Forensic medical practitioners and scientists have for several years sought improved decision sup... more Forensic medical practitioners and scientists have for several years sought improved decision support for determining and managing care and release of prisoners with mental health problems. Some of these prisoners can pose a serious threat of violence to society after release. It is, therefore, critical that the risk of violent reoffending is accurately measured and, more importantly, well managed with causal interventions to reduce this risk after release. The well-established predictors in this area of research are typically based on regression models or even some rule-based methods with no statistical composition, and these have proven to be unsuitable for simulating causal interventions for risk management. In collaboration with the medical practitioners of the Violence Prevention Research Unit (VPRU), Queen Mary University of London, we have developed a Bayesian network (BN) model for this purpose, which we call DSVM-P (Decision Support for Violence Management -Prisoners). The BN model captures the causal relationships between risk factors, interventions and violence and demonstrates significantly higher accuracy (cross-validated AUC score of 0.78) compared to well-established predictors (AUC scores ranging from 0.665 to 0.717) within this area of research, with respect to whether a prisoner is determined suitable for release. Even more important, however, the BN model also allows for specific risk factors to be targeted for causal intervention for risk management of future re-offending. Hence, unlike the previous predictors, this makes the model useful in terms of answering complex clinical questions that are based on unobserved evidence. Clinicians and probation officers who work in these areas would benefit from a system that takes account of these complex risk management considerations, since these decision support features are not available in the previous generation of models used by forensic psychiatrists.
Lower extremity vascular trauma (LEVT) is a major cause of amputation. A clear understanding of p... more Lower extremity vascular trauma (LEVT) is a major cause of amputation. A clear understanding of prognostic factors for amputation is important to inform surgical decision-making, patient counselling and risk stratification. The aim was to develop an understanding of prognostic factors for amputation following surgical repair of LEVT. A systematic review was conducted to identify potential prognostic factors. Bayesian meta-analysis was used to calculate an absolute (pooled proportion) and relative (pooled odds ratio, OR) measure of the amputation risk for each factor. Forty-five studies, totalling 3187 discrete LEVT repairs, were included. The overall amputation rate was 10·0 (95 per cent credible interval 7·4 to 13·1) per cent. Significant prognostic factors for secondary amputation included: associated major soft tissue injury (26 versus 8 per cent for no soft tissue injury; OR 5·80), compartment syndrome (28 versus 6 per cent; OR 5·11), multiple arterial injuries (18 versus 9 per ...
Studies in health technology and informatics, 2014
Many prognostic models are not adopted in clinical practice regardless of their reported accuracy... more Many prognostic models are not adopted in clinical practice regardless of their reported accuracy. Doubts about the basis of the model is considered to be a major reason for this as the evidence behind clinical models is often not clear to anyone other than their developers. We propose a framework for representing the evidence behind Bayesian networks (BN) developed for prognostic decision support. The aim of this evidence framework is to be able to present all the evidence alongside the BN itself. We illustrate this framework by a BN developed with clinical evidence to predict coagulation disorders in trauma care.
Complex clinical decisions require the decision maker to evaluate multiple factors that may inter... more Complex clinical decisions require the decision maker to evaluate multiple factors that may interact with each other. Many clinical studies, however, report 'univariate' relations between a single factor and outcome. Such univariate statistics are often insufficient to provide useful support for complex clinical decisions even when they are pooled using meta-analysis. More useful decision support could be provided by evidence-based models that take the interaction between factors into account. In this paper, we propose a method of integrating the univariate results of a meta-analysis with a clinical dataset and expert knowledge to construct multivariate Bayesian network (BN) models. The technique reduces the size of the dataset needed to learn the parameters of a model of a given complexity. Supplementing the data with the meta-analysis results avoids the need to either simplify the model - ignoring some complexities of the problem - or to gather more data. The method is illustrated by a clinical case study into the prediction of the viability of severely injured lower extremities. The case study illustrates the advantages of integrating combined evidence into BN development: the BN developed using our method outperformed four different data-driven structure learning methods, and a well-known scoring model (MESS) in this domain.
