|Abstract (english)|| |
This doctoral thesis presents the implemented methods and integrated software systems for the classification of asthma and chronic obstructive pulmonary disease (COPD). The implemented methods are based off of GINA and GOLD recommendations, as well as recommendations from other experts in the field of diagnosis of lung diseases. Within the methods applied, the process begins with the static assessment of the conditions of the subjects’ respiratory systems based on the interpretation of the results from the measurement of lung function by means of impulse oscillometry (IO) or spirometry. From the results of the static assessment of the respiratory patients and also the addition of their symptoms allows for the initial diagnosis. Defining symptoms and also classification of disease based on symptoms is implemented according to the GINA and GOLD recommendations. However, for the given solution to classify all the cases of COPD and asthma, it is necessary to implement the remaining steps that are used for clinical purposes to give a final diagnosis of lung disease. The applied method is also used for determining the dynamic assessment of the respiratory system of the respondents, which includes the use of bronchodilating test (BDT) and / or bronchial provocation test (BPT). Subsequently, on the basis of the static and dynamic assessment of respiratory patients and their symptoms, it becomes possible to classify the conditions of all cases of asthma and COPD. For the methods of pulmonary diagnostics to be used, it was necessary to develop an integrated software system that is easy for users. The system consists of programs for classification based on symptoms and programs for classification based on static and dynamic assessment of the subjects’ respiratory systems conditions. All these programs can work separately, and even doctors can offer their opinions after each of the steps have been performed. Ultimately, an integrated software system allows for the final classification based on the classification results of each of the separate sections of the program. The classification of diseases based on a patient’s symptoms is completed using Visual Studio and its graphic interface, and it is presented to the user in the same fashion also used in health care facilities during anamnesis. The results of the classification based on the symptoms are very important, and often prevailing in cases where it is difficult to assess whether it is asthma or COPD. The statistical assessment of a subjects’ respiratory systems conditions based on their results from impulse oscillometry measurements (IO) and spirometry were produced using a fuzzy logic system of decision-making, designed by the GINA and GOLD recommendations, as well as information from prior research. On the basis of this data as the only logical decision pertained, it was proposed that in most cases we need to use ∧ and Γ membership function, which is used when creating a fuzzy logic system of decision-making. The rules of fuzzy decision-making are defined by the value of the input data obtained by spirometry and / or impulse oscillometry. These values are stored in PDF documents. On the basis of these data systems, fuzzy decision-making has the opportunity to make a preliminary classification of the disease, if it is about a simple case. Otherwise, outputs of the fuzzy decision-making system represent input vectors of artificial neural networks. Using recommendations for the diagnosis of asthma and COPD, and using the implemented logic of the fuzzy system of decision-making, outputs are also defined on the basis of which yields a classification. In the case of input data IO as one of the outputs from the logic of the fuzzy decision-making system, NSCO-IV, NSCO-RV, NSPO, CO-IV, CO RV or PO can be obtained. If we use spirometric input data as one of the outputs, we can obtain normal, light, moderate or heavy results. In cases where it is not possible to classify only on the basis of the fuzzy logic system of decision-making, and only on the basis of static assessment of respiratory patients, it is necessary to use an artificial neural network. In this case, the artificial neural network "performs the role of a doctor", and on the basis of their knowledge or the degree experience, with greater or lesser success in classifying the incoming data, and provides the correct classification of the disease. When selecting the architecture of the neural network used, the recommendations of experts in this area were taken into account, where it was stated that acyclic and recurrent networks are most commonly used. As the recurrent network is used exclusively for data classification that basically represents dynamic systems, it was not necessary to train and test recurrent networks as it would be the same in the final elimination of feedback loops, which would be reduced to an acyclic network. The network is further divided into layers. The input layer is composed of the input network, which consists of the results of the classification of the fuzzy logic of decision-making system. Furthermore, it follows a hidden layer consisting of neurons or hidden units connected in a parallel fashion. For the purpose of examining the efficiency of this network, the effect of the number of neurons in the hidden layer is tested on the success of the classification of the disease. Thus in the hidden layer was positioned 1, 2, 5, 10 and 20 neurons, to help find the results of the classification. On this basis, it was concluded that the best results were achieved using 20 neurons in the hidden layer. With regards to performance and training adaptations that are extremely dependent on the complexity of the network and the number of neurons in the hidden layer, it was taken into account to avoid the so-called curse of dimensionality. This method ultimately selected 10 neurons in the hidden layer to implement neural networks. Each selected neuron accumulates