Extreme Precipitation Modelling Using Geostatistics and Machine Learning Algorithms

The paper presents an approach for mapping of precipitation data. The main goal is to perform spatial predictions and simulations of precipitation fields using geostatistical methods (ordinary kriging, kriging with external drift) as well as machine learning algorithms (neural networks). More practically, the objective is to reproduce simultaneously both the spatial patterns and the extreme values. This objective is best reached by models integrating geostatistics and machine learning algorithms. To demonstrate how such models work, two case studies have been considered: first, a 2-day accumulation of heavy precipitation and second, a 6-day accumulation of extreme orographic precipitation. The first example is used to compare the performance of two optimization algorithms (conjugate gradients and Levenberg-Marquardt) of a neural network for the reproduction of extreme values. Hybrid models, which combine geostatistical and machine learning algorithms, are also treated in this context. The second dataset is used to analyze the contribution of radar Doppler imagery when used as external drift or as input in the models (kriging with external drift and neural networks). Model assessment is carried out by comparing independent validation errors as well as analyzing data patterns.

Machine learning algorithms for spatial data. Case studies: Environmental pollution, natural hazards, renewable resources

This paper presents general problems and approaches for the spatial data analysis using machine learning algorithms. Machine learning is a very powerful approach to adaptive data analysis, modelling and visualisation. The key feature of the machine learning algorithms is that they learn from empirical data and can be used in cases when the modelled environmental phenomena are hidden, nonlinear, noisy and highly variable in space and in time. Most of the machines learning algorithms are universal and adaptive modelling tools developed to solve basic problems of learning from data: classification/pattern recognition, regression/mapping and probability density modelling. In the present report some of the widely used machine learning algorithms, namely artificial neural networks (ANN) of different architectures and Support Vector Machines (SVM), are adapted to the problems of the analysis and modelling of geo-spatial data. Machine learning algorithms have an important advantage over traditional models of spatial statistics when problems are considered in a high dimensional geo-feature spaces, when the dimension of space exceeds 5. Such features are usually generated, for example, from digital elevation models, remote sensing images, etc. An important extension of models concerns considering of real space constrains like geomorphology, networks, and other natural structures. Recent developments in semi-supervised learning can improve modelling of environmental phenomena taking into account on geo-manifolds. An important part of the study deals with the analysis of relevant variables and models' inputs. This problem is approached by using different feature selection/feature extraction nonlinear tools. To demonstrate the application of machine learning algorithms several interesting case studies are considered: digital soil mapping using SVM, automatic mapping of soil and water system pollution using ANN; natural hazards risk analysis (avalanches, landslides), assessments of renewable resources (wind fields) with SVM and ANN models, etc. The dimensionality of spaces considered varies from 2 to more than 30. Figures 1, 2, 3 demonstrate some results of the studies and their outputs. Finally, the results of environmental mapping are discussed and compared with traditional models of geostatistics.

Ruling out coronary heart disease in primary care patients with chest pain : a clinical prediction score

ABSTRACT: BACKGROUND: Chest pain raises concern for the possibility of coronary heart disease. Scoring methods have been developed to identify coronary heart disease in emergency settings, but not in primary care. METHODS: Data were collected from a multicenter Swiss clinical cohort study including 672 consecutive patients with chest pain, who had visited one of 59 family practitioners' offices. Using delayed diagnosis we derived a prediction rule to rule out coronary heart disease by means of a logistic regression model. Known cardiovascular risk factors, pain characteristics, and physical signs associated with coronary heart disease were explored to develop a clinical score. Patients diagnosed with angina or acute myocardial infarction within the year following their initial visit comprised the coronary heart disease group. RESULTS: The coronary heart disease score was derived from eight variables: age, gender, duration of chest pain from 1 to 60 minutes, substernal chest pain location, pain increases with exertion, absence of tenderness point at palpation, cardiovascular risks factors, and personal history of cardiovascular disease. Area under the receiver operating characteristics curve was of 0.95 with a 95% confidence interval of 0.92; 0.97. From this score, 413 patients were considered as low risk for values of percentile 5 of the coronary heart disease patients. Internal validity was confirmed by bootstrapping. External validation using data from a German cohort (Marburg, n = 774) revealed a receiver operating characteristics curve of 0.75 (95% confidence interval, 0.72; 0.81) with a sensitivity of 85.6% and a specificity of 47.2%. CONCLUSIONS: This score, based only on history and physical examination, is a complementary tool for ruling out coronary heart disease in primary care patients complaining of chest pain.

