Ensemble methods for environmental data modelling with support vector regression

This paper investigates the use of ensemble of predictors in order to improve the performance of spatial prediction methods. Support vector regression (SVR), a popular method from the field of statistical machine learning, is used. Several instances of SVR are combined using different data sampling schemes (bagging and boosting). Bagging shows good performance, and proves to be more computationally efficient than training a single SVR model while reducing error. Boosting, however, does not improve results on this specific problem.

Improvement of therapeutic drug monitoring of imatinib by Bayesian prediction of trough levels. Personalized drug dosage

Aims: Plasma concentrations of imatinib differ largely between patients despite same dosage, owing to large inter-individual variability in pharmacokinetic (PK) parameters. As the drug concentration at the end of the dosage interval (Cmin) correlates with treatment response and tolerability, monitoring of Cmin is suggested for therapeutic drug monitoring (TDM) of imatinib. Due to logistic difficulties, random sampling during the dosage interval is however often performed in clinical practice, thus rendering the respective results not informative regarding Cmin values.Objectives: (I) To extrapolate randomly measured imatinib concentrations to more informative Cmin using classical Bayesian forecasting. (II) To extend the classical Bayesian method to account for correlation between PK parameters. (III) To evaluate the predictive performance of both methods.Methods: 31 paired blood samples (random and trough levels) were obtained from 19 cancer patients under imatinib. Two Bayesian maximum a posteriori (MAP) methods were implemented: (A) a classical method ignoring correlation between PK parameters, and (B) an extended one accounting for correlation. Both methods were applied to estimate individual PK parameters, conditional on random observations and covariate-adjusted priors from a population PK model. The PK parameter estimates were used to calculate trough levels. Relative prediction errors (PE) were analyzed to evaluate accuracy (one-sample t-test) and to compare precision between the methods (F-test to compare variances).Results: Both Bayesian MAP methods allowed non-biased predictions of individual Cmin compared to observations: (A) - 7% mean PE (CI95% - 18 to 4 %, p = 0.15) and (B) - 4% mean PE (CI95% - 18 to 10 %, p = 0.69). Relative standard deviations of actual observations from predictions were 22% (A) and 30% (B), i.e. comparable to the intraindividual variability reported. Precision was not improved by taking into account correlation between PK parameters (p = 0.22).Conclusion: Clinical interpretation of randomly measured imatinib concentrations can be assisted by Bayesian extrapolation to maximum likelihood Cmin. Classical Bayesian estimation can be applied for TDM without the need to include correlation between PK parameters. Both methods could be adapted in the future to evaluate other individual pharmacokinetic measures correlated to clinical outcomes, such as area under the curve(AUC).

Prediction of cross-recognition of peptide-HLA A2 by Melan-A-specific cytotoxic T lymphocytes using three-dimensional quantitative structure-activity relationships.

The cross-recognition of peptides by cytotoxic T lymphocytes is a key element in immunology and in particular in peptide based immunotherapy. Here we develop three-dimensional (3D) quantitative structure-activity relationships (QSARs) to predict cross-recognition by Melan-A-specific cytotoxic T lymphocytes of peptides bound to HLA A*0201 (hereafter referred to as HLA A2). First, we predict the structure of a set of self- and pathogen-derived peptides bound to HLA A2 using a previously developed ab initio structure prediction approach [Fagerberg et al., J. Mol. Biol., 521-46 (2006)]. Second, shape and electrostatic energy calculations are performed on a 3D grid to produce similarity matrices which are combined with a genetic neural network method [So et al., J. Med. Chem., 4347-59 (1997)] to generate 3D-QSAR models. The models are extensively validated using several different approaches. During the model generation, the leave-one-out cross-validated correlation coefficient (q (2)) is used as the fitness criterion and all obtained models are evaluated based on their q (2) values. Moreover, the best model obtained for a partitioned data set is evaluated by its correlation coefficient (r = 0.92 for the external test set). The physical relevance of all models is tested using a functional dependence analysis and the robustness of the models obtained for the entire data set is confirmed using y-randomization. Finally, the validated models are tested for their utility in the setting of rational peptide design: their ability to discriminate between peptides that only contain side chain substitutions in a single secondary anchor position is evaluated. In addition, the predicted cross-recognition of the mono-substituted peptides is confirmed experimentally in chromium-release assays. These results underline the utility of 3D-QSARs in peptide mimetic design and suggest that the properties of the unbound epitope are sufficient to capture most of the information to determine the cross-recognition.

A novel approach to reducing number of sensing units for wearable gait analysis systems.

Gait analysis methods to estimate spatiotemporal measures, based on two, three or four gyroscopes attached on lower limbs have been discussed in the literature. The most common approach to reduce the number of sensing units is to simplify the underlying biomechanical gait model. In this study, we propose a novel method based on prediction of movements of thighs from movements of shanks. Datasets from three previous studies were used. Data from the first study (ten healthy subjects and ten with Parkinson's disease) were used to develop and calibrate a system with only two gyroscopes attached on shanks. Data from two other studies (36 subjects with hip replacement, seven subjects with coxarthrosis, and eight control subjects) were used for comparison with the other methods and for assessment of error compared to a motion capture system. Results show that the error of estimation of stride length compared to motion capture with the system with four gyroscopes and our new method based on two gyroscopes was close ( -0.8 ±6.6 versus 3.8 ±6.6 cm). An alternative with three sensing units did not show better results (error: -0.2 ±8.4 cm). Finally, a fourth that also used two units but with a simpler gait model had the highest bias compared to the reference (error: -25.6 ±7.6 cm). We concluded that it is feasible to estimate movements of thighs from movements of shanks to reduce number of needed sensing units from 4 to 2 in context of ambulatory gait analysis.

