Complex Predictive Models Research Articles

Missing Data Imputation for Multisite Rainfall Networks: A Comparison between Geostatistical Interpolation and Pattern-Based Estimation on Different Terrain Types

AbstractMissing rainfall data are a major limitation for distributed hydrological modeling and climate studies. Practitioners need reliable approaches that can be employed on a daily basis, often with too limited data in space to feed complex predictive models. In this study we compare different automatic approaches for missing data imputation, including geostatistical interpolation and pattern-based estimation algorithms. We introduce two pattern-based approaches based on the analysis of historical data patterns: (i) an iterative version of K-nearest neighbor (IKNN) and (ii) a new algorithm called vector sampling (VS) that combines concepts of multiple-point statistics and resampling. Both algorithms can draw estimations from variably incomplete data patterns, allowing the target dataset to be at the same time the training dataset. Tested on five case studies from Denmark, Australia, and Switzerland, the algorithms show a different performance that seems to be related to the terrain type: on flat terrains with spatially homogeneous rain events, geostatistical interpolation tends to minimize the average error, while in mountainous regions with nonstationary rainfall statistics, data mining can recover better the rainfall patterns. The VS algorithm, requiring minimal parameterization, turns out to be a convenient option for routine application on complex and poorly gauged terrains.

Multilevel-Modeling Interpretation of Trailing-Edge Noise Models for Wind Turbines with NACA 0012 Airfoil

Recently, good interest in renewable energy has been increased and thereby the need for noise prediction model is required for the design of wind turbines with less noise. In particular, trailing edge noise is a major noise source. A number of results using complex predictive models with machine learning techniques and artificial neural network have been reported, and the results are generally good, depending on the application. However, the rigorous verification of the model itself that produces such predictive results is very insufficient, and there is a lack of understanding of the process in which the results are produced, and an explanation of the various physical phenomena observed in the experiment is not easy. In this study, Brooks, Pope, and Marcolini (BPM) model, the most popular semi-empirical model for airfoil noise prediction, was subjected to an analysis using multilevel modeling. The multilevel model is a statistical model containing both fixed effects and random effects. We were able to reckon the interpretation of two issues that are left unexplained in BPM model and to improve our understanding of the noise phenomenon of wind turbines. With further research, multilevel modeling is expected to be an effective tool for better design and analysis of complex predictive models for wind turbines noise.

Improving Clinical Translation of Machine Learning Approaches Through Clinician-Tailored Visual Displays of Black Box Algorithms: Development and Validation.

BackgroundDespite the promise of machine learning (ML) to inform individualized medical care, the clinical utility of ML in medicine has been limited by the minimal interpretability and black box nature of these algorithms.ObjectiveThe study aimed to demonstrate a general and simple framework for generating clinically relevant and interpretable visualizations of black box predictions to aid in the clinical translation of ML.MethodsTo obtain improved transparency of ML, simplified models and visual displays can be generated using common methods from clinical practice such as decision trees and effect plots. We illustrated the approach based on postprocessing of ML predictions, in this case random forest predictions, and applied the method to data from the Left Ventricular (LV) Structural Predictors of Sudden Cardiac Death (SCD) Registry for individualized risk prediction of SCD, a leading cause of death.ResultsWith the LV Structural Predictors of SCD Registry data, SCD risk predictions are obtained from a random forest algorithm that identifies the most important predictors, nonlinearities, and interactions among a large number of variables while naturally accounting for missing data. The black box predictions are postprocessed using classification and regression trees into a clinically relevant and interpretable visualization. The method also quantifies the relative importance of an individual or a combination of predictors. Several risk factors (heart failure hospitalization, cardiac magnetic resonance imaging indices, and serum concentration of systemic inflammation) can be clearly visualized as branch points of a decision tree to discriminate between low-, intermediate-, and high-risk patients.ConclusionsThrough a clinically important example, we illustrate a general and simple approach to increase the clinical translation of ML through clinician-tailored visual displays of results from black box algorithms. We illustrate this general model-agnostic framework by applying it to SCD risk prediction. Although we illustrate the methods using SCD prediction with random forest, the methods presented are applicable more broadly to improving the clinical translation of ML, regardless of the specific ML algorithm or clinical application. As any trained predictive model can be summarized in this manner to a prespecified level of precision, we encourage the use of simplified visual displays as an adjunct to the complex predictive model. Overall, this framework can allow clinicians to peek inside the black box and develop a deeper understanding of the most important features from a model to gain trust in the predictions and confidence in applying them to clinical care.

