The problem of bias in training data in regression problems in medical decision support

Identifying and handling data bias within primary healthcare data using synthetic data generators

Advanced synthetic data generators can simulate data samples that closely resemble sensitive personal datasets while significantly reducing the risk of individual identification. The use of these advanced generators holds enormous potential in the medical field, as it allows for the simulation and sharing of sensitive patient data. This enables the development and rigorous validation of novel AI technologies for accurate diagnosis and efficient disease management. Despite the availability of massive ground truth datasets (such as UK-NHS databases that contain millions of patient records), the risk of biases being carried over to data generators still exists. These biases may arise from the under-representation of specific patient cohorts due to cultural sensitivities within certain communities or standardised data collection procedures. Machine learning models can exhibit bias in various forms, including the under-representation of certain groups in the data. This can lead to missing data and inaccurate correlations and distributions, which may also be reflected in synthetic data. Our paper aims to improve synthetic data generators by introducing probabilistic approaches to first detect difficult-to-predict data samples in ground truth data and then boost them when applying the generator. In addition, we explore strategies to generate synthetic data that can reduce bias and, at the same time, improve the performance of predictive models.

Advanced Sampling Technique in Radiology Free-Text Data for Efficiently Building Text Mining Models by Deep Learning in Vertebral Fracture

This study aims to establish advanced sampling methods in free-text data for efficiently building semantic text mining models using deep learning, such as identifying vertebral compression fracture (VCF) in radiology reports. We enrolled a total of 27,401 radiology free-text reports of X-ray examinations of the spine. The predictive effects were compared between text mining models built using supervised long short-term memory networks, independently derived by four sampling methods: vector sum minimization, vector sum maximization, stratified, and simple random sampling, using four fixed percentages. The drawn samples were applied to the training set, and the remaining samples were used to validate each group using different sampling methods and ratios. The predictive accuracy was measured using the area under the receiver operating characteristics (AUROC) to identify VCF. At the sampling ratios of 1/10, 1/20, 1/30, and 1/40, the highest AUROC was revealed in the sampling methods of vector sum minimization as confidence intervals of 0.981 (95%CIs: 0.980-0.983)/0.963 (95%CIs: 0.961-0.965)/0.907 (95%CIs: 0.904-0.911)/0.895 (95%CIs: 0.891-0.899), respectively. The lowest AUROC was demonstrated in the vector sum maximization. This study proposes an advanced sampling method, vector sum minimization, in free-text data that can be efficiently applied to build the text mining models by smartly drawing a small amount of critical representative samples.

Recent Advances in Bioimage Analysis Methods for Detecting Skeletal Deformities in Biomedical and Aquaculture Fish Species

Detecting skeletal or bone-related deformities in model and aquaculture fish is vital for numerous biomedical studies. In biomedical research, model fish with bone-related disorders are potential indicators of various chemically induced toxins in their environment or poor dietary conditions. In aquaculture, skeletal deformities are affecting fish health, and economic losses are incurred by fish farmers. This survey paper focuses on showcasing the cutting-edge image analysis tools and techniques based on artificial intelligence that are currently applied in the analysis of bone-related deformities in aquaculture and model fish. These methods and tools play a significant role in improving research by automating various aspects of the analysis. This paper also sheds light on some of the hurdles faced when dealing with high-content bioimages and explores potential solutions to overcome these challenges.

Evaluating the performance of machine-learning regression models for pharmacokinetic drug-drug interactions

Combination therapy or concomitant drug administration can be associated with pharmacokinetic drug-drug interactions, increasing the risk of adverse drug events and reduced drug efficacy. Thus far, machine-learning models have been developed that can classify drug-drug interactions. However, to enable quantification of the pharmacokinetic effects of a drug-drug interaction, regression-based machine learning should be explored. Therefore, this study investigated the use of regression-based machine learning to predict changes in drug exposure caused by pharmacokinetic drug-drug interactions. Fold changes in exposure relative to substrate drug monotherapy were collected from 120 clinical drug-drug interaction studies extracted from the Washington Drug Interaction Database and SimCYP compound library files. Drug characteristics (features) were collected such as structure, physicochemical properties, in vitro pharmacokinetic properties, cytochrome P450 metabolic activity, and population characteristics. Three different regression-based supervised machine-learning models were then applied to the prediction task: random forest, elastic net, and support vector regressor. Model performance was evaluated using fivefold cross-validation. Strongest performance was observed with support vector regression, with 78% of predictions within twofold of the observed exposure changes. The results show that changes in drug exposure can be predicted with reasonable accuracy using regression-based machine-learning models trained on data available early in drug discovery. This has potential applications in enabling earlier drug-drug interaction risk assessment for new drug candidates.

