Roland Orre

A Bayesian neural network method for adverse drug reaction signal generation

AbstractObjective: The database of adverse drug reactions (ADRs) held by the Uppsala Monitoring Centre on behalf of the 47 countries of the World Health Organization (WHO) Collaborating Programme for International Drug Monitoring contains nearly two million reports. It is the largest database of this sort in the world, and about 35 000 new reports are added quarterly. The task of trying to find new drug–ADR signals has been carried out by an expert panel, but with such a large volume of material the task is daunting. We have developed a flexible, automated procedure to find new signals with known probability difference from the background data. Method: Data mining, using various computational approaches, has been applied in a variety of disciplines. A Bayesian confidence propagation neural network (BCPNN) has been developed which can manage large data sets, is robust in handling incomplete data, and may be used with complex variables. Using information theory, such a tool is ideal for finding drug–ADR combinations with other variables, which are highly associated compared to the generality of the stored data, or a section of the stored data. The method is transparent for easy checking and flexible for different kinds of search. Results: Using the BCPNN, some time scan examples are given which show the power of the technique to find signals early (captopril–coughing) and to avoid false positives where a common drug and ADRs occur in the database (digoxin–acne; digoxin–rash). A routine application of the BCPNN to a quarterly update is also tested, showing that 1004 suspected drug–ADR combinations reached the 97.5% confidence level of difference from the generality. Of these, 307 were potentially serious ADRs, and of these 53 related to new drugs. Twelve of the latter were not recorded in the CD editions of The physicians Desk Reference orMartindales Extra Pharmacopoea and did not appear in Reactions Weekly online. Conclusion: The results indicate that the BCPNN can be used in the detection of significant signals from the data set of the WHO Programme on International Drug Monitoring. The BCPNN will be an extremely useful adjunct to the expert assessment of very large numbers of spontaneously reported ADRs.

A Data Mining Approach for Signal Detection and Analysis

The WHO database contains over 2.5 million case reports, analysis of this data set is performed with the intention of signal detection. This paper presents an overview of the quantitative method used to highlight dependencies in this data set.The method Bayesian confidence propagation neural network (BCPNN) is used to highlight dependencies in the data set. The method uses Bayesian statistics implemented in a neural network architecture to analyse all reported drug adverse reaction combinations.This method is now in routine use for drug adverse reaction signal detection. Also this approach has been extended to highlight drug group effects and look for higher order dependencies in the WHO data.Quantitatively unexpectedly strong relationships in the data are highlighted relative to general reporting of suspected adverse effects; these associations are then clinically assessed.

Bayesian neural networks with confidence estimations applied to data mining

An international database of case reports, each one describing a possible case of adverse drug reactions (ADRs), is maintained by the Uppsala Monitoring Centre (UMC), for the WHO international prog ...

Duplicate detection in adverse drug reaction surveillance

The WHO Collaborating Centre for International Drug Monitoring in Uppsala, Sweden, maintains and analyses the world’s largest database of reports on suspected adverse drug reaction (ADR) incidents that occur after drugs are on the market. The presence of duplicate case reports is an important data quality problem and their detection remains a formidable challenge, especially in the WHO drug safety database where reports are anonymised before submission. In this paper, we propose a duplicate detection method based on the hit-miss model for statistical record linkage described by Copas and Hilton, which handles the limited amount of training data well and is well suited for the available data (categorical and numerical rather than free text). We propose two extensions of the standard hit-miss model: a hit-miss mixture model for errors in numerical record fields and a new method to handle correlated record fields, and we demonstrate the effectiveness both at identifying the most likely duplicate for a given case report (94.7% accuracy) and at discriminating true duplicates from random matches (63% recall with 71% precision). The proposed method allows for more efficient data cleaning in post-marketing drug safety data sets, and perhaps other knowledge discovery applications as well.

