Robert F. Harrison

Introduction to neural networks

book for non-commercial use, as long as it is distributed as a whole in its original form, and the names of the authors and the University of Amsterdam are mentioned. Permission is also granted to use this book for non-commercial courses, provided the authors are notiied of this beforehand.

Free triiodothyronine has a distinct circadian rhythm that is delayed but parallels thyrotropin levels.

CONTEXT TSH is known to have a circadian rhythm, but the relationship between this and any rhythm in T(4) and T(3) has not been clearly demonstrated. OBJECTIVE With a view to optimizing thyroid hormone replacement therapy, we have used modern assays for free T(4) (FT4) and free T(3) (FT3) to investigate circadian rhythmicity. SETTING The study was performed at a university hospital. DESIGN AND SUBJECTS This was a cross-sectional study in 33 healthy individuals with 24-h blood sampling (TSH in 33 and FT4 and FT3 in 29 individuals) and cosinor analysis. RESULTS Of the individuals, 100% showed a sinusoidal signal in TSH, for FT4 76%, and for FT3 86% (P < 0.05). For FT4 and FT3, the amplitude was low. For TSH the acrophase occurred at a clock time of 0240 h, and for FT3 approximately 90 minutes later at 0404 h. The group cosinor model predicts that TSH hormone levels remain above the mesor between 2020 and 0820 h, and for FT3 from 2200-1000 h. Cross correlation of FT3 with TSH showed that the peak correlation occurred with a delay of 0.5-2.5 h. When time-adjusted profiles of TSH and FT3 were compared, there was a strong correlation between FT3 and TSH levels (rho = 0.80; P < 0.0001). In contrast, cross correlation revealed no temporal relationship between FT4 and TSH. CONCLUSIONS FT3 shows a circadian rhythm with a periodicity that lags behind TSH, suggesting that the periodic rhythm of FT3 is due to the proportion of T(3) derived from the thyroid. Optimizing thyroid hormone replacement may need to take these rhythms into account.

A modified fuzzy ARTMAP architecture for the approximation of noisy mappings

Abstract A neural architecture, fuzzy ARTMAP, is considered here as an alternative to standard feedforward networks for noisy mapping tasks. It is one of a series of architectures based upon adaptive resonance theory or ART. Like other ART-based systems, fuzzy ARTMAP has advantages over feedforward networks and is especially suited to classification-type problems. Here it is used to approximate a noisy continuous mapping. Results show that properties that confer useful advantages for classification problems do not necessarily confer similar advantages for noisy mapping problems. One particular feature, match tracking, is found to cause overlearning of the data. A modified variant is proposed, without match tracking, that stores probability information in the map field. This information is subsequently used to compute output estimates. The proposed fuzzy ARTMAP variant is found to outperform fuzzy ARTMAP in a mapping task.

Evaluation of machine-learning methods for ligand-based virtual screening

Machine-learning methods can be used for virtual screening by analysing the structural characteristics of molecules of known (in)activity, and we here discuss the use of kernel discrimination and naive Bayesian classifier (NBC) methods for this purpose. We report a kernel method that allows the processing of molecules represented by binary, integer and real-valued descriptors, and show that it is little different in screening performance from a previously described kernel that had been developed specifically for the analysis of binary fingerprint representations of molecular structure. We then evaluate the performance of an NBC when the training-set contains only a very few active molecules. In such cases, a simpler approach based on group fusion would appear to provide superior screening performance, especially when structurally heterogeneous datasets are to be processed.

An artificial neural network system for diagnosis of acute myocardial infarction (AMI) in the accident and emergency department: evaluation and comparison with serum myoglobin measurements

Recent studies have confirmed that artificial neural networks (ANNs) are adept at recognising patterns in sets of clinical data. The diagnosis of acute myocardial infarction (AMI) in patients presenting with chest pain remains one of the greatest challenges in emergency medicine. The aim of this study was to evaluate the performance of an ANN trained to analyse clinical data from chest pain patients. The ANN was compared with serum myoglobin measurements--cardiac damage is associated with increased circulating myoglobin levels, and this is widely used as an early marker for evolving AMI. We used 39 items of clinical and ECG data from the time of presentation to derive 53 binary inputs to a back propagation network. On test data (200 cases), overall accuracy, sensitivity, specificity and positive predictive value (PPV) of the ANN were 91.8, 91.2, 90.2 and 84.9% respectively. Corresponding figures using linear discriminant analysis were 81.0, 77.9, 82.6 and 69.7% (P < 0.01). Using a further test set from a different centre (91 cases), the accuracy, sensitivity, specificity and PPV for the admitting physicians were 65.1, 28.5, 76.9 and 28.6% respectively compared with 73.6, 52.4, 80.0 and 44.0% for the ANN. Although myoglobin at presentation was highly specific, it was only 38.0% sensitive, compared with 85.7% at 3 h. Simple strategies to combine clinical opinion, ANN output and myoglobin at presentation could greatly improve sensitivity and specificity of AMI diagnosis. The ideal support for emergency room physicians may come from a combination of computer-aided analysis of clinical factors and biochemical markers such as myoglobin. This study demonstrates that the two approaches could be usefully combined, the major benefit of the decision support system being in the first 3 h before biochemical markers have become abnormal.