Inter-electrode correlations measured with EEG predict individual differences in cognitive ability

Human brains share a broadly similar functional organization with consequential individual variation. This duality in brain function has primarily been observed when using techniques that consider the spatial organization of the brain, such as MRI. Here, we ask whether these common and unique signals of cognition are also present in temporally sensitive but spatially insensitive neural signals. To address this question, we compiled electroencephalogram (EEG) data from individuals of both sexes while they performed multiple working memory tasks at two different data-collection sites (n = 171 and 165). Results revealed that trial-averaged EEG activity exhibited inter-electrode correlations that were stable within individuals and unique across individuals. Furthermore, models based on these inter-electrode correlations generalized across datasets to predict participants' working memory capacity and general fluid intelligence. Thus, inter-electrode correlation patterns measured with EEG provide a signature of working memory and fluid intelligence in humans and a new framework for characterizing individual differences in cognitive abilities.

For a critical appraisal of artificial intelligence in healthcare: The problem of bias in mHealth

RATIONALE, AIMS AND OBJECTIVES

Artificial intelligence and big data are more and more used in medicine, either in prevention, diagnosis or treatment, and are clearly modifying the way medicine is thought and practiced. Some authors argue that the use of artificial intelligence techniques to analyze big data would even constitute a scientific revolution, in medicine as much as in other scientific disciplines. Moreover, artificial intelligence techniques, coupled with mobile health technologies, could furnish a personalized medicine, adapted to the individuality of each patient. In this paper we argue that this conception is largely a myth: what health professionals and patients need is not more data, but data that are critically appraised, especially to avoid bias.

METHODS

In this historical and conceptual article, we focus on two main problems: first, the data and the problem of its validity; second, the inference drawn from the data by AI, and the establishment of correlations through the use of algorithms. We use examples from the contemporary use of mobile health (mHealth), i.e. the practice of medicine and public health supported by mobile or wearable devices such as mobile phones or smart watches.

RESULTS

We show that the validity of the data and of the inferences drawn from these mHealth data are likely to be biased. As biases are insensitive to the size of the sample, even if the sample is the whole population, artificial intelligence and big data cannot avoid biases and even tend to increase them.

CONCLUSIONS

The large amount of data thus appears rather as a problem than a solution. What contemporary medicine needs is not more data or more algorithms, but a critical appraisal of the data and of the analysis of the data. Considering the history of epidemiology, we propose three research priorities concerning the use of artificial intelligence and big data in medicine.

AMR-Diag: Neural network based genotype-to-phenotype prediction of resistance towards β-lactams in Escherichia coli and Klebsiella pneumoniae

Antibiotic resistance poses a major threat to public health. More effective ways of the antibiotic prescription are needed to delay the spread of antibiotic resistance. Employment of sequencing technologies coupled with the use of trained neural network algorithms for genotype-to-phenotype prediction will reduce the time needed for antibiotic susceptibility profile identification from days to hours.

In this work, we have sequenced and phenotypically characterized 171 clinical isolates of Escherichia coli and Klebsiella pneumoniae from Norway and India. Based on the data, we have created neural networks to predict susceptibility for ampicillin, 3rd generation cephalosporins and carbapenems. All networks were trained on unassembled data, enabling prediction within minutes after the sequencing information becomes available. Moreover, they can be used both on Illumina and MinION generated data and do not require high genome coverage for phenotype prediction. We cross-checked our networks with previously published algorithms for genotype-to-phenotype prediction and their corresponding datasets. Besides, we also created an ensemble of networks trained on different datasets, which improved the cross-dataset prediction compared to a single network.

Additionally, we have used data from direct sequencing of spiked blood cultures and found that AMR-Diag networks, coupled with MinION sequencing, can predict bacterial species, resistome, and phenotype as fast as 1–8 h from the sequencing start. To our knowledge, this is the first study for genotype-to-phenotype prediction: (1) employing a neural network method; (2) using data from more than one sequencing platform; and (3) utilizing sequence data from spiked blood cultures.

