TIME SERIES ANALYSIS AS INPUT FOR PREDICTIVE MODELING: PREDICTING CARDIAC ARREST IN A PEDIATRIC INTENSIVE CARE UNIT

The first manuscript, entitled "Time-Series Analysis as Input for Clinical Predictive Modeling: Modeling Cardiac Arrest in a Pediatric ICU" lays out the theoretical background for the project. There are several core concepts presented in this paper. First, traditional multivariate models (where each variable is represented by only one value) provide single point-in-time snapshots of patient status: they are incapable of characterizing deterioration. Since deterioration is consistently identified as a precursor to cardiac arrests, we maintain that the traditional multivariate paradigm is insufficient for predicting arrests. We identify time series analysis as a method capable of characterizing deterioration in an objective, mathematical fashion, and describe how to build a general foundation for predictive modeling using time series analysis results as latent variables. Building a solid foundation for any given modeling task involves addressing a number of issues during the design phase. These include selecting the proper candidate features on which to base the model, and selecting the most appropriate tool to measure them. We also identified several unique design issues that are introduced when time series data elements are added to the set of candidate features. One such issue is in defining the duration and resolution of time series elements required to sufficiently characterize the time series phenomena being considered as candidate features for the predictive model. Once the duration and resolution are established, there must also be explicit mathematical or statistical operations that produce the time series analysis result to be used as a latent candidate feature. In synthesizing the comprehensive framework for building a predictive model based on time series data elements, we identified at least four classes of data that can be used in the model design. The first two classes are shared with traditional multivariate models: multivariate data and clinical latent features. Multivariate data is represented by the standard one value per variable paradigm and is widely employed in a host of clinical models and tools. These are often represented by a number present in a given cell of a table. Clinical latent features derived, rather than directly measured, data elements that more accurately represent a particular clinical phenomenon than any of the directly measured data elements in isolation. The second two classes are unique to the time series data elements. The first of these is the raw data elements. These are represented by multiple values per variable, and constitute the measured observations that are typically available to end users when they review time series data. These are often represented as dots on a graph. The final class of data results from performing time series analysis. This class of data represents the fundamental concept on which our hypothesis is based. The specific statistical or mathematical operations are up to the modeler to determine, but we generally recommend that a variety of analyses be performed in order to maximize the likelihood that a representation of the time series data elements is produced that is able to distinguish between two or more classes of outcomes. The second manuscript, entitled "Building Clinical Prediction Models Using Time Series Data: Modeling Cardiac Arrest in a Pediatric ICU" provides a detailed description, start to finish, of the methods required to prepare the data, build, and validate a predictive model that uses the time series data elements determined in the first paper. One of the fundamental tenets of the second paper is that manual implementations of time series based models are unfeasible due to the relatively large number of data elements and the complexity of preprocessing that must occur before data can be presented to the model. Each of the seventeen steps is analyzed from the perspective of how it may be automated, when necessary. We identify the general objectives and available strategies of each of the steps, and we present our rationale for choosing a specific strategy for each step in the case of predicting cardiac arrest in a pediatric intensive care unit. Another issue brought to light by the second paper is that the individual steps required to use time series data for predictive modeling are more numerous and more complex than those used for modeling with traditional multivariate data. Even after complexities attributable to the design phase (addressed in our first paper) have been accounted for, the management and manipulation of the time series elements (the preprocessing steps in particular) are issues that are not present in a traditional multivariate modeling paradigm. In our methods, we present the issues that arise from the time series data elements: defining a reference time; imputing and reducing time series data in order to conform to a predefined structure that was specified during the design phase; and normalizing variable families rather than individual variable instances. The final manuscript, entitled: "Using Time-Series Analysis to Predict Cardiac Arrest in a Pediatric Intensive Care Unit" presents the results that were obtained by applying the theoretical construct and its associated methods (detailed in the first two papers) to the case of cardiac arrest prediction in a pediatric intensive care unit. Our results showed that utilizing the trend analysis from the time series data elements reduced the number of classification errors by 73%. The area under the Receiver Operating Characteristic curve increased from a baseline of 87% to 98% by including the trend analysis. In addition to the performance measures, we were also able to demonstrate that adding raw time series data elements without their associated trend analyses improved classification accuracy as compared to the baseline multivariate model, but diminished classification accuracy as compared to when just the trend analysis features were added (ie, without adding the raw time series data elements). We believe this phenomenon was largely attributable to overfitting, which is known to increase as the ratio of candidate features to class examples rises. Furthermore, although we employed several feature reduction strategies to counteract the overfitting problem, they failed to improve the performance beyond that which was achieved by exclusion of the raw time series elements. Finally, our data demonstrated that pulse oximetry and systolic blood pressure readings tend to start diminishing about 10-20 minutes before an arrest, whereas heart rates tend to diminish rapidly less than 5 minutes before an arrest.

