Roland Fried

Medical device alarms

Abstract The high number of false positive alarms has long been known to be a serious problem in critical care medicine – yet it remains unresolved. At the same time, threats to patient safety due to missing or suppressed alarms are being reported. The purpose of this paper is to present results from a workshop titled “Too many alarms? Too few alarms?” organized by the Section Patient Monitoring and the Workgroup Alarms of the German Association of Biomedical Engineering of the Association for Electrical, Electronic and Information Technologies. The current situation regarding alarms and their problems in intensive care, such as lack of clinical relevance, alarm fatigue, workload increases due to clinically irrelevant alarms, usability problems in alarm systems, problems with manuals and training, and missing alarms due to operator error are outlined, followed by a discussion of solutions and strategies to improve the current situation. Finally, the need for more research and development, focusing on signal quality considerations, networking of medical devices at the bedside, diagnostic alarms and predictive warnings, usability of alarm systems, education of healthcare providers, creation of annotated clinical databases for testing, standardization efforts, and patient monitoring in the regular ward, are called for.

Interventions in INGARCH Processes

We study the problem of intervention effects generating various types of outliers in a linear count time series model. This model belongs to the class of observation driven models and extends the class of Gaussian linear time series models within the exponential family framework. Studies about effects of covariates and interventions for count time series models have largely fallen behind due to the fact that the underlying process, whose behavior determines the dynamics of the observed process, is not observed. We suggest a computationally feasible approach to these problems, focusing especially on the detection and estimation of sudden shifts and outliers. To identify successfully such unusual events we employ the maximum of score tests, whose critical values in finite samples are determined by parametric bootstrap. The usefulness of the proposed methods is illustrated using simulated and real data examples.

Robust filtering of time series with trends

We develop and test a robust procedure for extracting an underlying signal in form of a time-varying trend from very noisy time series. The application we have in mind is online monitoring data measured in intensive care, where we find periods of relative constancy, slow monotonic trends, level shifts and many measurement artifacts. A method is needed which allows a fast and reliable denoising of the data and which distinguishes artifacts from clinically relevant changes in the patients condition. We use robust regression functionals for local approximation of the trend in a moving time window. For further improving the robustness of the procedure, we investigate online outlier replacement by e.g. trimming or winsorization based on robust scale estimators. The performance of several versions of the procedure is compared in important data situations and applications to real and simulated data are given.

Modified repeated median filters

We discuss moving window techniques for fast extraction of a signal composed of monotonic trends and abrupt shifts from a noisy time series with irrelevant spikes. Running medians remove spikes and preserve shifts, but they deteriorate in trend periods. Modified trimmed mean filters use a robust scale estimate such as the median absolute deviation about the median (MAD) to select an adaptive amount of trimming. Application of robust regression, particularly of the repeated median, has been suggested for improving upon the median in trend periods. We combine these ideas and construct modified filters based on the repeated median offering better shift preservation. All these filters are compared w.r.t. fundamental analytical properties and in basic data situations. An algorithm for the update of the MAD running in time O(log n) for window width n is presented as well.

Recursive Filtering for Zero Offset Correction of Diving Depth Time Series with GNU R Package diveMove

Zero offset correction of diving depth measured by time-depth recorders is required to remove artifacts arising from temporal changes in accuracy of pressure transducers. Currently used methods for this procedure are in the proprietary software domain, where researchers cannot study it in sufficient detail, so they have little or no control over how their data were changed. GNU R package diveMove implements a procedure in the Free Software domain that consists of recursively smoothing and filtering the input time series using moving quantiles. This paper describes, demonstrates, and evaluates the proposed method by using a “perfect” data set, which is subsequently corrupted to provide input for the proposed procedure. The method is evaluated by comparing the corrected time series to the original, uncorrupted, data set from an Antarctic fur seal (Arctocephalus gazella Peters, 1875). The Root Mean Square Error of the corrected data set, relative to the “perfect” data set, was nearly identical to the magnitude of noise introduced into the latter. The method, thus, provides a flexible, reliable, and efficient mechanism to perform zero offset correction for analyses of diving behaviour. We illustrate applications of the method to data sets from four species with large differences in diving behaviour, measured using different sampling protocols and instrument characteristics.

Interventions in log-linear Poisson autoregression:

We consider the problem of estimating and detecting outliers in count time series data following a log-linear observation driven model. Log-linear models for count time series arise naturally because they correspond to the canonical link function of the Poisson distribution. They yield both positive and negative dependence, and covariate information can be conveniently incorporated. Within this framework, we establish test procedures for detection of unusual events (‘interventions’) leading to different kinds of outliers, we implement joint maximum likelihood estimation of model parameters and outlier sizes and we derive formulae for correcting the data for detected interventions. The effectiveness of the proposed methodology is illustrated with two real data examples. The first example offers a fresh data analytic point of view towards the polio data. Our methodology identifies different forms of outliers in these data by an observation-driven model. The second example deals with some campylobacterosis data which we analyzed in a previous communication, by a different model. The results are reconfirmed by the new model that we put forward in this communication. The reliability of the procedure is verified using artificial data examples.