Diving Deep into Short-Term Electricity Load Forecasting: Comparative Analysis and a Novel Framework

In this article, we present an in-depth comparative analysis of the conventional and sequential learning algorithms for electricity load forecasting and optimally select the most appropriate algorithm for energy consumption prediction (ECP). ECP reduces the misusage and wastage of energy using mathematical modeling and supervised learning algorithms. However, the existing ECP research lacks comparative analysis of various algorithms to reach the optimal model with real-world implementation potentials and convincingly reduced error rates. Furthermore, these methods are less friendly towards the energy management chain between the smart grids and residential buildings, with limited contributions in saving energy resources and maintaining an appropriate equilibrium between energy producers and consumers. Considering these limitations, we dive deep into load forecasting methods, analyze their performance, and finally, present a novel three-tier framework for ECP. The first tier applies data preprocessing for its refinement and organization, prior to the actual training, facilitating its effective output generation. The second tier is the learning process, employing ensemble learning algorithms (ELAs) and sequential learning techniques to train over energy consumption data. In the third tier, we obtain the final ECP model and evaluate our method; we visualize the data for energy data analysts. We experimentally prove that deep sequential learning models are dominant over mathematical modeling techniques and its several invariants by utilizing available residential electricity consumption data to reach an optimal proposed model with smallest mean square error (MSE) of value 0.1661 and root mean square error (RMSE) of value 0.4075 against the recent rivals.

Machine Learning Methods for Preterm Birth Prediction: A Review

Preterm births affect around 15 million children a year worldwide. Current medical efforts focus on mitigating the effects of prematurity, not on preventing it. Diagnostic methods are based on parent traits and transvaginal ultrasound, during which the length of the cervix is examined. Approximately 30% of preterm births are not correctly predicted due to the complexity of this process and its subjective assessment. Based on recent research, there is hope that machine learning can be a helpful tool to support the diagnosis of preterm births. The objective of this study is to present various machine learning algorithms applied to preterm birth prediction. The wide spectrum of analysed data sets is the advantage of this survey. They range from electrohysterogram signals through electronic health records to transvaginal ultrasounds. Reviews of works on preterm birth already exist; however, this is the first review that includes works that are based on a transvaginal ultrasound examination. In this work, we present a critical appraisal of popular methods that have employed machine learning methods for preterm birth prediction. Moreover, we summarise the most common challenges incurred and discuss their possible application in the future.

CEGAN: Classification Enhancement Generative Adversarial Networks for unraveling data imbalance problems

The data imbalance problem in classification is a frequent but challenging task. In real-world datasets, numerous class distributions are imbalanced and the classification result under such condition reveals extreme bias in the majority data class. Recently, the potential of GAN as a data augmentation method on minority data has been studied. In this paper, we propose a classification enhancement generative adversarial networks (CEGAN) to enhance the quality of generated synthetic minority data and more importantly, to improve the prediction accuracy in data imbalanced condition. In addition, we propose an ambiguity reduction method using the generated synthetic minority data for the case of multiple similar classes that are degenerating the classification accuracy. The proposed method is demonstrated with five benchmark datasets. The results indicate that approximating the real data distribution using CEGAN improves the classification performance significantly in data imbalanced conditions compared with various standard data augmentation methods.

Dynamic Prediction in Clinical Survival Analysis Using Temporal Convolutional Networks

Accurate prediction of disease trajectories is critical for early identification and timely treatment of patients at risk. Conventional methods in survival analysis are often constrained by strong parametric assumptions and limited in their ability to learn from high-dimensional data. This paper develops a novel convolutional approach that addresses the drawbacks of both traditional statistical approaches as well as recent neural network models for survival. We present Match-Net: a missingness-aware temporal convolutional hitting-time network, designed to capture temporal dependencies and heterogeneous interactions in covariate trajectories and patterns of missingness. To the best of our knowledge, this is the first investigation of temporal convolutions in the context of dynamic prediction for personalized risk prognosis. Using real-world data from the Alzheimer's disease neuroimaging initiative, we demonstrate state-of-the-art performance without making any assumptions regarding underlying longitudinal or time-to-event processes-attesting to the model's potential utility in clinical decision support.