the weighted sum of the input, which is then passed to the nonlinear activation function σ, also known as neuronal function. In the case of solutions presented in this paper, it uses the tansig activation function for the hidden layer, which is equivalent to the hyperbolic tangent function. This function is used in highly nonlinear classification data, as is the case in the evidence given by the system shown in this study. The output of the network is formed by the following weighted sum of the output neurons in the hidden layer. The sum of the output is called the output layer. For the output layer, in the case presented in this study, the linear activation function was selected since it is commonly used for regression problems. For the purposes of training for the artificial neural network, Levenberg - Marquardt algorithm was used. Training and validation were conducted using estimation and validation data. The set for training consisted of 1000 findings with previous diagnosis obtained from the company CareFusion. To determine the best ratio of the necessary data for Estimation and validation, the set examined different combinations of data. We examined the rate of correct diagnosis in the case of estimation data set, i.e. the safety of the LMA, and the percentage of correct diagnosis when using the validation data set. Based on the two obtained results, the average percentage of correct diagnosis was determined, which will also assess which combination will gain the best results. On the basis of the percentage of correct diagnoses in the estimation set through all the tested combination, it can be concluded that the LMA provides the best results. To the greatest percentage of the correct diagnoses in the validation set came by using an estimation data set of 800 and validation data set sizes of 200. According to results of testing, it was concluded that the most successful combination of training artificial neural networks was using estimation data sets of 800 and validation data set size of 200. Ultimately, training artificial neural networks and validation of the implemented system was implemented in this selected combination. After that, the selected structure of the neural network is trained with all 1000 findings. The validation of the train system was done with 455 patients within the Lung Department of the Clinical Center University of Sarajevo over the period of 6 months. Doctors worked in accordance to its procedures and instructions for the use of the developed system for examining patients and making a diagnosis. Doctors at the stage of diagnosing patients with asthma and COPD, first entered the symptoms, and then check the patient’s heart using a statoscope. Then they approached the measurement of lung function using spirometry tests and IO-E. In some typical cases of respiratory diseases, it was necessary, and the method of measuring was body plethysmography. Each of these methods of testing lung function gives the characteristic parameters necessary for the diagnosis of specific diseases. However, there are some correlations between parameters obtained by different methods of lung function tests. For the purpose of this doctoral thesis, the impact of input parameters IO's and spirometry in relation to the outputs from the logic of the system of fuzzy decision-making was tested, and so was the impact of the input parameters IO's and spirometry output of the artificial neural network. Based on these results, we determine the exact boundaries of the measured values of the input parameters that can assume some of the diagnosis using the developed fuzzy logic system of decision-making or in the end the system of artificial neural networks. All these studies on correlations between impulse oscillometry and spirometry, as well as recommendations GINA and GOLD on the diagnosis of asthma and COPD, served as motivation to integrate the software system based on a combination of measurements and spirometry IO’s. The integrated software system developed as part of the doctoral thesis in contrast to previous studies is able to classify the two diseases, asthma and COPD, as well as to determine whether the respondent is healthy. After using only the results of standard tests, spirometry and IO-E, the system has performed the correct classification in 87.65% asthmatics and 85.50% COPD patients. Once the recommendations of system access bronchodilator and / or bronchial provocation test, correct classification is achieved in 99.41% asthmatics and 99.19% COPD patients. Considering the fact that the integrated software system consists of a system of decisions and an artificial neural network, all results, obtained just based on classification of systems of decisions, are confirmed through the artificial neural network. And for simple cases where the diagnosis can be determined immediately after the classification system of fuzzy decision-making, the result of the forward and artificial neural network acts as the role of a doctor. In this way we are able to get a complete picture of the dynamic state of respiratory patients in contrast to all other solutions that offer only static assessment respiratory patients. The developed integrated software system is fully adapted to the end user, ie. physician and other health care personnel. Entering symptoms was designed in the form of multiple choice in order to process the input and the classification based on symptoms making it faster and easier to do. Also, the result of classification of fuzzy decision-making, and static assessment of the respiratory system of the respondents, is provided as text, and pictorial. Also, the results of the classification of artificial neural networks are clearly and concisely defined, and to provide instructions what to do next, and that the final classification for each patient. Further research is planned to develop an integrated software system that will be able to diagnose and some other respiratory diseases.