Evaluation of organ-specific peripheral doses after 2-dimensional, 3-dimensional and hybrid intensity modulated radiation therapy for breast cancer based on Monte Carlo and convolution/superposition algorithms: implications for secondary cancer risk assessment.

To make a comprehensive evaluation of organ-specific out-of-field doses using Monte Carlo (MC) simulations for different breast cancer irradiation techniques and to compare results with a commercial treatment planning system (TPS). Three breast radiotherapy techniques using 6MV tangential photon beams were compared: (a) 2DRT (open rectangular fields), (b) 3DCRT (conformal wedged fields), and (c) hybrid IMRT (open conformal+modulated fields). Over 35 organs were contoured in a whole-body CT scan and organ-specific dose distributions were determined with MC and the TPS. Large differences in out-of-field doses were observed between MC and TPS calculations, even for organs close to the target volume such as the heart, the lungs and the contralateral breast (up to 70% difference). MC simulations showed that a large fraction of the out-of-field dose comes from the out-of-field head scatter fluence (>40%) which is not adequately modeled by the TPS. Based on MC simulations, the 3DCRT technique using external wedges yielded significantly higher doses (up to a factor 4-5 in the pelvis) than the 2DRT and the hybrid IMRT techniques which yielded similar out-of-field doses. In sharp contrast to popular belief, the IMRT technique investigated here does not increase the out-of-field dose compared to conventional techniques and may offer the most optimal plan. The 3DCRT technique with external wedges yields the largest out-of-field doses. For accurate out-of-field dose assessment, a commercial TPS should not be used, even for organs near the target volume (contralateral breast, lungs, heart).

Current genetic data do not improve the prediction of type 2 diabetes mellitus: the CoLaus study.

CONTEXT: Several genetic risk scores to identify asymptomatic subjects at high risk of developing type 2 diabetes mellitus (T2DM) have been proposed, but it is unclear whether they add extra information to risk scores based on clinical and biological data. OBJECTIVE: The objective of the study was to assess the extra clinical value of genetic risk scores in predicting the occurrence of T2DM. DESIGN: This was a prospective study, with a mean follow-up time of 5 yr. SETTING AND SUBJECTS: The study included 2824 nondiabetic participants (1548 women, 52 ± 10 yr). MAIN OUTCOME MEASURE: Six genetic risk scores for T2DM were tested. Four were derived from the literature and two were created combining all (n = 24) or shared (n = 9) single-nucleotide polymorphisms of the previous scores. A previously validated clinic + biological risk score for T2DM was used as reference. RESULTS: Two hundred seven participants (7.3%) developed T2DM during follow-up. On bivariate analysis, no differences were found for all but one genetic score between nondiabetic and diabetic participants. After adjusting for the validated clinic + biological risk score, none of the genetic scores improved discrimination, as assessed by changes in the area under the receiver-operating characteristic curve (range -0.4 to -0.1%), sensitivity (-2.9 to -1.0%), specificity (0.0-0.1%), and positive (-6.6 to +0.7%) and negative (-0.2 to 0.0%) predictive values. Similarly, no improvement in T2DM risk prediction was found: net reclassification index ranging from -5.3 to -1.6% and nonsignificant (P ≥ 0.49) integrated discrimination improvement. CONCLUSIONS: In this study, adding genetic information to a previously validated clinic + biological score does not seem to improve the prediction of T2DM.

Prediction of recanalization in acute stroke patients receiving intravenous and endovascular revascularization therapy.