Expression of receptor activator of nuclear factor κβ ligand (RANKL) and tumour necrosis factor related, apoptosis inducing ligand (TRAIL) in breast cancer, and their relations with osteoprotegerin, oestrogen receptor, and clinicopathological variables

Background: Receptor activator of nuclear factor κβ ligand (RANKL) has an important role in bone remodelling, and tumour necrosis factor related, apoptosis inducing ligand (TRAIL) can induce apoptosis in cancer cells. Their functions are linked by their interactions with osteoprotegerin (OPG). Objective: To investigate the expression of RANKL and TRAIL in a large series of unselected breast cancers and to analyse the relations between these expressions and the expression of OPG, oestrogen receptor, and clinicopathological variables. Methods: 395 breast cancers were sampled into tissue microarrays and immunohistochemistry undertaken for RANKL and TRAIL. Results: There was strong expression of RANKL in 14% of the cancers and strong expression of TRAIL in 30%. Expression of RANKL had a negative association with expression of oestrogen receptor (p = 0.036). Expression of TRAIL had a negative association with the Nottingham Prognostic Index (p = 0.021). There was a significant negative relation between expression of RANKL and TRAIL (p<0.005). Unsupervised cluster analysis produced a dendrogram that showed a clear division into two groups, and the expression of oestrogen receptor was significantly higher in one of those groups (p = 0.012). Conclusions: There is apparent loss of expression of RANKL in 86% of breast cancers; those tumours that retain expression tend to be oestrogen receptor negative and of a high histological grade. There is strong expression of TRAIL in 30% of breast cancers and these tend to be of better prognostic type. These results may be important in the processes of metastasis to bone and the apoptotic cell death pathway in cancer.

Application of autonomous neural network systems to medical pattern classification tasks

This paper presents a study of the application of autonomously learning multiple neural network systems to medical pattern classification tasks. In our earlier work, a hybrid neural network architecture has been developed for on-line learning and probability estimation tasks. The network has been shown to be capable of asymptotically achieving the Bayes optimal classification rates, on-line, in a number of benchmark classification experiments. In the context of pattern classification, however, the concept of multiple classifier systems has been proposed to improve the performance of a single classifier. Thus, three decision combination algorithms have been implemented to produce a multiple neural network classifier system. Here the applicability of the system is assessed using patient records in two medical domains. The first task is the prognosis of patients admitted to coronary care units; whereas the second is the prediction of survival in trauma patients. The results are compared with those from logistic regression models, and implications of the system as a useful clinical diagnostic tool are discussed.

The early diagnosis of heart attacks: a neurocomputational approach

A multilayered perceptron (MLP) was trained to diagnose the presence of acute myocardial infarction (heart attack) in patients admitted to an emergency unit with acute chest pain. Two learning algorithms, based on mean-square-error and the log-likelihood function, are compared. Their performance does not differ significantly, but the latter rule converges much more rapidly. Performance in excess of that of the admitting clinicians was achieved for a number of performance indicators, and a protocol for combining the networks diagnosis with that of the clinician is proposed. This results in further improvements in performance, indicating that the MLP can act as a useful decision aid in an emergency context.<<ETX>>

Virtual screening using binary kernel discrimination : Analysis of pesticide data

This paper discusses the use of binary kernel discrimination (BKD) for identifying potential active compounds in lead-discovery programs. BKD was compared with established virtual screening methods in a series of experiments using pesticide data from the Syngenta corporate database. It was found to be superior to methods based on similarity searching and substructural analysis but inferior to a support vector machine. Similar conclusions resulted from application of the methods to a pesticide data set for which categorical activity data were available.

Ownership and uses of human tissue: what are the opinions of surgical in-patients?

Aims: To investigate whether patient opinion about the uses of tissue removed at therapeutic operations has changed since the adverse publicity surrounding the Alder Hey and Bristol Royal Infirmary Inquiries, and to see whether it aligns with the Human Tissue Act 2004. Methods: A questionnaire was given to 220 postoperative patients in a teaching hospital during an 11 week period. Aggregated responses to each question were ranked in frequency order. Unweighted centroid linkage hierarchical clustering analysis was performed with dendrogram display for the main data on tissue usage. Results: 203 completed questionnaires were collected (compliance rate 92.3%). 96.3% of patients indicated that they would not object to their tissue being used in research, significantly higher than in the 1996 study (89.1%) with no overlap of the 95% CIs. 29.1% of patients believed that the hospital had ownership of tissue once it has been removed during surgery, 23.2% believed they had ownership, 19.7% believed that the pathology laboratory had ownership, and 15.3% believed that nobody had ownership rights in the case of tissue samples. Conclusions: This new survey indicates that despite a turbulent decade for those involved in human tissue retention in the UK, public support for a wide range of human tissue based activities, especially biomedical research, has not diminished and that patient opinion aligns well with the Human Tissue Act 2004.