The precision-recall plot is more informative than the ROC plot when evaluating binary classifiers on imbalanced datasets

Binary classifiers are routinely evaluated with performance measures such as sensitivity and specificity, and performance is frequently illustrated with Receiver Operating Characteristics (ROC) plots. Alternative measures such as positive predictive value (PPV) and the associated Precision/Recall (PRC) plots are used less frequently. Many bioinformatics studies develop and evaluate classifiers that are to be applied to strongly imbalanced datasets in which the number of negatives outweighs the number of positives significantly. While ROC plots are visually appealing and provide an overview of a classifier's performance across a wide range of specificities, one can ask whether ROC plots could be misleading when applied in imbalanced classification scenarios. We show here that the visual interpretability of ROC plots in the context of imbalanced datasets can be deceptive with respect to conclusions about the reliability of classification performance, owing to an intuitive but wrong interpretation of specificity. PRC plots, on the other hand, can provide the viewer with an accurate prediction of future classification performance due to the fact that they evaluate the fraction of true positives among positive predictions. Our findings have potential implications for the interpretation of a large number of studies that use ROC plots on imbalanced datasets.

Incremental learning with SVM for multimodal classification of prostatic adenocarcinoma

Robust detection of prostatic cancer is a challenge due to the multitude of variants and their representation in MR images. We propose a pattern recognition system with an incremental learning ensemble algorithm using support vector machines (SVM) tackling this problem employing multimodal MR images and a texture-based information strategy. The proposed system integrates anatomic, texture, and functional features. The data set was preprocessed using B-Spline interpolation, bias field correction and intensity standardization. First- and second-order angular independent statistical approaches and rotation invariant local phase quantization (RI-LPQ) were utilized to quantify texture information. An incremental learning ensemble SVM was implemented to suit working conditions in medical applications and to improve effectiveness and robustness of the system. The probability estimation of cancer structures was calculated using SVM and the corresponding optimization was carried out with a heuristic method together with a 3-fold cross-validation methodology. We achieved an average sensitivity of 0.844 ± 0.068 and a specificity of 0.780 ± 0.038, which yielded superior or similar performance to current state of the art using a total database of only 41 slices from twelve patients with histological confirmed information, including cancerous, unhealthy non-cancerous and healthy prostate tissue. Our results show the feasibility of an ensemble SVM being able to learn additional information from new data while preserving previously acquired knowledge and preventing unlearning. The use of texture descriptors provides more salient discriminative patterns than the functional information used. Furthermore, the system improves selection of information, efficiency and robustness of the classification. The generated probability map enables radiologists to have a lower variability in diagnosis, decrease false negative rates and reduce the time to recognize and delineate structures in the prostate.

Channel selection and classification of electroencephalogram signals: an artificial neural network and genetic algorithm-based approach

OBJECTIVE

An electroencephalogram-based (EEG-based) brain-computer-interface (BCI) provides a new communication channel between the human brain and a computer. Amongst the various available techniques, artificial neural networks (ANNs) are well established in BCI research and have numerous successful applications. However, one of the drawbacks of conventional ANNs is the lack of an explicit input optimization mechanism. In addition, results of ANN learning are usually not easily interpretable. In this paper, we have applied an ANN-based method, the genetic neural mathematic method (GNMM), to two EEG channel selection and classification problems, aiming to address the issues above.

METHODS AND MATERIALS

Pre-processing steps include: least-square (LS) approximation to determine the overall signal increase/decrease rate; locally weighted polynomial regression (Loess) and fast Fourier transform (FFT) to smooth the signals to determine the signal strength and variations. The GNMM method consists of three successive steps: (1) a genetic algorithm-based (GA-based) input selection process; (2) multi-layer perceptron-based (MLP-based) modelling; and (3) rule extraction based upon successful training. The fitness function used in the GA is the training error when an MLP is trained for a limited number of epochs. By averaging the appearance of a particular channel in the winning chromosome over several runs, we were able to minimize the error due to randomness and to obtain an energy distribution around the scalp. In the second step, a threshold was used to select a subset of channels to be fed into an MLP, which performed modelling with a large number of iterations, thus fine-tuning the input/output relationship. Upon successful training, neurons in the input layer are divided into four sub-spaces to produce if-then rules (step 3). Two datasets were used as case studies to perform three classifications. The first data were electrocorticography (ECoG) recordings that have been used in the BCI competition III. The data belonged to two categories, imagined movements of either a finger or the tongue. The data were recorded using an 8 × 8 ECoG platinum electrode grid at a sampling rate of 1000 Hz for a total of 378 trials. The second dataset consisted of a 32-channel, 256 Hz EEG recording of 960 trials where participants had to execute a left- or right-hand button-press in response to left- or right-pointing arrow stimuli. The data were used to classify correct/incorrect responses and left/right hand movements.

