Refine
Year of publication
Document Type
- Article (19)
- Postprint (10)
- Doctoral Thesis (9)
- Conference Proceeding (1)
- Review (1)
Keywords
- prediction (40) (remove)
Institute
- Department Psychologie (5)
- Institut für Biochemie und Biologie (5)
- Institut für Physik und Astronomie (4)
- Hasso-Plattner-Institut für Digital Engineering gGmbH (3)
- Institut für Umweltwissenschaften und Geographie (3)
- Mathematisch-Naturwissenschaftliche Fakultät (3)
- Department Linguistik (2)
- Hochschulambulanz (2)
- Institut für Ernährungswissenschaft (2)
- Institut für Mathematik (2)
Based on an analysis of continuous monitoring of farm animal behavior in the region of the 2016 M6.6 Norcia earthquake in Italy, Wikelski et al., 2020; (Seismol Res Lett, 89, 2020, 1238) conclude that animal activity can be anticipated with subsequent seismic activity and that this finding might help to design a "short-term earthquake forecasting method." We show that this result is based on an incomplete analysis and misleading interpretations. Applying state-of-the-art methods of statistics, we demonstrate that the proposed anticipatory patterns cannot be distinguished from random patterns, and consequently, the observed anomalies in animal activity do not have any forecasting power.
We investigate spatio-temporal properties of earthquake patterns in the San Jacinto fault zone (SJFZ), California, between Cajon Pass and the Superstition Hill Fault, using a long record of simulated seismicity constrained by available seismological and geological data. The model provides an effective realization of a large segmented strike-slip fault zone in a 3D elastic half-space, with heterogeneous distribution of static friction chosen to represent several clear step-overs at the surface. The simulated synthetic catalog reproduces well the basic statistical features of the instrumental seismicity recorded at the SJFZ area since 1981. The model also produces events larger than those included in the short instrumental record, consistent with paleo-earthquakes documented at sites along the SJFZ for the last 1,400 years. The general agreement between the synthetic and observed data allows us to address with the long-simulated seismicity questions related to large earthquakes and expected seismic hazard. The interaction between m a parts per thousand yen 7 events on different sections of the SJFZ is found to be close to random. The hazard associated with m a parts per thousand yen 7 events on the SJFZ increases significantly if the long record of simulated seismicity is taken into account. The model simulations indicate that the recent increased number of observed intermediate SJFZ earthquakes is a robust statistical feature heralding the occurrence of m a parts per thousand yen 7 earthquakes. The hypocenters of the m a parts per thousand yen 5 events in the simulation results move progressively towards the hypocenter of the upcoming m a parts per thousand yen 7 earthquake.
Species respond to environmental change by dynamically adjusting their geographical ranges. Robust predictions of these changes are prerequisites to inform dynamic and sustainable conservation strategies. Correlative species distribution models (SDMs) relate species’ occurrence records to prevailing environmental factors to describe the environmental niche. They have been widely applied in global change context as they have comparably low data requirements and allow for rapid assessments of potential future species’ distributions. However, due to their static nature, transient responses to environmental change are essentially ignored in SDMs. Furthermore, neither dispersal nor demographic processes and biotic interactions are explicitly incorporated. Therefore, it has often been suggested to link statistical and mechanistic modelling approaches in order to make more realistic predictions of species’ distributions for scenarios of environmental change. In this thesis, I present two different ways of such linkage. (i) Mechanistic modelling can act as virtual playground for testing statistical models and allows extensive exploration of specific questions. I promote this ‘virtual ecologist’ approach as a powerful evaluation framework for testing sampling protocols, analyses and modelling tools. Also, I employ such an approach to systematically assess the effects of transient dynamics and ecological properties and processes on the prediction accuracy of SDMs for climate change projections. That way, relevant mechanisms are identified that shape the species’ response to altered environmental conditions and which should hence be considered when trying to project species’ distribution through time. (ii) I supplement SDM projections of potential future habitat for black grouse in Switzerland with an individual-based population model. By explicitly considering complex interactions between habitat availability and demographic processes, this allows for a more direct assessment of expected population response to environmental change and associated extinction risks. However, predictions were highly variable across simulations emphasising the need for principal evaluation tools like sensitivity analysis to assess uncertainty and robustness in dynamic range predictions. Furthermore, I identify data coverage of the environmental niche as a likely cause for contrasted range predictions between SDM algorithms. SDMs may fail to make reliable predictions for truncated and edge niches, meaning that portions of the niche are not represented in the data or niche edges coincide with data limits. Overall, my thesis contributes to an improved understanding of uncertainty factors in predictions of range dynamics and presents ways how to deal with these. Finally I provide preliminary guidelines for predictive modelling of dynamic species’ response to environmental change, identify key challenges for future research and discuss emerging developments.