Event trees are a popular technique for modelling accidents in system safety analyses. Bayesian n... more Event trees are a popular technique for modelling accidents in system safety analyses. Bayesian networks are a probabilistic modelling technique representing influences between uncertain variables. Although popular in expert systems, Bayesian networks are not used widely for safety. Using a train derailment case study, we show how an event tree can be viewed as a Bayesian network, making it clearer when one event affects a later one. Since this effect needs to be understood to construct an event tree correctly, we argue that the two notations should be used together. We then show how the Bayesian Network enables the factors that influence the outcome of events to be represented explicitly. In the case study, this allowed the train derailment model to be generalised and applied in more circumstances. Although the resulting model is no longer just an event tree, the familiar event tree notation remains useful.
ABSTRACT Fault trees are used to model how failures lead to hazards and so to estimate the freque... more ABSTRACT Fault trees are used to model how failures lead to hazards and so to estimate the frequencies of the identified hazards of a system. Large systems, such as a rail network, do not give rise to endless different hazards. Rather, similar hazards arise repeatedly but with different frequency depending on factors such as location. Several authors have identified the need to build models to estimate both system-wide average hazard frequencies and hazard frequencies in specific situations. Fault trees can be used for this but they grow as additional factors are considered. In this paper, we describe a compact model using Bayesian networks. The fault tree notation is retained; with base events parameterised by variables in the Bayesian net to represent a mixture of related fault trees compactly. We use a simple example to describe the model structure and report on ongoing work on a model of train derailment.
Probabilistic Safety Assessment and Management, 2004
We describe a method of modelling organisational causes of accidents, using Bayesian Networks. A ... more We describe a method of modelling organisational causes of accidents, using Bayesian Networks. A rigorous method is used to relate interactions within the organisation operating the system to causal factors for accidents. Using examples from a model of the causes of SPAD incidents in the UK railways, we describe how such a model can be used for risk assessment.
Proceedings of the Institution of Mechanical Engineers, Part F: Journal of Rail and Rapid Transit, 2013
ABSTRACT The UK rail industry collects and analyses incident data to assess the risks experienced... more ABSTRACT The UK rail industry collects and analyses incident data to assess the risks experienced by passengers, the railway workforce and members of the public. The current analysis mainly compares current and historic incident rates and, for each type of incident, looks at the railway as a whole. However, changes to reduce risk are often made locally, at least at first, so a modelling approach is needed that is able to analyse local risk. In this paper we present a form of model that is able to make local risk estimates from incident data, using a case study - boarding and alighting incidents at stations. Using a Bayesian network (BN), we analyse the incident data with expert judgements about causal factors. The BN cannot be directly leant from data because the dataset contains no entries for the overwhelming majority of cases where the railway is used without incident. Instead, the incident data is analysed in the context of a model of the use of the railway, so that the prevalence of the causal factors in normal use can be compared with that in the incident data. The usage model is derived from multiple data sets; the use of a BN allows this to be done with approximations. Finally, we show how the model allows local risk to be estimated and consider the wider applicability of the techniques described.
ABSTRACT The graphical structure of a Bayesian network (BN) makes it a technology well-suited for... more ABSTRACT The graphical structure of a Bayesian network (BN) makes it a technology well-suited for developing decision support models from a combination of domain knowledge and data. The domain knowledge of experts is used to determine the graphical structure of the BN, corresponding to the relationships and between variables, and data is used for learning the strength of these relationships. However, the available data seldom match the variables in the structure that is elicited from experts, whose models may be quite detailed; consequently, the structure needs to be abstracted to match the data. Up to now, this abstraction has been informal, loosening the link between the final model and the experts’ knowledge. In this paper, we propose a method for abstracting the BN structure by using four ‘abstraction’ operations: node removal, node merging, state-space collapsing and edge removal. Some of these steps introduce approximations, which can be identified from changes in the set of conditional independence (CI) assertions of a network.
