Assisting the Diagnosis of Thyroid Diseases with Bayesian-Type and SOM-Type Neural Networks Making Use of Routine Test Data

Abstract

Both classification problems and QSAR analyses of drugs have been better studied by the multi-layer feedforward neural networks with back-propagation learning than the classical methods in pattern recognition such as linear multiple regression or adaptive least square method, since those problems often involve nonlinear relationships between descriptors and the class (/activity). [1][2][3] Recently Bayesian regularized neural networks (BRNN), [4][5][6][7] which extends back-propagation learning algorithm in order to overcome its defects such as the problems of local trapping, overfitting etc. by introducing probabilistic treatment of the Bayesian inference technique for the synaptic weights, has been successfully applied for QSAR studies 8,9) even for massive sample data. 10)Analytical chemistry is often concerned with the classification of samples into groups on the basis of chemical results while such classification process is called diagnosis if these groups are pathological cases. In the previous work 11) we analyzed the human thyroid data as the three class classification problem, that is, diagnosis by using two notable approaches, the Bayesian regularized neural networks and the self-organizing map. [12][13][14] The former presented the high classification rates and a nice soft pruning if ARD method is swichted on. The SOM enabled nicely to grasp characteric features of thyroid function from the map of well clustered thyroid patients owing to data-visualization. These results ensured that both approaches are superior to statistical analysis based on conventional multivariate analysis techniques, and would be helpful in diagnosing thyroid function based on laboratory tests. Therefore, we now try to assist the diagnosis of thyroid diseases making use of the recent clinical examination data.Conventionally, multivariate analysis has been frequently used for analyzing data on medical care.15) This kind of analysis is implicitly based on the assumption that the sample shows a normal distribution. However, this requirement is often not satisfied because of small sample size or other reasons. The relationship of the routine test data, disease history and physical findings to the diagnosis (disease name) for individual patients seems to assume a non-linear form in many cases. When neural networks are used for analysis, there is no need to assume a normal distribution of the sample. This kind of analysis is applicable even when the sample size is large or small. Furthermore, analysis using neural networks is valid for non-linear relationships.In Japan, people who go to a hospital because they are complaining of some illness usually undergo routine testing as an early step in the diagnostic process. The data from these routine tests point to one or several possible diagnoses. Full-scale diagnosis is then started with the conduct of detailed tests focused on the diseases suspected on the basis of the disease history, findings from observations and data from the routine tests. However, sometimes the findings are negati...

Los cirujanos sobreestiman el riesgo de malignidad de los nódulos tiroideos, evaluación de los estimados subjetivos usando un análisis bayesiano

“…We only found 8 studies in this field, most of them using information from imaging tests or pathology reports. We considered laboratory information as a secondary step that should be applied after defining the risk of malignancy of the nodule (22)(23)(24)(25)(26)(27)(28)(29) If we consider individual variables, as shown in table 3, lower risk factors, as may be seen in middle-aged female patients, are assigned a malignancy probability of 19%, a patient without risk factors, such as radiotherapy or family history, is assigned a probability of 25% or a patient with multiple nodules that are smaller than 1 cm and have a soft consistency is assigned a probability of 20%, which are clearly higher than those reported in the literature, of finding carcinoma in an index nodule, which is approximately 5 to 15%. This was corroborated by the clinical cases, in which a patient without any risk factors is assigned a probability of malignancy of 19%, and in the final results that were obtained by the network, where a patient without risk factors has a basal probability of malignancy of 33%.…”

Thyroid dysfunction diagnosis from routine laboratory tests based on machine learning

“…We used 5 explanatory variables as the baseline set of features (referred to as Feature set 2): sex, AST, ALT, γ-GTP, and total cholesterol, of which four features were the required laboratory tests conducted in the Japanese national health screening program called Specific Health Checkups. In addition, hemoglobin (Hb), RBC, creatinine (S-Cr), ALP, and UA, which are the tests measured only at the doctor's discretion, are reported to be highly relevant to thyroid dysfunction 14,15 , these were added to the above items. We also included the UA/S-Cr ratio in this study considering that a reduction in S-Cr has been reported in hyperthyroidism, while UA has not been confirmed to fluctuate with thyroid dysfunction.…”

“…We used 11 variables as explanatory variables in this study as the first experimented set of features (referred to as Feature set 1) in this study, of which eight tests are tests measured in routine health checkup: sex, AST, ALT, γ-GTP, total cholesterol, Hemoglobin (Hb), RBC, and creatinine (S-Cr). In addition, since ALP, UA, and S-Cr ratio are reported to be highly relevant to thyroid dysfunction (24,25), these were added to the above items. We also included UA/S-Cr ratio in this study considering that the reduction of S-Cr has been reported in hyperthyroidism, while UA has not been confirmed to fluctuate with thyroid dysfunction.…”

“…In another study, the incidence of myocardial infarction or cerebral infarction was predicted using the results of a health checkup 10 . Numerous studies have also attempted to assess the efficacy of detecting misdiagnosed diseases, including thyroid dysfunction [11][12][13][14][15][16][17] . Aoki et al 16,17 found that there were strong, multiple correlations between a set of routine clinical parameters and FT4 in patients with both overt hyperthyroidism and overt hypothyroidism.…”

An In Silico Approach to Cluster CAM Kinase Protein Sequences

“…Self Organizing Map (also known as Kohonen Map) is a unsupervised learning algorithm (Kohonen, 2001) used for clustering and reducing dimensions of complex data with-J Proteomics Bioinform Volume 2(2) : 097-107 (2009) -098 ISSN:0974-276X JPB, an open access journal out loosing 'essence' of the data and is capable of organizing data based on the similarity by putting entities geometrically close to each other. SOMs have been applied in diverse fields like assessment of water quality (Walley et al, 2000), classification of communities (Chon et al, 1996;Arab et al, 2004;Tison et al, 2005), gene expression studies (Tamayo et al, 1999), disease diagnosis (Chen et al, 2000;Hoshi et al, 2006), medical imaging (Chuang et al, 2007), biochemical profiling (Kaartinen et al, 1998) and epidemiology (Murty and Arora, 2007). Self organizing maps have been earlier used in classification of families (Andrade et al, 1997), secondary structure determination(Unneberg et al, 2001) and pattern recognition in proteins (Hanke et al 1996).Owing to its use for multidimensional data visualization, SOM has aptly become the method of choice in bioinformatics studies (Hsu et al, 2003).…”