Atmospheric and seawater carbon dioxide at time series station DYFAMED

A time series of fCO2, SST, and fluorescence data was collected between 1995 and 1997 by a CARIOCA buoy moored at the DyFAMed station (Dynamique des Flux Atmospheriques en Mediterranée) located in the northwestern Mediterranean Sea. On seasonal timescales, the spring phytoplankton bloom decreases the surface water fCO2 to approximately 290 µatm, followed by summer heating and a strong increase in fCO2 to a maximum of approximately 510 µatm. While the DELTA fCO2 shows strong variations on seasonal timescales, the annual average air-sea disequilibrium is only 2 µatm. Temperature-normalized fCO2 shows a continued decrease in dissolved CO2 throughout the summer and fall at a rate of approximately 0.6 µatm/d. The calculated annual air-sea CO2 transfer rate is -0.10 to -0.15 moles CO2 m-2 y-1, with these low values reflecting the relatively weak wind speed regime and small annual air-sea fCO2 disequilibrium. Extrapolating this rate over the whole Mediterranean Sea would lead to a flux of approximately -3 * 10**12 to -4.5 * 10**12 grams C/y, in good agreement with other estimates. An analysis of the effects of sampling frequency on annual air-sea CO2 flux estimates showed that monthly sampling is adequate to resolve the annual CO2 flux to within approximately ±10 - 18% at this site. Annual flux estimates made using temperature-derived fCO2 based on the measured fCO2-SST correlations are in agreement with measurement-based calculations to within ± 7-10% (depending on the gas transfer parameterization used), and suggest that annual CO2 flux estimates may be reasonably well predicted in this region from satellite or model-derived SST and wind speed information.

Oxygen concentration at time series station DYNAPROC

Continuous measurements between 0 and 200 m depth were performed every 2 h over two separate periods of four days at a station in the open northwestern Mediterranean Sea (Dyfamed Station) during the Dynaproc cruise in May 1995. Estimates of the daily variations in profiles of temperature, partial pressure of CO2, oxygen, chlorophyll a and nutrients were obtained. The distributions of the various physical and chemical properties were clearly different during the two time series, which were separated by a period of 11 days during which a wind event occurred. The mean daily utilization or production due to biological processes of dissolved inorganic carbon (DIC), nitrate+nitrite and oxygen were calculated along isopycnals using a vertical diffusion model. Between the surface and about 20 m depth, DIC was consumed and O2 released during the two time series while the nitrate+nitrite concentrations as well as supplies were zero. After the wind event, the O2 : C : N ratios of consumption (or production) were, on average, near the Redfield ratios, but during the first time series, the C : N utilization ratio between 20 and 35 m was two to three times that of Redfield stoichiometry and the oxygen release was low. The integrated net community production (NCP) in terms of carbon was equivalent during the two time series, whereas the chlorophyll a biomass was twice as high, on average, during the first time series but did decrease. These results imply that the production systems were different during the two periods. The first time series corresponds to a period at the end of production, due to the nutrient depletion in the euphotic layer. The formation of degradation products of the living material in dissolved organic form is probably important as indicated by the high C : N utilization ratios. The second time series corresponds to a reactivation of the primary production due to the upward shift of nutrients after the wind event.

Land Surface Temperature continuous time series from SSM/I passive microwave radiometers, over northern latitudes (Latitudes > 45°N), during summer snow free period (2000-2011), with links to NetCDF files