DALEX: Explainers for Complex Predictive Models in R

DALEX: Explainers for Complex Predictive Models in R

Геоинформационные системы в популяционном анализе распространения депрессивных расстройств в Хабаровске

The research is devoted to the use of modern geoinformation technologies for the analysis of spatial medical and demographic data. On the example of the medical and ecological geoinformation system (MEGIS) of Khabarovsk the possibilities of geoinformation technologies in the study of the spread of depressive disorders in a large city and the analysis of cause-and-effect relationships between this disease and some socio-economic factors are shown. The features of mathematical support of MEGIS necessary for population epidemiological analysis are considered. The possibilities of using correlation-regression and cluster-discriminant analysis for these purposes are shown. At the stage of the initial manifestation of symptoms of depression, statistically significant binary risk factors for depressive disorders were established, which were used in the diagnostic questionnaire. Developed complex recognition, classification and predictive models on the individual level, to assess the risk of developing depressive disorders and to predict the potential severity of the disease. At the population level, risk groups for depressive disorders in patients who have suffered depressive episodes in childhood should be formed. Based on the results of multi-level GIS and classification and predictive modeling based on individual clinical dynamically, socio-psychological, transcultural and environmental health risk factors formulated practical recommendations to improve the prevention of depressive disorders. The study of clinical and pathodynamic, socio-psychological, transcultural and medico-ecological aspects of non-psychotic depressive disorders with the use of geoinformation systems was carried out for the prevention of depression and mental health stabilization of the region population. The results of this research can be used by specialists in geoinformation systems, medical demography, medical psychology and social psychiatry, sociology, psychiatry, psychology and family medicine, conflictology, information modeling and system analysis in health care.

Child‐related attributions of hostile intent and harsh discipline: Moderating effects of anger

Two studies tested whether attributions of hostile intent (AHI) and anger for ambiguous child behaviors uniquely predict parents' harsh verbal discipline (HVD) and harsh physical discipline (HPD) or whether AHI only predicts harsh discipline when AHI occurs in conjunction with anger, as suggested by the Integrated Cognitive Model of general aggression (Wilkowski & Robinson, 2010, J. Pers., 78, 9-38). Replicating previous studies, the expected positive bivariate relationships between AHI, anger, HVD, and HPD were found in Study 1 (N = 493). However, when AHI, anger, and the AHI by anger interaction were considered simultaneously, patterns of prediction varied by discipline type. For HVD, high levels of AHI, high levels of anger, and the AHI by anger interaction (at high levels of AHI and anger) each were significant predictors. For HPD, low levels of AHI, high levels of anger, and the AHI by anger interaction (at high levels of AHI and anger) were significant predictors, suggesting a possible suppressor effect for AHI. These results were replicated in Study 2 ( N = 503). Thus, there was support for the contention that AHI only predicts harsh discipline when AHI occurs in conjunction with anger with respect to HPD but this was not the case for HVD. Since AHI suppressor effects were observed for HPD in two large sample studies, variables that may account for the AHI suppressor effects and the need for testing more complex interactive predictive models of harsh parenting behaviors are discussed.

Cardiovascular disease risk prediction using automated machine learning: A prospective study of 423,604 UK Biobank participants.