Forensic medical practitioners and scientists have for several years sought improved decision sup... more Forensic medical practitioners and scientists have for several years sought improved decision support for determining and managing care and release of prisoners with mental health problems. Some of these prisoners can pose a serious threat of violence to society after release. It is, therefore, critical that the risk of violent reoffending is accurately measured and, more importantly, well managed with causal interventions to reduce this risk after release. The well-established predictors in this area of research are typically based on regression models or even some rule-based methods with no statistical composition, and these have proven to be unsuitable for simulating causal interventions for risk management. In collaboration with the medical practitioners of the Violence Prevention Research Unit (VPRU), Queen Mary University of London, we have developed a Bayesian network (BN) model for this purpose, which we call DSVM-P (Decision Support for Violence Management -Prisoners). The BN model captures the causal relationships between risk factors, interventions and violence and demonstrates significantly higher accuracy (cross-validated AUC score of 0.78) compared to well-established predictors (AUC scores ranging from 0.665 to 0.717) within this area of research, with respect to whether a prisoner is determined suitable for release. Even more important, however, the BN model also allows for specific risk factors to be targeted for causal intervention for risk management of future re-offending. Hence, unlike the previous predictors, this makes the model useful in terms of answering complex clinical questions that are based on unobserved evidence. Clinicians and probation officers who work in these areas would benefit from a system that takes account of these complex risk management considerations, since these decision support features are not available in the previous generation of models used by forensic psychiatrists.
Lower extremity vascular trauma (LEVT) is a major cause of amputation. A clear understanding of p... more Lower extremity vascular trauma (LEVT) is a major cause of amputation. A clear understanding of prognostic factors for amputation is important to inform surgical decision-making, patient counselling and risk stratification. The aim was to develop an understanding of prognostic factors for amputation following surgical repair of LEVT. A systematic review was conducted to identify potential prognostic factors. Bayesian meta-analysis was used to calculate an absolute (pooled proportion) and relative (pooled odds ratio, OR) measure of the amputation risk for each factor. Forty-five studies, totalling 3187 discrete LEVT repairs, were included. The overall amputation rate was 10·0 (95 per cent credible interval 7·4 to 13·1) per cent. Significant prognostic factors for secondary amputation included: associated major soft tissue injury (26 versus 8 per cent for no soft tissue injury; OR 5·80), compartment syndrome (28 versus 6 per cent; OR 5·11), multiple arterial injuries (18 versus 9 per ...
Studies in health technology and informatics, 2014
Many prognostic models are not adopted in clinical practice regardless of their reported accuracy... more Many prognostic models are not adopted in clinical practice regardless of their reported accuracy. Doubts about the basis of the model is considered to be a major reason for this as the evidence behind clinical models is often not clear to anyone other than their developers. We propose a framework for representing the evidence behind Bayesian networks (BN) developed for prognostic decision support. The aim of this evidence framework is to be able to present all the evidence alongside the BN itself. We illustrate this framework by a BN developed with clinical evidence to predict coagulation disorders in trauma care.
Complex clinical decisions require the decision maker to evaluate multiple factors that may inter... more Complex clinical decisions require the decision maker to evaluate multiple factors that may interact with each other. Many clinical studies, however, report 'univariate' relations between a single factor and outcome. Such univariate statistics are often insufficient to provide useful support for complex clinical decisions even when they are pooled using meta-analysis. More useful decision support could be provided by evidence-based models that take the interaction between factors into account. In this paper, we propose a method of integrating the univariate results of a meta-analysis with a clinical dataset and expert knowledge to construct multivariate Bayesian network (BN) models. The technique reduces the size of the dataset needed to learn the parameters of a model of a given complexity. Supplementing the data with the meta-analysis results avoids the need to either simplify the model - ignoring some complexities of the problem - or to gather more data. The method is illustrated by a clinical case study into the prediction of the viability of severely injured lower extremities. The case study illustrates the advantages of integrating combined evidence into BN development: the BN developed using our method outperformed four different data-driven structure learning methods, and a well-known scoring model (MESS) in this domain.