Elliptical graphical modelling

We propose elliptical graphical models based on conditional uncorrelatedness as a robust generalization of Gaussian graphical models. Letting the population distribution be elliptical instead of normal allows the fitting of data with arbitrarily heavy tails. We study the class of proportionally affine equivariant scatter estimators and show how they can be used to perform elliptical graphical modelling. This leads to a new class of partial correlation estimators and analogues of the classical deviance test. General expressions for the asymptotic variance of partial correlation estimators, unconstrained and under decomposable models, are given, and the asymptotic chi square approximation for the pseudo-deviance test statistic is proved. The feasibility of our approach is demonstrated by a simulation study, using, among others, Tylers scatter estimator, which is distribution-free within the elliptical model. Copyright 2011, Oxford University Press.

The Crying Wolf: Still Crying?

Roland Fried, PhD† In 1994, Lawless compared the situation of alarms in the intensive care unit (ICU) with the boy who cried wolf in the famous fable by Aesop, alluding to the danger of desensitization of caregivers to true medical device alarms through the overwhelming number of false medical device alarms that he observed on a pediatric ICU. Alarm limits may be set dangerously broad, or alarms may even be completely disabled to reduce the nuisance from false alarms. Even at these settings, clinicians may tolerate an alarm for up to 10 min before taking action. This situation cries for immediate remedy. The sad reality, although, is that not much seems to have changed over the nearly 15 yr since Lawless’ publication. The current literature and ongoing research efforts (reviewed in Ref. 6) as well as recent data from our own group, show that still the vast majority of medical device alarms are false positives. Interestingly, there is no scarcity of research addressing the problem of medical device alarms. Many different approaches have been studied in the fields of statistics and artificial intelligence as well as biomedical and human factors engineering. Several approaches have shown efficacy and effectiveness in reducing the rate of false alarms in clinical study. Still, very little has been implemented in commercially available medical devices. In this situation Görges et al. promise hope in their article published in this issue of Anesthesia & Analgesia. In their study, they first acquired comprehensive clinical data on medical device alarms and then investigated two approaches to reduce the number of false-positive alarms. The authors must be commended for their efforts, as we know from other researchers and our own experience how much stamina it takes to acquire alarm data and consistently annotate sufficiently large numbers of medical device alarms. Görges et al. confirm that only the minority of medical device alarms are clinically relevant—in their study, 23% of all alarms. They also found that not only were six alarms activated per hour per bed, but also alarms were sounding 31⁄2 min per hour per bed. Extrapolating to a 10-bed ICU, this means that a false alarm is active, i.e., making some noise or “crying,” nearly 50% of the time, day and night, 24/7. These numbers are in line with other studies. If we keep in mind that it took the boy in Aesop’s fable only two false alarms to make the shepherds ignore the third but true and deadly alarm, the current situation of medical device alarms seems mindboggling. Of course, the study by Görges et al. has distinct weaknesses, most of which the authors diligently discuss: night shifts were not included in the study, the physical presence of the observer may have induced a Hawthorne effect, clinical annotations of alarms were subjective, and there may have been significant intraand interobserver variability. Moreover, clinical practice patterns in the study ICU may differ from other institutions, which may further affect the generalizability of the reported results, as may the differences in annotation schemata between different studies, as pointed out by the authors. But this is true for each and every clinical alarm study published as of today. And still, all studies come to similar conclusions despite their differences in methodology and clinical settings, actually strengthening rather than weakening our point about the inadequacy of current device alarms. From the *Department for Medical Informatics, Biometrics and Epidemiology, Ruhr-University, Bochum, Germany; and †Department of Statistics, Technical University Dortmund, Dortmund, Germany. Accepted for publication January 19, 2009. M.I. and R.F. have received research grants from the German Research Foundation (DFG SFB475). MI has received consulting fees from Draeger Medical and is managing member of Boston MedTech Advisors Europe. Address correspondence and reprint requests to Dr. Michael Imhoff, Am Pastorenwäldchen 2, D-44229 Dortmund, Germany. Address e-mail to mike@imhoff.de. Copyright

Methods and algorithms for robust filtering

We discuss filtering procedures for robust extraction of a signal from noisy time series. Moving averages and running medians are standard methods for this, but they have shortcomings when large spikes (outliers) respectively trends occur. Modified trimmed means and linear median hybrid filters combine advantages of both approaches, but they do not completely overcome the difficulties. Improvements can be achieved by using robust regression methods, which work even in real time because of increased computational power and faster algorithms. Extending recent work we present filters for robust online signal extraction and discuss their merits for preserving trends, abrupt shifts and extremes and for the removal of spikes.

Asymptotic distribution of two-sample empirical U-quantiles with applications to robust tests for shifts in location

We derive the asymptotical distributions of two-sample U-statistics and two-sample empirical U-quantiles in the case of weakly dependent data. Our results apply to observations that can be represented as functionals of absolutely regular processes, including e.g. many classical time series models as well as data from chaotic dynamical systems. Based on these theoretical results we propose a new robust nonparametric test for the two-sample location problem, which is constructed from the median of pairwise differences between the two samples. We inspect the properties of the test in the case of weakly dependent data and compare the performance with classical tests such as the t-test and Wilcoxons two-sample rank test with corrections for dependencies. Simulations indicate that the new test offers better power than the Wilcoxon test in case of skewed and heavy tailed distributions, when at least one of the two samples is not very large. The test is then applied for detecting shifts of location in some weakly dependent time series, which are contaminated by outliers.