Remote sensing instruments are key players to map land surface temperature (LST) at large temporal and spatial scales. In this paper, we present how we combine passive microwave and thermal infrared data to estimate LST during summer snow-free periods over northern high latitudes. The methodology is based on the SSM/I-SSMIS 37 GHz measurements at both vertical and horizontal polarizations on a 25 km × 25 km grid size. LST is retrieved from brightness temperatures introducing an empirical linear relationship between emissivities at both polarizations as described in Royer and Poirier (2010). This relationship is calibrated at pixel scale, using cloud-free independent LST data from MODIS instruments. The SSM/I-SSMIS and MODIS data are synchronized by fitting a diurnal cycle model built on skin temperature reanalysis provided by the European Centre for Medium-Range Weather Forecasts (ECMWF). The resulting temperature dataset is provided at 25 km scale and at an hourly time step during the ten-year analysis period (2000-2011). This new product was locally evaluated at five experimental sites of the EU-PAGE21 project against air temperature measurements and meteorological model reanalysis, and compared to the MODIS LST product at both local and circumpolar scale. The results giving a mean RMSE of the order of 2.2 K demonstrate the usefulness of the microwave product, which is unaffected by clouds as opposed to thermal infrared products and offers a better resolution compared to model reanalysis.

Environmental variables modeling based on remote sensing data using time series analysis: forestry and agricultural applications

En la actualidad, el seguimiento de la dinámica de los procesos medio ambientales está considerado como un punto de gran interés en el campo medioambiental. La cobertura espacio temporal de los datos de teledetección proporciona información continua con una alta frecuencia temporal, permitiendo el análisis de la evolución de los ecosistemas desde diferentes escalas espacio-temporales. Aunque el valor de la teledetección ha sido ampliamente probado, en la actualidad solo existe un número reducido de metodologías que permiten su análisis de una forma cuantitativa. En la presente tesis se propone un esquema de trabajo para explotar las series temporales de datos de teledetección, basado en la combinación del análisis estadístico de series de tiempo y la fenometría. El objetivo principal es demostrar el uso de las series temporales de datos de teledetección para analizar la dinámica de variables medio ambientales de una forma cuantitativa. Los objetivos específicos son: (1) evaluar dichas variables medio ambientales y (2) desarrollar modelos empíricos para predecir su comportamiento futuro. Estos objetivos se materializan en cuatro aplicaciones cuyos objetivos específicos son: (1) evaluar y cartografiar estados fenológicos del cultivo del algodón mediante análisis espectral y fenometría, (2) evaluar y modelizar la estacionalidad de incendios forestales en dos regiones bioclimáticas mediante modelos dinámicos, (3) predecir el riesgo de incendios forestales a nivel pixel utilizando modelos dinámicos y (4) evaluar el funcionamiento de la vegetación en base a la autocorrelación temporal y la fenometría. Los resultados de esta tesis muestran la utilidad del ajuste de funciones para modelizar los índices espectrales AS1 y AS2. Los parámetros fenológicos derivados del ajuste de funciones permiten la identificación de distintos estados fenológicos del cultivo del algodón. El análisis espectral ha demostrado, de una forma cuantitativa, la presencia de un ciclo en el índice AS2 y de dos ciclos en el AS1 así como el comportamiento unimodal y bimodal de la estacionalidad de incendios en las regiones mediterránea y templada respectivamente. Modelos autorregresivos han sido utilizados para caracterizar la dinámica de la estacionalidad de incendios y para predecir de una forma muy precisa el riesgo de incendios forestales a nivel pixel. Ha sido demostrada la utilidad de la autocorrelación temporal para definir y caracterizar el funcionamiento de la vegetación a nivel pixel. Finalmente el concepto “Optical Functional Type” ha sido definido, donde se propone que los pixeles deberían ser considerados como unidades temporales y analizados en función de su dinámica temporal. ix SUMMARY A good understanding of land surface processes is considered as a key subject in environmental sciences. The spatial-temporal coverage of remote sensing data provides continuous observations with a high temporal frequency allowing the assessment of ecosystem evolution at different temporal and spatial scales. Although the value of remote sensing time series has been firmly proved, only few time series methods have been developed for analyzing this data in a quantitative and continuous manner. In the present dissertation a working framework to exploit Remote Sensing time series is proposed based on the combination of Time Series Analysis and phenometric approach. The main goal is to demonstrate the use of remote sensing time series to analyze quantitatively environmental variable dynamics. The specific objectives are (1) to assess environmental variables based on remote sensing time series and (2) to develop empirical models to forecast environmental variables. These objectives have been achieved in four applications which specific objectives are (1) assessing and mapping cotton crop phenological stages using spectral and phenometric analyses, (2) assessing and modeling fire seasonality in two different ecoregions by dynamic models, (3) forecasting forest fire risk on a pixel basis by dynamic models, and (4) assessing vegetation functioning based on temporal autocorrelation and phenometric analysis. The results of this dissertation show the usefulness of function fitting procedures to model AS1 and AS2. Phenometrics derived from function fitting procedure makes it possible to identify cotton crop phenological stages. Spectral analysis has demonstrated quantitatively the presence of one cycle in AS2 and two in AS1 and the unimodal and bimodal behaviour of fire seasonality in the Mediterranean and temperate ecoregions respectively. Autoregressive models has been used to characterize the dynamics of fire seasonality in two ecoregions and to forecasts accurately fire risk on a pixel basis. The usefulness of temporal autocorrelation to define and characterized land surface functioning has been demonstrated. And finally the “Optical Functional Types” concept has been proposed, in this approach pixels could be as temporal unities based on its temporal dynamics or functioning.