BackgroundIdentifying people at risk of cardiovascular diseases (CVD) is a cornerstone of preventative cardiology. Risk prediction models currently recommended by clinical guidelines are typically based on a limited number of predictors with sub-optimal performance across all patient groups. Data-driven techniques based on machine learning (ML) might improve the performance of risk predictions by agnostically discovering novel risk predictors and learning the complex interactions between them. We tested (1) whether ML techniques based on a state-of-the-art automated ML framework (AutoPrognosis) could improve CVD risk prediction compared to traditional approaches, and (2) whether considering non-traditional variables could increase the accuracy of CVD risk predictions.Methods and findingsUsing data on 423,604 participants without CVD at baseline in UK Biobank, we developed a ML-based model for predicting CVD risk based on 473 available variables. Our ML-based model was derived using AutoPrognosis, an algorithmic tool that automatically selects and tunes ensembles of ML modeling pipelines (comprising data imputation, feature processing, classification and calibration algorithms). We compared our model with a well-established risk prediction algorithm based on conventional CVD risk factors (Framingham score), a Cox proportional hazards (PH) model based on familiar risk factors (i.e, age, gender, smoking status, systolic blood pressure, history of diabetes, reception of treatments for hypertension and body mass index), and a Cox PH model based on all of the 473 available variables. Predictive performances were assessed using area under the receiver operating characteristic curve (AUC-ROC). Overall, our AutoPrognosis model improved risk prediction (AUC-ROC: 0.774, 95% CI: 0.768-0.780) compared to Framingham score (AUC-ROC: 0.724, 95% CI: 0.720-0.728, p < 0.001), Cox PH model with conventional risk factors (AUC-ROC: 0.734, 95% CI: 0.729-0.739, p < 0.001), and Cox PH model with all UK Biobank variables (AUC-ROC: 0.758, 95% CI: 0.753-0.763, p < 0.001). Out of 4,801 CVD cases recorded within 5 years of baseline, AutoPrognosis was able to correctly predict 368 more cases compared to the Framingham score. Our AutoPrognosis model included predictors that are not usually considered in existing risk prediction models, such as the individuals’ usual walking pace and their self-reported overall health rating. Furthermore, our model improved risk prediction in potentially relevant sub-populations, such as in individuals with history of diabetes. We also highlight the relative benefits accrued from including more information into a predictive model (information gain) as compared to the benefits of using more complex models (modeling gain).ConclusionsOur AutoPrognosis model improves the accuracy of CVD risk prediction in the UK Biobank population. This approach performs well in traditionally poorly served patient subgroups. Additionally, AutoPrognosis uncovered novel predictors for CVD disease that may now be tested in prospective studies. We found that the “information gain” achieved by considering more risk factors in the predictive model was significantly higher than the “modeling gain” achieved by adopting complex predictive models.

Predicting scheduled hospital attendance with artificial intelligence

Failure to attend scheduled hospital appointments disrupts clinical management and consumes resource estimated at £1 billion annually in the United Kingdom National Health Service alone. Accurate stratification of absence risk can maximize the yield of preventative interventions. The wide multiplicity of potential causes, and the poor performance of systems based on simple, linear, low-dimensional models, suggests complex predictive models of attendance are needed. Here, we quantify the effect of using complex, non-linear, high-dimensional models enabled by machine learning. Models systematically varying in complexity based on logistic regression, support vector machines, random forests, AdaBoost, or gradient boosting machines were trained and evaluated on an unselected set of 22,318 consecutive scheduled magnetic resonance imaging appointments at two UCL hospitals. High-dimensional Gradient Boosting Machine-based models achieved the best performance reported in the literature, exhibiting an area under the receiver operating characteristic curve of 0.852 and average precision of 0.511. Optimal predictive performance required 81 variables. Simulations showed net potential benefit across a wide range of attendance characteristics, peaking at £3.15 per appointment at current prevalence and call efficiency. Optimal attendance prediction requires more complex models than have hitherto been applied in the field, reflecting the complex interplay of patient, environmental, and operational causal factors. Far from an exotic luxury, high-dimensional models based on machine learning are likely essential to optimal scheduling amongst other operational aspects of hospital care. High predictive performance is achievable with data from a single institution, obviating the need for aggregating large-scale sensitive data across governance boundaries.

A Path to Prediction of Outcomes in Juvenile Idiopathic Inflammatory Myopathy.

Humans have an innate desire to observe and subsequently dissect an event into component pieces in an effort to better characterize the event. We then examine these pieces individually and in combinations using this information to determine the outcome of future similar events and the likelihood of their recurrence. Practically, this attempt to foretell an occurrence and predict its outcomes is evident in multiple disciplines ranging from meteorology to sociologic studies. In this manuscript we share the historical and present-day tools to predict course and outcome in juvenile idiopathic inflammatory myopathy including clinical features, testing, and biomarkers. Further we discuss considerations for building more complex predictive models of outcome especially in diseases such as juvenile idiopathic inflammatory myopathy where patients numbers are low. Many of the barriers to developing risk prediction models for juvenile idiopathic inflammatory myopathy outcomes have improved with many remaining challenges being addressed.