RESULTS

For the first dataset, 100 samples were reserved for testing, and those remaining were for training and validation with a ratio of 90%:10% using K-fold cross-validation. Using the top 10 channels selected by GNMM, we achieved a classification accuracy of 0.80 ± 0.04 for the testing dataset, which compares favourably with results reported in the literature. For the second case, we performed multi-time-windows pre-processing over a single trial. By selecting 6 channels out of 32, we were able to achieve a classification accuracy of about 0.86 for the response correctness classification and 0.82 for the actual responding hand classification, respectively. Furthermore, 139 regression rules were identified after training was completed.

CONCLUSIONS

We demonstrate that GNMM is able to perform effective channel selections/reductions, which not only reduces the difficulty of data collection, but also greatly improves the generalization of the classifier. An important step that affects the effectiveness of GNMM is the pre-processing method. In this paper, we also highlight the importance of choosing an appropriate time window position.

Randomized trial of model predictive control for improved anemia management

BACKGROUND AND OBJECTIVES

Variable hemoglobin (Hb) response to erythropoiesis stimulating agents may result in adverse outcomes. The utility of model predictive control for drug dosing was previously demonstrated.

DESIGN, SETTING, PARTICIPANTS, & MEASUREMENTS

This was a double-blinded, randomized, controlled trial to test model predictive control for dosing erythropoietin in ESRD patients. The trial included 60 hemodialysis patients who were randomized into a treatment arm (30 subjects) that received erythropoietin doses on the basis of the computer recommendations or a control arm (30 subjects) that received erythropoietin doses on the basis of recommendations from a standard anemia management protocol (control). The subjects were followed for 8 months, and the proportions of measured Hb within the target of 11 to 12 g/dl and outside 9 to 13 g/dl were measured. Variability of the Hb level was measured by the absolute difference between the achieved Hb and the target Hb of 11.5 g/dl as well as the area under the Hb curve.

RESULTS

Model predictive control resulted in 15 observations >13 or <9 g/dl (outliers), a mean absolute difference between achieved Hb and 11.5 g/dl of 0.98 +/- 0.08 g/dl, and an area under the Hb curve of 2.86 +/- 1.46. The control group algorithm resulted in 30 Hb outliers (P = 0.051), produced a mean absolute difference between achieved Hb and 11.5 g/dl of 1.18 +/- 0.18 g/dl (P < 0.001 difference in variance), and an area under the Hb curve of 3.38 +/- 2.69 (P = 0.025 difference in variance).

CONCLUSIONS

Model predictive control of erythropoietin administration improves anemia management.

A comparative analysis of multi-level computer-assisted decision making systems for traumatic injuries

BACKGROUND

This paper focuses on the creation of a predictive computer-assisted decision making system for traumatic injury using machine learning algorithms. Trauma experts must make several difficult decisions based on a large number of patient attributes, usually in a short period of time. The aim is to compare the existing machine learning methods available for medical informatics, and develop reliable, rule-based computer-assisted decision-making systems that provide recommendations for the course of treatment for new patients, based on previously seen cases in trauma databases. Datasets of traumatic brain injury (TBI) patients are used to train and test the decision making algorithm. The work is also applicable to patients with traumatic pelvic injuries.

METHODS

Decision-making rules are created by processing patterns discovered in the datasets, using machine learning techniques. More specifically, CART and C4.5 are used, as they provide grammatical expressions of knowledge extracted by applying logical operations to the available features. The resulting rule sets are tested against other machine learning methods, including AdaBoost and SVM. The rule creation algorithm is applied to multiple datasets, both with and without prior filtering to discover significant variables. This filtering is performed via logistic regression prior to the rule discovery process.

RESULTS

For survival prediction using all variables, CART outperformed the other machine learning methods. When using only significant variables, neural networks performed best. A reliable rule-base was generated using combined C4.5/CART. The average predictive rule performance was 82% when using all variables, and approximately 84% when using significant variables only. The average performance of the combined C4.5 and CART system using significant variables was 89.7% in predicting the exact outcome (home or rehabilitation), and 93.1% in predicting the ICU length of stay for airlifted TBI patients.

CONCLUSION

This study creates an efficient computer-aided rule-based system that can be employed in decision making in TBI cases. The rule-bases apply methods that combine CART and C4.5 with logistic regression to improve rule performance and quality. For final outcome prediction for TBI cases, the resulting rule-bases outperform systems that utilize all available variables.

