurological diseases and to extract relevant parameters from the database of an otoneurological expert system ONE. The records of patients with benign positional vertigo, Meniere's disease, sudden deafness, traumatic vertigo, vestibular neuritis and vestibular schwannoma were retrieved from the database of ONE, and for each disease, decision trees were constructed. The study shows that decision tree induction is a useful technique for acquiring diagnostic knowledge for otoneurological diseases and for extracting relevant parameters from a large set of parameters.