A new algorithm to accelerate harmonic analysis of time series

The Lomb periodogram has been traditionally a tool that allows us to elucidate if a frequency turns out to be important for explaining the behaviour of a given time series. Many linear and nonlinear reiterative harmonic processes that are used for studying the spectral content of a time series take into account this periodogram in order to avoid including spurious frequencies in their models due to the leakage problem of energy from one frequency to others. However, the estimation of the periodogram requires long computation time that makes the harmonic analysis slower when we deal with certain time series. Here we propose an algorithm that accelerates the extraction of the most remarkable frequencies from the periodogram, avoiding its whole estimation of the harmonic process at each iteration. This algorithm allows the user to perform a specific analysis of a given scalar time series. As a result, we obtain a functional model made of (1) a trend component, (2) a linear combination of Fourier terms, and (3) the so-called mixed secular terms by reducing the computation time of the estimation of the periodogram.

Detection and mapping of burnt areas from time series of MODIS-derived NDVI data in a Mediterranean region

Moderate resolution remote sensing data, as provided by MODIS, can be used to detect and map active or past wildfires from daily records of suitable combinations of reflectance bands. The objective of the present work was to develop and test simple algorithms and variations for automatic or semiautomatic detection of burnt areas from time series data of MODIS biweekly vegetation indices for a Mediterranean region. MODIS-derived NDVI 250m time series data for the Valencia region, East Spain, were subjected to a two-step process for the detection of candidate burnt areas, and the results compared with available fire event records from the Valencia Regional Government. For each pixel and date in the data series, a model was fitted to both the previous and posterior time series data. Combining drops between two consecutive points and 1-year average drops, we used discrepancies or jumps between the pre and post models to identify seed pixels, and then delimitated fire scars for each potential wildfire using an extension algorithm from the seed pixels. The resulting maps of the detected burnt areas showed a very good agreement with the perimeters registered in the database of fire records used as reference. Overall accuracies and indices of agreement were very high, and omission and commission errors were similar or lower than in previous studies that used automatic or semiautomatic fire scar detection based on remote sensing. This supports the effectiveness of the method for detecting and mapping burnt areas in the Mediterranean region.

Use of an Inverse Method for Time Series to Estimate the Dynamics of and Management Strategies for the Box Jellyfish Carybdea marsupialis

Frequently, population ecology of marine organisms uses a descriptive approach in which their sizes and densities are plotted over time. This approach has limited usefulness for design strategies in management or modelling different scenarios. Population projection matrix models are among the most widely used tools in ecology. Unfortunately, for the majority of pelagic marine organisms, it is difficult to mark individuals and follow them over time to determine their vital rates and built a population projection matrix model. Nevertheless, it is possible to get time-series data to calculate size structure and densities of each size, in order to determine the matrix parameters. This approach is known as a “demographic inverse problem” and it is based on quadratic programming methods, but it has rarely been used on aquatic organisms. We used unpublished field data of a population of cubomedusae Carybdea marsupialis to construct a population projection matrix model and compare two different management strategies to lower population to values before year 2008 when there was no significant interaction with bathers. Those strategies were by direct removal of medusae and by reducing prey. Our results showed that removal of jellyfish from all size classes was more effective than removing only juveniles or adults. When reducing prey, the highest efficiency to lower the C. marsupialis population occurred when prey depletion affected prey of all medusae sizes. Our model fit well with the field data and may serve to design an efficient management strategy or build hypothetical scenarios such as removal of individuals or reducing prey. TThis This sdfsdshis method is applicable to other marine or terrestrial species, for which density and population structure over time are available.