Automated detection of prostatic adenocarcinoma from high-resolution ex vivo MRI

Prostatic adenocarcinoma is the most commonly occurring cancer among men in the United States, second only to skin cancer. Currently, the only definitive method to ascertain the presence of prostatic cancer is by trans-rectal ultrasound (TRUS) directed biopsy. Owing to the poor image quality of ultrasound, the accuracy of TRUS is only 20%-25%. High-resolution magnetic resonance imaging (MRI) has been shown to have a higher accuracy of prostate cancer detection compared to ultrasound. Consequently, several researchers have been exploring the use of high resolution MRI in performing prostate biopsies. Visual detection of prostate cancer, however, continues to be difficult owing to its apparent lack of shape, and the fact that several malignant and benign structures have overlapping intensity and texture characteristics. In this paper, we present a fully automated computer-aided detection (CAD) system for detecting prostatic adenocarcinoma from 4 Tesla ex vivo magnetic resonance (MR) imagery of the prostate. After the acquired MR images have been corrected for background inhomogeneity and nonstandardness, novel three-dimensional (3-D) texture features are extracted from the 3-D MRI scene. A Bayesian classifier then assigns each image voxel a "likelihood" of malignancy for each feature independently. The "likelihood" images generated in this fashion are then combined using an optimally weighted feature combination scheme. Quantitative evaluation was performed by comparing the CAD results with the manually ascertained ground truth for the tumor on the MRI. The tumor labels on the MR slices were determined manually by an expert by visually registering the MR slices with the corresponding regions on the histology slices. We evaluated our CAD system on a total of 33 two-dimensional (2-D) MR slices from five different 3-D MR prostate studies. Five slices from two different glands were used for training. Our feature combination scheme was found to outperform the individual texture features, and also other popularly used feature combination methods, including AdaBoost, ensemble averaging, and majority voting. Further, in several instances our CAD system performed better than the experts in terms of accuracy, the expert segmentations being determined solely from visual inspection of the MRI data. In addition, the intrasystem variability (changes in CAD accuracy with changes in values of system parameters) was significantly lower than the corresponding intraobserver and interobserver variability. CAD performance was found to be very similar for different training sets. Future work will focus on extending the methodology to guide high-resolution MRI-assisted in vivo prostate biopsies.

Computer-aided tumor detection in endoscopic video using color wavelet features

We present an approach to the detection of tumors in colonoscopic video. It is based on a new color feature extraction scheme to represent the different regions in the frame sequence. This scheme is built on the wavelet decomposition. The features named as color wavelet covariance (CWC) are based on the covariances of second-order textural measures and an optimum subset of them is proposed after the application of a selection algorithm. The proposed approach is supported by a linear discriminant analysis (LDA) procedure for the characterization of the image regions along the video frames. The whole methodology has been applied on real data sets of color colonoscopic videos. The performance in the detection of abnormal colonic regions corresponding to adenomatous polyps has been estimated high, reaching 97% specificity and 90% sensitivity.

Urinary nucleosides as potential tumor markers evaluated by learning vector quantization

Modified nucleosides were recently presented as potential tumor markers for breast cancer. The patterns of the levels of urinary nucleosides are different for tumor bearing individuals and for healthy individuals. Thus, a powerful pattern recognition method is needed. Although backpropagation (BP) neural networks are becoming increasingly common in medical literature for pattern recognition, it has been shown that often-superior methods exist like learning vector quantization (LVQ) and support vector machines (SVM). The aim of this feasibility study is to get an indication of the performance of urinary nucleoside levels evaluated by LVQ in contrast to the evaluation the popular BP and SVM networks. Urine samples were collected from female breast cancer patients and from healthy females. Twelve different ribonucleosides were isolated and quantified by a high performance liquid chromatography (HPLC) procedure. LVQ, SVM and BP networks were trained and the performance was evaluated by the classification of the test sets into the categories "cancer" and "healthy". All methods showed a good classification with a sensitivity ranging from 58.8 to 70.6% at a specificity of 88.4-94.2% for the test patterns. Although the classification performance of all methods is comparable, the LVQ implementations are superior in terms of more qualitative features: the results of LVQ networks are more reproducible, as the initialization is deterministic. The LVQ networks can be trained by unbalanced sizes of the different classes. LVQ networks are fast during training, need only few parameters adjusted for training and can be retrained by patterns of "local individuals". As at least some of these features play an important role in an implementation into a medical decision support system, it is recommended to use LVQ for an extended study.