Background:
COVID-19 has infected millions of people worldwide and is responsible for several hundred thousand fatalities. The COVID-19 pandemic has necessitated thoughtful resource allocation and early identification of high-risk patients. However, effective methods to meet these needs are lacking.
Objective:
The aims of this study were to analyze the electronic health records (EHRs) of patients who tested positive for COVID-19 and were admitted to hospitals in the Mount Sinai Health System in New York City; to develop machine learning models for making predictions about the hospital course of the patients over clinically meaningful time horizons based on patient characteristics at admission; and to assess the performance of these models at multiple hospitals and time points.
Methods:
We used Extreme Gradient Boosting (XGBoost) and baseline comparator models to predict in-hospital mortality and critical events at time windows of 3, 5, 7, and 10 days from admission. Our study population included harmonized EHR data from five hospitals in New York City for 4098 COVID-19-positive patients admitted from March 15 to May 22, 2020. The models were first trained on patients from a single hospital (n=1514) before or on May 1, externally validated on patients from four other hospitals (n=2201) before or on May 1, and prospectively validated on all patients after May 1 (n=383). Finally, we established model interpretability to identify and rank variables that drive model predictions.
Results:
Upon cross-validation, the XGBoost classifier outperformed baseline models, with an area under the receiver operating characteristic curve (AUC-ROC) for mortality of 0.89 at 3 days, 0.85 at 5 and 7 days, and 0.84 at 10 days. XGBoost also performed well for critical event prediction, with an AUC-ROC of 0.80 at 3 days, 0.79 at 5 days, 0.80 at 7 days, and 0.81 at 10 days. In external validation, XGBoost achieved an AUC-ROC of 0.88 at 3 days, 0.86 at 5 days, 0.86 at 7 days, and 0.84 at 10 days for mortality prediction. Similarly, the unimputed XGBoost model achieved an AUC-ROC of 0.78 at 3 days, 0.79 at 5 days, 0.80 at 7 days, and 0.81 at 10 days. Trends in performance on prospective validation sets were similar. At 7 days, acute kidney injury on admission, elevated LDH, tachypnea, and hyperglycemia were the strongest drivers of critical event prediction, while higher age, anion gap, and C-reactive protein were the strongest drivers of mortality prediction.
Conclusions:
We externally and prospectively trained and validated machine learning models for mortality and critical events for patients with COVID-19 at different time horizons. These models identified at-risk patients and uncovered underlying relationships that predicted outcomes.
Background and objectives
AKI treated with dialysis initiation is a common complication of coronavirus disease 2019 (COVID-19) among hospitalized patients. However, dialysis supplies and personnel are often limited.
Design, setting, participants, & measurements
Using data from adult patients hospitalized with COVID-19 from five hospitals from theMount Sinai Health System who were admitted between March 10 and December 26, 2020, we developed and validated several models (logistic regression, Least Absolute Shrinkage and Selection Operator (LASSO), random forest, and eXtreme GradientBoosting [XGBoost; with and without imputation]) for predicting treatment with dialysis or death at various time horizons (1, 3, 5, and 7 days) after hospital admission. Patients admitted to theMount Sinai Hospital were used for internal validation, whereas the other hospitals formed part of the external validation cohort. Features included demographics, comorbidities, and laboratory and vital signs within 12 hours of hospital admission.
Results
A total of 6093 patients (2442 in training and 3651 in external validation) were included in the final cohort. Of the different modeling approaches used, XGBoost without imputation had the highest area under the receiver operating characteristic (AUROC) curve on internal validation (range of 0.93-0.98) and area under the precisionrecall curve (AUPRC; range of 0.78-0.82) for all time points. XGBoost without imputation also had the highest test parameters on external validation (AUROC range of 0.85-0.87, and AUPRC range of 0.27-0.54) across all time windows. XGBoost without imputation outperformed all models with higher precision and recall (mean difference in AUROC of 0.04; mean difference in AUPRC of 0.15). Features of creatinine, BUN, and red cell distribution width were major drivers of the model's prediction.