Evaluation of the potential of InSAR time series to study the spatio-temporal evolution of piezometric levels in the Madrid aquifer

The Tertiary detritic aquifer of Madrid (TDAM), with an average thickness of 1500 m and a heterogeneous, anisotropic structure, supplies water to Madrid, the most populated city of Spain (3.2 million inhabitants in the metropolitan area). Besides its complex structure, a previous work focused in the north-northwest of Madrid city showed that the aquifer behaves quasi elastically trough extraction/recovery cycles and ground uplifting during recovery periods compensates most of the ground subsidence measured during previous extraction periods (Ezquerro et al., 2014). Therefore, the relationship between ground deformation and groundwater level through time can be simulated using simple elastic models. In this work, we model the temporal evolution of the piezometric level in 19 wells of the TDAM in the period 1997–2010. Using InSAR and piezometric time series spanning the studied period, we first estimate the elastic storage coefficient (Ske) for every well. Both, the Ske of each well and the average Ske of all wells, are used to predict hydraulic heads at the different well locations during the study period and compared against the measured hydraulic heads, leading to very similar errors when using the Ske of each well and the average Ske of all wells: 14 and 16 % on average respectively. This result suggests that an average Ske can be used to estimate piezometric level variations in all the points where ground deformation has been measured by InSAR, thus allowing production of piezometric level maps for the different extraction/recovery cycles in the TDAM.

Biological and environmental time series obtained from Point B, Bay of Villefranche

Ecological succession provides a widely accepted description of seasonal changes in phytoplankton and mesozooplankton assemblages in the natural environment, but concurrent changes in smaller (i.e. microbes) and larger (i.e. macroplankton) organisms are not included in the model because plankton ranging from bacteria to jellies are seldom sampled and analyzed simultaneously. Here we studied, for the first time in the aquatic literature, the succession of marine plankton in the whole-plankton assemblage that spanned 5 orders of magnitude in size from microbes to macroplankton predators (not including fish or fish larvae, for which no consistent data were available). Samples were collected in the northwestern Mediterranean Sea (Bay of Villefranche) weekly during 10 months. Simultaneously collected samples were analyzed by flow cytometry, inverse microscopy, FlowCam, and ZooScan. The whole-plankton assemblage underwent sharp reorganizations that corresponded to bottom-up events of vertical mixing in the water-column, and its development was top-down controlled by large gelatinous filter feeders and predators. Based on the results provided by our novel whole-plankton assemblage approach, we propose a new comprehensive conceptual model of the annual plankton succession (i.e. whole plankton model) characterized by both stepwise stacking of four broad trophic communities from early spring through summer, which is a new concept, and progressive replacement of ecological plankton categories within the different trophic communities, as recognised traditionally.

Quantifying ice-shelf basal melt and ice-stream dynamics using high-resolution DEM and GPS time series

Thesis (Ph.D.)--University of Washington, 2016-06

Genetic time-series analysis identifies a major QTL for in vivo alcohol metabolism not predicted by in vitro studies of structural protein polymorphism at the ADH1B or ADH1C loci

After ingestion of a standardized dose of ethanol, alcohol concentrations were assessed, over 3.5 hours from blood (six readings) and breath (10 readings) in a sample of 412 MZ and DZ twins who took part in an Alcohol Challenge Twin Study (ACTS). Nearly all participants were subsequently genotyped on two polymorphic SNPs in the ADH1B and ADH1C loci known to affect in vitro ADH activity. In the DZ pairs, 14 microsatellite markers covering a 20.5 cM region on chromosome 4 that includes the ADH gene family were assessed, Variation in the timed series of autocorrelated blood and breath alcohol readings was studied using a bivariate simplex design. The contribution of a quantitative trait locus (QTL) or QTL's linked to the ADH region was estimated via a mixture of likelihoods weighted by identity-by-descent probabilities. The effects of allelic substitution at the ADH1B and ADH1C loci were estimated in the means part of the model simultaneously with the effects sex and age. There was a major contribution to variance in alcohol metabolism due to a QTL which accounted for about 64% of the additive genetic covariation common to both blood and breath alcohol readings at the first time point. No effects of the ADH1B*47His or ADH1C*349Ile alleles on in vivo metabolism were observed, although these have been shown to have major effects in vitro. This implies that there is a major determinant of variation for in vivo alcohol metabolism in the ADH region that is not accounted for by these polymorphisms. Earlier analyses of these data suggested that alcohol metabolism is related to drinking behavior and imply that this QTL may be protective against alcohol dependence.