Event trees are a popular technique for modelling accidents in system safety analyses. Bayesian n... more Event trees are a popular technique for modelling accidents in system safety analyses. Bayesian networks are a probabilistic modelling technique representing influences between uncertain variables. Although popular in expert systems, Bayesian networks are not used widely for safety. Using a train derailment case study, we show how an event tree can be viewed as a Bayesian network, making it clearer when one event affects a later one. Since this effect needs to be understood to construct an event tree correctly, we argue that the two notations should be used together. We then show how the Bayesian Network enables the factors that influence the outcome of events to be represented explicitly. In the case study, this allowed the train derailment model to be generalised and applied in more circumstances. Although the resulting model is no longer just an event tree, the familiar event tree notation remains useful.
ABSTRACT Fault trees are used to model how failures lead to hazards and so to estimate the freque... more ABSTRACT Fault trees are used to model how failures lead to hazards and so to estimate the frequencies of the identified hazards of a system. Large systems, such as a rail network, do not give rise to endless different hazards. Rather, similar hazards arise repeatedly but with different frequency depending on factors such as location. Several authors have identified the need to build models to estimate both system-wide average hazard frequencies and hazard frequencies in specific situations. Fault trees can be used for this but they grow as additional factors are considered. In this paper, we describe a compact model using Bayesian networks. The fault tree notation is retained; with base events parameterised by variables in the Bayesian net to represent a mixture of related fault trees compactly. We use a simple example to describe the model structure and report on ongoing work on a model of train derailment.
Probabilistic Safety Assessment and Management, 2004
We describe a method of modelling organisational causes of accidents, using Bayesian Networks. A ... more We describe a method of modelling organisational causes of accidents, using Bayesian Networks. A rigorous method is used to relate interactions within the organisation operating the system to causal factors for accidents. Using examples from a model of the causes of SPAD incidents in the UK railways, we describe how such a model can be used for risk assessment.
Proceedings of the Institution of Mechanical Engineers, Part F: Journal of Rail and Rapid Transit, 2013
ABSTRACT The UK rail industry collects and analyses incident data to assess the risks experienced... more ABSTRACT The UK rail industry collects and analyses incident data to assess the risks experienced by passengers, the railway workforce and members of the public. The current analysis mainly compares current and historic incident rates and, for each type of incident, looks at the railway as a whole. However, changes to reduce risk are often made locally, at least at first, so a modelling approach is needed that is able to analyse local risk. In this paper we present a form of model that is able to make local risk estimates from incident data, using a case study - boarding and alighting incidents at stations. Using a Bayesian network (BN), we analyse the incident data with expert judgements about causal factors. The BN cannot be directly leant from data because the dataset contains no entries for the overwhelming majority of cases where the railway is used without incident. Instead, the incident data is analysed in the context of a model of the use of the railway, so that the prevalence of the causal factors in normal use can be compared with that in the incident data. The usage model is derived from multiple data sets; the use of a BN allows this to be done with approximations. Finally, we show how the model allows local risk to be estimated and consider the wider applicability of the techniques described.
ABSTRACT The graphical structure of a Bayesian network (BN) makes it a technology well-suited for... more ABSTRACT The graphical structure of a Bayesian network (BN) makes it a technology well-suited for developing decision support models from a combination of domain knowledge and data. The domain knowledge of experts is used to determine the graphical structure of the BN, corresponding to the relationships and between variables, and data is used for learning the strength of these relationships. However, the available data seldom match the variables in the structure that is elicited from experts, whose models may be quite detailed; consequently, the structure needs to be abstracted to match the data. Up to now, this abstraction has been informal, loosening the link between the final model and the experts’ knowledge. In this paper, we propose a method for abstracting the BN structure by using four ‘abstraction’ operations: node removal, node merging, state-space collapsing and edge removal. Some of these steps introduce approximations, which can be identified from changes in the set of conditional independence (CI) assertions of a network.
Uploads
Papers by William Marsh