Soft Matter Informatics: Current Progress and Challenges

AbstractThe number of applications of informatics or data‐driven discovery is growing in many fields, including materials science. The large amount of data that is readily available, combined with high‐level statistical algorithms, is proving to be extremely useful in developing complex predictive models with little to no human supervision or bias. However, in the field of soft matter, which includes complex materials such as polymers, liquids, emulsions, colloids, and gels, there is a slower adoption of informatics strategies than in adjacent fields. Here, the current state of soft matter informatics is discussed. Challenges specific to soft materials, including data classification, various degrees of organization at multiple length scales, and process‐dependent properties require unique approaches by researchers in order to develop robust informatics approaches in soft matter. The current ability to extract and analyze the information from the PoLyInfo database is demonstrated by the fitting of the Flory–Fox equation for glass transition temperature for several polymers. This Progress Report serves to introduce and excite the scientific community about the remarkable potential of informatics for exploring the properties of soft materials.

An energy-efficient internet of things (IoT) architecture for preventive conservation of cultural heritage

Internet of Things (IoT) technologies can facilitate the preventive conservation of cultural heritage (CH) by enabling the management of data collected from electronic sensors. This work presents an IoT architecture for this purpose. Firstly, we discuss the requirements from the artwork standpoint, data acquisition, cloud processing and data visualization to the end user. The results presented in this work focuses on the most critical aspect of the architecture, which are the sensor nodes. We designed a solution based on LoRa and Sigfox technologies to produce the minimum impact in the artwork, achieving a lifespan of more than 10 years. The solution will be capable of scaling the processing and storage resources, deployed either in a public or on-premise cloud, embedding complex predictive models. This combination of technologies can cope with different types of cultural heritage environments.

Family Impact and Parenting Styles in Families of Children with ADHD

Attention deficit hyperactivity disorder (ADHD) begins in childhood and is characterized by attention deficits, hyperactivity, or impulsiveness that is inconsistent with a child’s developmental level. The effects of ADHD are not limited to the child alone but can affect their familial context, particularly parenting styles. Using data from 68 parents of 6–11-year-old ADHD-diagnosed children, we attempted to identify the predictive variables of two parenting styles: criticism-rejection and permissiveness-indulgence. We analyzed two complex predictive models using structural equation modeling. We hypothesized that family impact variables would mediate the relation between the child’s behavior and parenting. The data showed that the child’s ADHD was only indirectly related to parenting styles, whereas child’s behavior problems had a direct relationship. The results stressed the central role of the child’s behavior on family social life, parents’ marital relationship, and parents’ feelings about their children. These variables mediated the relationship between the children’s disorders and parenting styles. On the other hand, perceived social support had an inverse relationship with this negative family impact, and it even had relevant indirect effects on criticism and permissiveness.

Impact of environmental variables on Dubas bug infestation rate: A case study from the Sultanate of Oman.

Date palm cultivation is economically important in the Sultanate of Oman, with significant financial investment coming from both the government and from private individuals. However, a global infestation of Dubas bug (Ommatissus lybicus Bergevin) has impacted the Middle East region, and infestations of date palms have been widespread. In this study, spatial analysis and geostatistical techniques were used to model the spatial distribution of Dubas bug infestations to (a) identify correlations between Dubas bug densities and different environmental variables, and (b) predict the locations of future Dubas bug infestations in Oman. Firstly, we considered individual environmental variables and their correlations with infestation locations. Then, we applied more complex predictive models and regression analysis techniques to investigate the combinations of environmental factors most conducive to the survival and spread of the Dubas bug. Environmental variables including elevation, geology, and distance to drainage pathways were found to significantly affect Dubas bug infestations. In contrast, aspect and hillshade did not significantly impact on Dubas bug infestations. Understanding their distribution and therefore applying targeted controls on their spread is important for effective mapping, control and management (e.g., resource allocation) of Dubas bug infestations.

Impact of functional somatic symptoms on 5–7-year-olds' healthcare use and costs

ObjectiveKnowledge about childhood functional somatic symptoms (FSS) and healthcare costs is scarce. This study aims to assess whether FSS in children aged 5–7 years are associated with increased future primary...