BACKGROUND AND PURPOSE: The study aims to assess the recanalization rate in acute ischemic stroke patients who received no revascularization therapy, intravenous thrombolysis, and endovascular treatment, respectively, and to identify best clinical and imaging predictors of recanalization in each treatment group. METHODS: Clinical and imaging data were collected in 103 patients with acute ischemic stroke caused by anterior circulation arterial occlusion. We recorded demographics and vascular risk factors. We reviewed the noncontrast head computed tomographies to assess for hyperdense middle cerebral artery and its computed tomography density. We reviewed the computed tomography angiograms and the raw images to determine the site and degree of arterial occlusion, collateral score, clot burden score, and the density of the clot. Recanalization status was assessed on recanalization imaging using Thrombolysis in Myocardial Ischemia. Multivariate logistic regressions were utilized to determine the best predictors of outcome in each treatment group. RESULTS: Among the 103 study patients, 43 (42%) received intravenous thrombolysis, 34 (33%) received endovascular thrombolysis, and 26 (25%) did not receive any revascularization therapy. In the patients with intravenous thrombolysis or no revascularization therapy, recanalization of the vessel was more likely with intravenous thrombolysis (P = 0·046) and when M1/A1 was occluded (P = 0·001). In this subgroup of patients, clot burden score, cervical degree of stenosis (North American Symptomatic Carotid Endarterectomy Trial), and hyperlipidemia status added information to the aforementioned likelihood of recanalization at the patient level (P < 0·001). In patients with endovascular thrombolysis, recanalization of the vessel was more likely in the case of a higher computed tomography angiogram clot density (P = 0·012), and in this subgroup of patients gender added information to the likelihood of recanalization at the patient level (P = 0·044). CONCLUSION: The overall likelihood of recanalization was the highest in the endovascular group, and higher for intravenous thrombolysis compared with no revascularization therapy. However, our statistical models of recanalization for each individual patient indicate significant variability between treatment options, suggesting the need to include this prediction in the personalized treatment selection.

The prediction of cardiovascular events is improved by combination of Framingham risk score and subclinical atherosclerosis imaging

Purpose: In primary prevention of cardiovascular disease (CVD), it is accepted that the intensity of risk factor treatment should be guided by the magnitude of absolute risk. Risk factors tools like Framingham risk score (FHS) or noninvasive atherosclerosis imaging tests are available to detect high risk subjects. However, these methods are imperfect and may misclassify a large number of individuals. The purpose of this prospective study was to evaluate whether the prediction of future cardiovascular events (CVE) can be improved when subclinical imaging atherosclerosis (SCATS) is combined with the FRS in asymptomatic subjects. Methods: Overall, 1038 asymptomatic subjects (413 women, 625 men, mean age 49.1±12.8 years) were assessed for their cardiovascular risk using the FRS. B-mode ultrasonography on carotid and femoral arteries was performed by two investigators to detect atherosclerotic plaques (focal thickening of intima-media > 1.2 mm) and to measure carotid intima-media thickness (C-IMT). The severity of SCATS was expressed by an ATS-burden Score (ABS) reflecting the number of the arterial sites with >1 plaques (range 0-4). CVE were defined as fatal or non fatal acute coronary syndrome, stroke, or angioplasty for peripheral artery disease. Results: during a mean follow-up of 4.9±3.1 years, 61 CVE were recorded. Event rates the rate of CVE increased significantly from 2.7% to 39.1% according to the ABS (p<0.001) and from 4% to 24.6% according to the quartiles of C-IMT. Similarly, FRS predicted CVE (p<0.001). When computing the angiographic markers of SCATS in addition of FRS, we observed an improvement of net reclassification rate of 16.6% (p< 0.04) for ABS as compared to 5.5% (p = 0.26) for C-IMT. Conclusion: these results indicate that the detection of subjects requiring more attention to prevent CVE can be significantly improved when using both FRS and SCATS imaging.

