Anders Holst

A HIGHER ORDER BAYESIAN NEURAL NETWORK WITH SPIKING UNITS

We treat a Bayesian confidence propagation neural network, primarily in a classifier context. The one-layer version of the network implements a naive Bayesian classifier, which requires the input attributes to be independent. This limitation is overcome by a higher order network. The higher order Bayesian neural network is evaluated on a real world task of diagnosing a telephone exchange computer. By introducing stochastic spiking units, and soft interval coding, it is also possible to handle uncertain as well as continuous valued inputs.

Analyses of cerebral microdialysis in patients with traumatic brain injury: relations to intracranial pressure, cerebral perfusion pressure and catheter placement.

BackgroundCerebral microdialysis (MD) is used to monitor local brain chemistry of patients with traumatic brain injury (TBI). Despite an extensive literature on cerebral MD in the clinical setting, it remains unclear how individual levels of real-time MD data are to be interpreted. Intracranial pressure (ICP) and cerebral perfusion pressure (CPP) are important continuous brain monitors in neurointensive care. They are used as surrogate monitors of cerebral blood flow and have an established relation to outcome. The purpose of this study was to investigate the relations between MD parameters and ICP and/or CPP in patients with TBI.MethodsCerebral MD, ICP and CPP were monitored in 90 patients with TBI. Data were extensively analyzed, using over 7,350 samples of complete (hourly) MD data sets (glucose, lactate, pyruvate and glycerol) to seek representations of ICP, CPP and MD that were best correlated. MD catheter positions were located on computed tomography scans as pericontusional or nonpericontusional. MD markers were analyzed for correlations to ICP and CPP using time series regression analysis, mixed effects models and nonlinear (artificial neural networks) computer-based pattern recognition methods.ResultsDespite much data indicating highly perturbed metabolism, MD shows weak correlations to ICP and CPP. In contrast, the autocorrelation of MD is high for all markers, even at up to 30 future hours. Consequently, subject identity alone explains 52% to 75% of MD marker variance. This indicates that the dominant metabolic processes monitored with MD are long-term, spanning days or longer. In comparison, short-term (differenced or Δ) changes of MD vs. CPP are significantly correlated in pericontusional locations, but with less than 1% explained variance. Moreover, CPP and ICP were significantly related to outcome based on Glasgow Outcome Scale scores, while no significant relations were found between outcome and MD.ConclusionsThe multitude of highly perturbed local chemistry seen with MD in patients with TBI predominately represents long-term metabolic patterns and is weakly correlated to ICP and CPP. This suggests that disturbances other than pressure and/or flow have a dominant influence on MD levels in patients with TBI.

Filaments of meaning in word space

Word space models, in the sense of vector space models built on distributional data taken from texts, are used to model semantic relations between words. We argue that the high dimensionality of typical vector space models lead to unintuitive effects on modeling likeness of meaning and that the local structure of word spaces is where interesting semantic relations reside.We show that the local structure of word spaces has substantially different dimensionality and character than the global space and that this structure shows potential to be exploited for further semantic analysis using methods for local analysis of vector space structure rather than globally scoped methods typically in use today such as singular value decomposition or principal component analysis.

Multivariate Outcome Prediction in Traumatic Brain Injury with Focus on Laboratory Values

Traumatic brain injury (TBI) is a major cause of morbidity and mortality. Identifying factors relevant to outcome can provide a better understanding of TBI pathophysiology, in addition to aiding prognostication. Many common laboratory variables have been related to outcome but may not be independent predictors in a multivariate setting. In this study, 757 patients were identified in the Karolinska TBI database who had retrievable early laboratory variables. These were analyzed towards a dichotomized Glasgow Outcome Scale (GOS) with logistic regression and relevance vector machines, a non-linear machine learning method, univariately and controlled for the known important predictors in TBI outcome: age, Glasgow Coma Score (GCS), pupil response, and computed tomography (CT) score. Accuracy was assessed with Nagelkerkes pseudo R². Of the 18 investigated laboratory variables, 15 were found significant (p<0.05) towards outcome in univariate analyses. In contrast, when adjusting for other predictors, few remained significant. Creatinine was found an independent predictor of TBI outcome. Glucose, albumin, and osmolarity levels were also identified as predictors, depending on analysis method. A worse outcome related to increasing osmolarity may warrant further study. Importantly, hemoglobin was not found significant when adjusted for post-resuscitation GCS as opposed to an admission GCS, and timing of GCS can thus have a major impact on conclusions. In total, laboratory variables added an additional 1.3-4.4% to pseudo R².

A FLEXIBLE AND FAULT TOLERANT QUERY-REPLY SYSTEM BASED ON A BAYESIAN NEURAL NETWORK

A query-reply system based on a Bayesian neural network is described. Strategies for generating questions which make the system both efficient and highly fault tolerant are presented. This involves having one phase of question generation intended to quickly reach a hypothesis followed by a phase where verification of the hypothesis is attempted. In addition, both phases have strategies for detecting and removing inconsistencies in the replies from the user. Also described is an explanatory mechanism which gives information related to why a certain hypotheses is reached or question asked. Specific examples of the systems behavior as well as the results of a statistical evaluation are presented.

Dependency derivation in industrial process data

In many industrial processes, finding dependencies and the creation of dependency graphs can increase the understanding of the system significantly. This knowledge can then be used for further optimization and variable selection. Most of the measured attributes in these cases come in the form of time series. There are several ways of determining correlation between series, most of them suffering from specific problems when applied to real-world data. Here, a well performing measure based on the mutual information rate is derived and discussed with results from both synthetic and real data.

A Bayesian Parametric Statistical Anomaly Detection Method for Finding Trends and Patterns in Criminal Behavior

In this paper we describe how Bayesian Principal Anomaly Detection (BPAD) can be used for detecting long and short term trends and anomalies in geographically tagged alarm data. We elaborate on how the detection of such deviations can be used for high-lighting suspected criminal behavior and activities. BPAD has previously been successively deployed and evaluated in several similar domains, including Maritime Domain Awareness, Train Fleet Maintenance, and Alarm filtering. Similar as for those applications, we argue in the paper that the deployment of BPAD in area of crime monitoring potentially can improve the situation awareness of criminal activities, by providing automatic detection of suspicious behaviors, and uncovering large scale patterns.

Fault-tolerant incremental diagnosis with limited historical data

We describe a novel incremental diagnostic system based on a statistical model that is trained from empirical data. The system guides the user by calculating what additional information would be most helpful for the diagnosis. We show that our diagnostic system can produce satisfactory classification rates, using only small amounts of available background information, such that the need of collecting vast quantities of initial training data is reduced. Further, we show that incorporation of inconsistency-checking mechanisms in our diagnostic system reduces the number of incorrect diagnoses caused by erroneous input.

Mode tracking using multiple data streams

Most existing work in information fusion focuses on combining information with well-defined meaning towards a concrete, pre-specified goal. In contradistinction, we instead aim for autonomous disco ...

SALES PREDICTION WITH MARKETING DATA INTEGRATION FOR SUPPLY CHAINS

Predicting customer demand, manifested as future sales, is an important task for supply chain optimisation. The ability to predict future sales can be used to do more efficient resource allocation and planning , avoiding large stocks and long delays in delivery. Sales prediction is inherently difficult, and can benefit greatly from including other information than historical sales data. Here the fore casting module of the DAMASCOS project, D3S2, will be presented, which both incorporates marketing information and customer demand in the sales prediction as well as supplying the user with options and information that makes the prediction much easier to use in practice.