Severity Scores in Emergency Department Patients With Presumed Infection: A Prospective Validation Study.

The objectives of this study were to 1) validate a number of severity of illness scores in a large cohort of emergency department patients admitted with presumed infection and 2) compare the performance of scores in patient subgroups with increasing mortality: infection without systemic inflammatory response syndrome, sepsis, severe sepsis, and septic shock. Prospective, observational study. Adult emergency department in a metropolitan tertiary, university-affiliated hospital. Emergency department patients admitted with presumed infection. None. Consecutive emergency department patients admitted with presumed infection were identified over 160 weeks in two periods between 2007 and 2011. Clinical and laboratory data sufficient to calculate Mortality in Emergency Department Sepsis score, Acute Physiology and Chronic Health Evaluation II, Simplified Acute Physiology Score II, Sequential Organ Failure Assessment, and the Severe Sepsis Score were entered into a database. Model discrimination was quantified using area under the receiver operating curve. Calibration was assessed using visual plots, Hosmer-Lemeshow statistics, and linear regressions of observed and predicted values. A total of 8,871 patients were enrolled with 30-day mortality of 3.7%. Area under the receiver operating curve values for the entire cohort were: Mortality in Emergency Department Sepsis score of 0.92, Simplified Acute Physiology Score II and Acute Physiology and Chronic Health Evaluation II scores of 0.90, Sequential Organ Failure Assessment score of 0.86, and Severe Sepsis Score of 0.82. Discrimination decreased in subgroups with greater mortality for each score. All scores overestimated mortality, but closest concordance between predicted and observed mortality was seen with Mortality in Emergency Department Sepsis score. The decrease in area under the receiver operating curve seen in subgroups with increasing mortality may explain some variation in results seen in previous validation studies. Scores developed in intensive care settings overestimated mortality in the emergency department. Our results underscore the importance of employing predictive models developed in similar patient populations. The Mortality in Emergency Department Sepsis score outperformed more complex predictive models and would be the most appropriate scoring system for use in similar emergency department populations with a wide spectrum of mortality risk.

Modelling of an Extended Brutedl Algorithm for Rule Extraction

Complex predictive models obtain very high predictive performance; however, it is difficult to explain their complex mathematical design. Rule extraction techniques help to understand their designs by generating structures like decision list. BruteDL algorithm generates decision list from a dataset, and also addresses the overlapping rule problem of most decision list learners. However; it does not harness the power of complex predictive model. It also performs poorly with small dataset. Hence, this work aimed to create rule extraction technique by extending BruteDL and to address its poor performance with small dataset. A rule extraction technique named BruteDL-RET (Brute Decision List-Rule Extraction Technique) was modeled and implemented. A finite state automaton was used to model the technique. A functionality to generate supporting training set was included. Artificial Neural Networks (ANN) is chosen as the complex predictive model which serves as the oracle because it decides the class of each example. Decision list was generated using both the predictive model and the dataset it was trained with. The implementation was done using Java programming language. We prove that on the average BruteDL-RET is able to generate more accurate rules than BruteDL. We report on the performance of our model using dataset of UCI repository.

Analysis of Low Intensity Laser Therapy as adjuvant to Photodynamic Therapy in Nonmelanoma Skin Cancer.

The combination of Low Intensity Laser Therapy and Photodynamic Therapy constitutes a minimal invasive and highly selective alternative for the treatment of skin tumors. Photodynamic effectiveness, specially the absence of recurrence, highly depends on the type of tumor, as well as on several parameters such as those related with the optical source. Therefore it is important to have predictive tools to provide an appropriate treatment planning. This work proposes the use of a complex predictive model that includes a three-dimensional meshing of the considered tumor geometry and the optical propagation therein by a three-dimensional Monte Carlo method. The proposed tool is applied to three different types of skin tumors, squamous cell carcinoma, nodular and infiltrative basal cell carcinomas. The estimated volume of treatment, depending on the tumor type and source parameters, allows an appropriate treatment planning in order to avoid undesirable tumor recurrence.

Self-rated health and hospital services use in the Spanish National Health System: a longitudinal study.