Debris flow hazard modelling on medium scale: Valtellina di Tirano, Italy

Debris flow hazard modelling at medium (regional) scale has been subject of various studies in recent years. In this study, hazard zonation was carried out, incorporating information about debris flow initiation probability (spatial and temporal), and the delimitation of the potential runout areas. Debris flow hazard zonation was carried out in the area of the Consortium of Mountain Municipalities of Valtellina di Tirano (Central Alps, Italy). The complexity of the phenomenon, the scale of the study, the variability of local conditioning factors, and the lacking data limited the use of process-based models for the runout zone delimitation. Firstly, a map of hazard initiation probabilities was prepared for the study area, based on the available susceptibility zoning information, and the analysis of two sets of aerial photographs for the temporal probability estimation. Afterwards, the hazard initiation map was used as one of the inputs for an empirical GIS-based model (Flow-R), developed at the University of Lausanne (Switzerland). An estimation of the debris flow magnitude was neglected as the main aim of the analysis was to prepare a debris flow hazard map at medium scale. A digital elevation model, with a 10 m resolution, was used together with landuse, geology and debris flow hazard initiation maps as inputs of the Flow-R model to restrict potential areas within each hazard initiation probability class to locations where debris flows are most likely to initiate. Afterwards, runout areas were calculated using multiple flow direction and energy based algorithms. Maximum probable runout zones were calibrated using documented past events and aerial photographs. Finally, two debris flow hazard maps were prepared. The first simply delimits five hazard zones, while the second incorporates the information about debris flow spreading direction probabilities, showing areas more likely to be affected by future debris flows. Limitations of the modelling arise mainly from the models applied and analysis scale, which are neglecting local controlling factors of debris flow hazard. The presented approach of debris flow hazard analysis, associating automatic detection of the source areas and a simple assessment of the debris flow spreading, provided results for consequent hazard and risk studies. However, for the validation and transferability of the parameters and results to other study areas, more testing is needed.

Flow-R, a model for susceptibility mapping of debris flows and other gravitational hazards at a regional scale: some case studies

Every year, debris flows cause huge damage in mountainous areas. Due to population pressure in hazardous zones, the socio-economic impact is much higher than in the past. Therefore, the development of indicative susceptibility hazard maps is of primary importance, particularly in developing countries. However, the complexity of the phenomenon and the variability of local controlling factors limit the use of processbased models for a first assessment. A debris flow model has been developed for regional susceptibility assessments using digital elevation model (DEM) with a GIS-based approach.. The automatic identification of source areas and the estimation of debris flow spreading, based on GIS tools, provide a substantial basis for a preliminary susceptibility assessment at a regional scale. One of the main advantages of this model is its workability. In fact, everything is open to the user, from the data choice to the selection of the algorithms and their parameters. The Flow-R model was tested in three different contexts: two in Switzerland and one in Pakistan, for indicative susceptibility hazard mapping. It was shown that the quality of the DEM is the most important parameter to obtain reliable results for propagation, but also to identify the potential debris flows sources.

Trabecular bone score improves fracture risk prediction in non-osteoporotic women: the OFELY study.

The use of areal bone mineral density (aBMD) for fracture prediction may be enhanced by considering bone microarchitectural deterioration. Trabecular bone score (TBS) helped in redefining a significant subset of non-osteoporotic women as a higher risk group. INTRODUCTION: TBS is an index of bone microarchitecture. Our goal was to assess the ability of TBS to predict incident fracture. METHODS: TBS was assessed in 560 postmenopausal women from the Os des Femmes de Lyon cohort, who had a lumbar spine (LS) DXA scan (QDR 4500A, Hologic) between years 2000 and 2001. During a mean follow-up of 7.8 ± 1.3 years, 94 women sustained 112 fragility fractures. RESULTS: At the time of baseline DXA scan, women with incident fracture were significantly older (70 ± 9 vs. 65 ± 8 years) and had a lower LS_aBMD and LS_TBS (both -0.4SD, p < 0.001) than women without fracture. The magnitude of fracture prediction was similar for LS_aBMD and LS_TBS (odds ratio [95 % confidence interval] = 1.4 [1.2;1.7] and 1.6 [1.2;2.0]). After adjustment for age and prevalent fracture, LS_TBS remained predictive of an increased risk of fracture. Yet, its addition to age, prevalent fracture, and LS_aBMD did not reach the level of significance to improve the fracture prediction. When using the WHO classification, 39 % of fractures occurred in osteoporotic women, 46 % in osteopenic women, and 15 % in women with T-score > -1. Thirty-seven percent of fractures occurred in the lowest quartile of LS_TBS, regardless of BMD. Moreover, 35 % of fractures that occurred in osteopenic women were classified below this LS_TBS threshold. CONCLUSION: In conclusion, LS_aBMD and LS_TBS predicted fractures equally well. In our cohort, the addition of LS_TBS to age and LS_aBMD added only limited information on fracture risk prediction. However, using the lowest quartile of LS_TBS helped in redefining a significant subset of non-osteoporotic women as a higher risk group which is important for patient management.