Conclusions
An XGBoost model without imputation for prediction of a composite outcome of either death or dialysis in patients positive for COVID-19 had the best performance, as compared with standard and other machine learning models.
A large body of research now supports the presence of both syntactic and lexical predictions in sentence processing. Lexical predictions, in particular, are considered to indicate a deep level of predictive processing that extends past the structural features of a necessary word (e.g. noun), right down to the phonological features of the lexical identity of a specific word (e.g. /kite/; DeLong et al., 2005). However, evidence for lexical predictions typically focuses on predictions in very local environments, such as the adjacent word or words (DeLong et al., 2005; Van Berkum et al., 2005; Wicha et al., 2004). Predictions in such local environments may be indistinguishable from lexical priming, which is transient and uncontrolled, and as such may prime lexical items that are not compatible with the context (e.g. Kukona et al., 2014). Predictive processing has been argued to be a controlled process, with top-down information guiding preactivation of plausible upcoming lexical items (Kuperberg & Jaeger, 2016). One way to distinguish lexical priming from prediction is to demonstrate that preactivated lexical content can be maintained over longer distances.
In this dissertation, separable German particle verbs are used to demonstrate that preactivation of lexical items can be maintained over multi-word distances. A self-paced reading time and an eye tracking experiment provide some support for the idea that particle preactivation triggered by a verb and its context can be observed by holding the sentence context constant and manipulating the predictabilty of the particle. Although evidence of an effect of particle predictability was only seen in eye tracking, this is consistent with previous evidence suggesting that predictive processing facilitates only some eye tracking measures to which the self-paced reading modality may not be sensitive (Staub, 2015; Rayner1998). Interestingly, manipulating the distance between the verb and the particle did not affect reading times, suggesting that the surprisal-predicted faster reading times at long distance may only occur when the additional distance is created by information that adds information about the lexical identity of a distant element (Levy, 2008; Grodner & Gibson, 2005). Furthermore, the results provide support for models proposing that temporal decay is not major influence on word processing (Lewandowsky et al., 2009; Vasishth et al., 2019).
In the third and fourth experiments, event-related potentials were used as a method for detecting specific lexical predictions. In the initial ERP experiment, we found some support for the presence of lexical predictions when the sentence context constrained the number of plausible particles to a single particle. This was suggested by a frontal post-N400 positivity (PNP) that was elicited when a lexical prediction had been violated, but not to violations when more than one particle had been plausible. The results of this study were highly consistent with previous research suggesting that the PNP might be a much sought-after ERP marker of prediction failure (DeLong et al., 2011; DeLong et al., 2014; Van Petten & Luka, 2012; Thornhill & Van Petten, 2012; Kuperberg et al., 2019). However, a second experiment in a larger sample experiment failed to replicate the effect, but did suggest the relationship of the PNP to predictive processing may not yet be fully understood. Evidence for long-distance lexical predictions was inconclusive.
The conclusion drawn from the four experiments is that preactivation of the lexical entries of plausible upcoming particles did occur and was maintained over long distances. The facilitatory effect of this preactivation at the particle site therefore did not appear to be the result of transient lexical priming. However, the question of whether this preactivation can also lead to lexical predictions of a specific particle remains unanswered. Of particular interest to future research on predictive processing is further characterisation of the PNP. Implications for models of sentence processing may be the inclusion of long-distance lexical predictions, or the possibility that preactivation of lexical material can facilitate reading times and ERP amplitude without commitment to a specific lexical item.
Ecologists urgently need a better ability to predict how environmental change affects biodiversity. We examine individual-based ecology (IBE), a research paradigm that promises better a predictive ability by using individual-based models (IBMs) to represent ecological dynamics as arising from how individuals interact with their environment and with each other. A key advantage of IBMs is that the basis for predictions-fitness maximization by individual organisms-is more general and reliable than the empirical relationships that other models depend on. Case studies illustrate the usefulness and predictive success of long-term IBE programs. The pioneering programs had three phases: conceptualization, implementation, and diversification. Continued validation of models runs throughout these phases. The breakthroughs that make IBE more productive include standards for describing and validating IBMs, improved and standardized theory for individual traits and behavior, software tools, and generalized instead of system-specific IBMs. We provide guidelines for pursuing IBE and a vision for future IBE research.