BackgroundSelf-rated health is a subjective measure that has been related to indicators such as mortality, morbidity, functional capacity, and the use of health services. In Spain, there are few longitudinal studies associating self-rated health with hospital services use. The purpose of this study is to analyze the association between self-rated health and socioeconomic, demographic, and health variables, and the use of hospital services among the general population in the Region of Valencia, Spain.MethodsLongitudinal study of 5,275 adults who were included in the 2005 Region of Valencia Health Survey and linked to the Minimum Hospital Data Set between 2006 and 2009. Logistic regression models were used to calculate the odds ratios between use of hospital services and self-rated health, sex, age, educational level, employment status, income, country of birth, chronic conditions, disability and previous use of hospital services.ResultsBy the end of a 4-year follow-up period, 1,184 participants (22.4 %) had used hospital services. Use of hospital services was associated with poor self-rated health among both men and women. In men, it was also associated with unemployment, low income, and the presence of a chronic disease. In women, it was associated with low educational level, the presence of a disability, previous hospital services use, and the presence of chronic disease. Interactions were detected between self-rated health and chronic disease in men and between self-rated health and educational level in women.ConclusionsSelf-rated health acts as a predictor of hospital services use. Various health and socioeconomic variables provide additional predictive capacity. Interactions were detected between self-rated health and other variables that may reflect different complex predictive models, by gender.

The Application of Sovereign Bond Spreads: The Case of United Kingdom, Iceland, Norway, Switzerland and Russia

The yield curve – specifically the spread between long term and short term interest rates is a valuable forecasting tool. It is simple to use and significantly outperforms other financial and macroeconomic indicators in predicting recessions two to six quarters ahead. The steepness of the yield curve should be an excellent indicator of a possible future economic activity. A rise in the short rate tends to flatten the yield curve as well as to slow real growth the near term. This paper aims to analyze the dependence between slope of the yield curve and an economic activity ofUnited Kingdom, Iceland, Norway, Switzerland and Russia between the years 2000 and 2013. This time period was divided on two samples in order to prove the change of predictive power of the model. The slope of the yield curve can be measured as the yield spread between sovereign 10-year bonds and sovereign 3-month bonds. The natural and probably the most popular measure of economic growth is by GDP growth, taken quarterly. The results showed that the prediction ability of the GDP growth or decrease was proven after year 2008 (financial crisis) in Iceland, Russia and United Kingdom. Certainly the simple yield curve growth forecast should not serve as a replacement for the complex predictive models, it does, however, provide enough information to serve as a useful check on the more sophisticated forecasts. These findings can be beneficial for investors and provide further evidence of the potential usefulness of the yield curve spreads as indicators of the future economic activity.

Estimating Storm-Induced Dune Erosion and Overtopping along U.S. West Coast Beaches

Mull, J. and Ruggiero, P., 2014. Estimating storm-induced dune erosion and overtopping along U.S. West Coast beaches.Coastal foredunes protect lives, infrastructure, and ecosystems during severe storms. A range of approaches, ranging from simple (e.g., geometric) to complex (e.g., process based) predictive models, have been developed to quantify overtopping and foredune retreat during storms. At present, however, there is no widely accepted approach for assessing the vulnerability of coastal foredunes to erosion and overtopping hazards. Because different coastal regions have distinct storm and geomorphic characteristics, models need to be assessed and possibly adapted before they can be successfully applied to each unique region. In this study we apply a total water level (TWL) model and three simple foredune erosion models and assess their suitability for U.S. West Coast vulnerability analyses. The erosion models include a geometric model, an equilibrium profile model, and a wave impact model. We discuss the assumptions required to implement each model and force them with hydrodynamic conditions associated with a large-scale laboratory dune erosion experiment and a major winter storm approximately equivalent to the 30 year return period (TWL) event in the U.S. Pacific Northwest. The models are each applied with beach and foredune characteristics extracted from airborne topographic LiDAR (light detection and ranging) data collected along the coasts of Oregon and Washington. Sensitivity tests reveal distinct differences in model dependence on beach slope, a critical parameter in determining storm water level elevations and ultimately the extent of foredune retreat. Without detailed, reliable field observations of storm-induced dune erosion from the region, the accuracy of each model is determined using results of the dune erosion experiment. Estimates of both overtopping and erosion extent are normalized by foredune dimensions, enabling comparisons of relative vulnerability between different reaches of coast.