Data-mining analyses of pharmacovigilance signals in relation to relevant comparison drugs

Abstract Objective. The aim of this paper is to demonstrate the usefulness of the Bayesian Confidence Propagation Neural Network (BCPNN) in the detection of drug-specific and drug-group effects in the database of adverse drug reactions of the World Health Organization Programme for International Drug Monitoring. Methods. Examples of drug–adverse reaction combinations highlighted by the BCPNN as quantitative associations were selected. The anatomical therapeutic chemical (ATC) group to which the drug belonged was then identified, and the information component (IC) was calculated for this ATC group and the adverse drug reaction (ADR). The IC of the ATC group with the ADR was then compared with the IC of the drug–ADR by plotting the change in IC and its 95% confidence limit over time for both. Results. The chosen examples show that the BCPNN data-mining approach can identify drug-specific as well as group effects. In the known examples that served as test cases, beta-blocking agents other than practolol are not associated with sclerosing peritonitis, but all angiotensin-converting enzyme inhibitors are associated with coughing, as are antihistamines with heart-rhythm disorders and antipsychotics with myocarditis. The recently identified association between antipsychotics and myocarditis remains even after consideration of concomitant medication. Conclusion. The BCPNN can be used to improve the ability of a signal detection system to highlight group and drug-specific effects.

A hit-miss model for duplicate detection in the WHO drug safety database

The WHO Collaborating Centre for International Drug Monitoring in Uppsala, Sweden, maintains and analyses the worlds largest database of reports on suspected adverse drug reaction incidents that occur after drugs are introduced on the market. As in other post-marketing drug safety data sets, the presence of duplicate records is an important data quality problem and the detection of duplicates in the WHO drug safety database remains a formidable challenge, especially since the reports are anonymised before submitted to the database. However, to our knowledge no work has been published on methods for duplicate detection in post-marketing drug safety data. In this paper, we propose a method for probabilistic duplicate detection based on the hit-miss model for statistical record linkage described by Copas & Hilton. We present two new generalisations of the standard hit-miss model: a hit-miss mixture model for errors in numerical record fields and a new method to handle correlated record fields. We demonstrate the effectiveness of the hit-miss model for duplicate detection in the WHO drug safety database both at identifying the most likely duplicate for a given record (94.7% accuracy) and at discriminating duplicates from random matches (63% recall with 71% precision). The proposed method allows for more efficient data cleaning in post-marketing drug safety data sets, and perhaps other applications throughout the KDD community.

Case Based Imprecision Estimates for Bayes Classifiers with the Bayesian Bootstrap

This article outlines a Bayesian bootstrap method for case based imprecision estimates in Bayes classification. We argue that this approach is an important complement to methods such as k-fold cross validation that are based on overall error rates. It is shown how case based imprecision estimates may be used to improve Bayes classifiers under asymmetrical loss functions. In addition, other approaches to making use of case based imprecision estimates are discussed and illustrated on two real world data sets. Contrary to the common assumption, Bayesian bootstrap simulations indicate that the uncertainty associated with the output of a Bayes classifier is often far from normally distributed.

Bayesian Neural Networks used to Find Adverse Drug Combinations and Drug Related Syndromes

The data mining task we are interrested in is to find associations between variables in a large database. The method we have earlier proposed to find outstanding associations is to compare estimated frequencies of combinations of variables with the frequencies that would be predicted assuming there were no dependencies. The method we now propose use the same strategy as an efficient way of finding complex dependencies, i.e. certain combinations of explanatory variables, mainly medical drugs, which may be highly associated with certain outcome events or combinations of adverse drug reactions (ADRs). Such combinations of ADRs may also be recognized as syndromes.

Process Modelling Using Artificial Neural Networks

We model a part of a process in pulp to paper production using feed forward connected neural networks. A set of parameters related to paper quality is predicted from a set of process values. The predicted values are results from laboratory experiments which are time consuming. The number of training vectors were rather limited. Therefore, our work was focused on finding the relevant inputs for each signal and to find the architecture that was most efficient for each output. The output vector is separated into single values which are predicted on different architectures adapted to each output. A strategy that continuously adapts the process model seems to be useful. In this work the backprop learning algorithm has been used.

A Bayesian Network for Temporal Segmentation

Abstract A recurrent network which segments an unlabeled externally timed sequence of data is presented. The proposed method uses a Bayesian learning scheme earlier investigated, where the relaxation scheme is modified with a few extra parameters, a pairwise correlation threshold and a pairwise conditional probability threshold. The method studied is able to find start and end positions of words which are in an unlabeled continuous stream of characters. The robustness against noise during both learning and recall is studied.