Predicting the actions of other individuals is crucial for our daily interactions. Recent evidence suggests that the prediction of object-directed arm and full-body actions employs the dorsal premotor cortex (PMd). Thus, the neural substrate involved in action control may also be essential for action prediction. Here, we aimed to address this issue and hypothesized that disrupting the PMd impairs action prediction. Using fMRI-guided coil navigation, rTMS (five pulses, 10Hz) was applied over the left PMd and over the vertex (control region) while participants observed everyday actions in video clips that were transiently occluded for 1s. The participants detected manipulations in the time course of occluded actions, which required them to internally predict the actions during occlusion. To differentiate between functional roles that the PMd could play in prediction, rTMS was either delivered at occluder-onset (TMS-early), affecting the initiation of action prediction, or 300 ms later during occlusion(TMS-late), affecting the maintenance of anongoing prediction. TMS-early over the left PMd produced more prediction errors than TMS-early over the vertex. TMS-late had no effect on prediction performance, suggesting that the left PMd might be involved particularly during the initiation of internally guided action prediction but may play a subordinate role in maintaining ongoing prediction. These findings open a new perspective on the role of the left PMd in action prediction which is in line with its functions in action control and in cognitive tasks. In the discussion, there levance of the left PMd for integrating external action parameters with the observer's motor repertoire is emphasized. Overall, the results are in line with the notion that premotor functions are employed in both action control and action observation.
Predicting Paris: Multi-Method Approaches to Forecast the Outcomes of Global Climate Negotiations
(2016)
We examine the negotiations held under the auspices of the United Nations Framework Convention of Climate Change in Paris, December 2015. Prior to these negotiations, there was considerable uncertainty about whether an agreement would be reached, particularly given that the world’s leaders failed to do so in the 2009 negotiations held in Copenhagen. Amid this uncertainty, we applied three different methods to predict the outcomes: an expert survey and two negotiation simulation models, namely the Exchange Model and the Predictioneer’s Game. After the event, these predictions were assessed against the coded texts that were agreed in Paris. The evidence suggests that combining experts’ predictions to reach a collective expert prediction makes for significantly more accurate predictions than individual experts’ predictions. The differences in the performance between the two different negotiation simulation models were not statistically significant.
The Limpopo Basin in southern Africa is prone to droughts which affect the livelihood of millions of people in South Africa, Botswana, Zimbabwe and Mozambique. Seasonal drought early warning is thus vital for the whole region. In this study, the predictability of hydrological droughts during the main runoff period from December to May is assessed using statistical approaches. Three methods (multiple linear models, artificial neural networks, random forest regression trees) are compared in terms of their ability to forecast streamflow with up to 12 months of lead time. The following four main findings result from the study.
1. There are stations in the basin at which standardised streamflow is predictable with lead times up to 12 months. The results show high inter-station differences of forecast skill but reach a coefficient of determination as high as 0.73 (cross validated).
2. A large range of potential predictors is considered in this study, comprising well-established climate indices, customised teleconnection indices derived from sea surface temperatures and antecedent streamflow as a proxy of catchment conditions. El Nino and customised indices, representing sea surface temperature in the Atlantic and Indian oceans, prove to be important teleconnection predictors for the region. Antecedent streamflow is a strong predictor in small catchments (with median 42% explained variance), whereas teleconnections exert a stronger influence in large catchments.
3. Multiple linear models show the best forecast skill in this study and the greatest robustness compared to artificial neural networks and random forest regression trees, despite their capabilities to represent nonlinear relationships.
4. Employed in early warning, the models can be used to forecast a specific drought level. Even if the coefficient of determination is low, the forecast models have a skill better than a climatological forecast, which is shown by analysis of receiver operating characteristics (ROCs). Seasonal statistical forecasts in the Limpopo show promising results, and thus it is recommended to employ them as complementary to existing forecasts in order to strengthen preparedness for droughts.