A comparison of Bayesian and frequentist approaches to incorporating external information for the prediction of prostate cancer risk.

We present the most comprehensive comparison to date of the predictive benefit of genetics in addition to currently used clinical variables, using genotype data for 33 single-nucleotide polymorphisms (SNPs) in 1,547 Caucasian men from the placebo arm of the REduction by DUtasteride of prostate Cancer Events (REDUCE®) trial. Moreover, we conducted a detailed comparison of three techniques for incorporating genetics into clinical risk prediction. The first method was a standard logistic regression model, which included separate terms for the clinical covariates and for each of the genetic markers. This approach ignores a substantial amount of external information concerning effect sizes for these Genome Wide Association Study (GWAS)-replicated SNPs. The second and third methods investigated two possible approaches to incorporating meta-analysed external SNP effect estimates - one via a weighted PCa 'risk' score based solely on the meta analysis estimates, and the other incorporating both the current and prior data via informative priors in a Bayesian logistic regression model. All methods demonstrated a slight improvement in predictive performance upon incorporation of genetics. The two methods that incorporated external information showed the greatest receiver-operating-characteristic AUCs increase from 0.61 to 0.64. The value of our methods comparison is likely to lie in observations of performance similarities, rather than difference, between three approaches of very different resource requirements. The two methods that included external information performed best, but only marginally despite substantial differences in complexity.

Comparison of an EMG-based and a stress-based method to predict shoulder muscle forces.

The estimation of muscle forces in musculoskeletal shoulder models is still controversial. Two different methods are widely used to solve the indeterminacy of the system: electromyography (EMG)-based methods and stress-based methods. The goal of this work was to evaluate the influence of these two methods on the prediction of muscle forces, glenohumeral load and joint stability after total shoulder arthroplasty. An EMG-based and a stress-based method were implemented into the same musculoskeletal shoulder model. The model replicated the glenohumeral joint after total shoulder arthroplasty. It contained the scapula, the humerus, the joint prosthesis, the rotator cuff muscles supraspinatus, subscapularis and infraspinatus and the middle, anterior and posterior deltoid muscles. A movement of abduction was simulated in the plane of the scapula. The EMG-based method replicated muscular activity of experimentally measured EMG. The stress-based method minimised a cost function based on muscle stresses. We compared muscle forces, joint reaction force, articular contact pressure and translation of the humeral head. The stress-based method predicted a lower force of the rotator cuff muscles. This was partly counter-balanced by a higher force of the middle part of the deltoid muscle. As a consequence, the stress-based method predicted a lower joint load (16% reduced) and a higher superior-inferior translation of the humeral head (increased by 1.2 mm). The EMG-based method has the advantage of replicating the observed cocontraction of stabilising muscles of the rotator cuff. This method is, however, limited to available EMG measurements. The stress-based method has thus an advantage of flexibility, but may overestimate glenohumeral subluxation.

Spatial prediction of monthly wind speeds in complex terrain with adaptive general regression neural networks

This paper presents the general regression neural networks (GRNN) as a nonlinear regression method for the interpolation of monthly wind speeds in complex Alpine orography. GRNN is trained using data coming from Swiss meteorological networks to learn the statistical relationship between topographic features and wind speed. The terrain convexity, slope and exposure are considered by extracting features from the digital elevation model at different spatial scales using specialised convolution filters. A database of gridded monthly wind speeds is then constructed by applying GRNN in prediction mode during the period 1968-2008. This study demonstrates that using topographic features as inputs in GRNN significantly reduces cross-validation errors with respect to low-dimensional models integrating only geographical coordinates and terrain height for the interpolation of wind speed. The spatial predictability of wind speed is found to be lower in summer than in winter due to more complex and weaker wind-topography relationships. The relevance of these relationships is studied using an adaptive version of the GRNN algorithm which allows to select the useful terrain features by eliminating the noisy ones. This research provides a framework for extending the low-dimensional interpolation models to high-dimensional spaces by integrating additional features accounting for the topographic conditions at multiple spatial scales. Copyright (c) 2012 Royal Meteorological Society.

Use of anisotropic modelling in electrical impedance tomography: description of method and preliminary assessment of utility in imaging brain function in the adult human head.

Electrical Impedance Tomography (EIT) is an imaging method which enables a volume conductivity map of a subject to be produced from multiple impedance measurements. It has the potential to become a portable non-invasive imaging technique of particular use in imaging brain function. Accurate numerical forward models may be used to improve image reconstruction but, until now, have employed an assumption of isotropic tissue conductivity. This may be expected to introduce inaccuracy, as body tissues, especially those such as white matter and the skull in head imaging, are highly anisotropic. The purpose of this study was, for the first time, to develop a method for incorporating anisotropy in a forward numerical model for EIT of the head and assess the resulting improvement in image quality in the case of linear reconstruction of one example of the human head. A realistic Finite Element Model (FEM) of an adult human head with segments for the scalp, skull, CSF, and brain was produced from a structural MRI. Anisotropy of the brain was estimated from a diffusion tensor-MRI of the same subject and anisotropy of the skull was approximated from the structural information. A method for incorporation of anisotropy in the forward model and its use in image reconstruction was produced. The improvement in reconstructed image quality was assessed in computer simulation by producing forward data, and then linear reconstruction using a sensitivity matrix approach. The mean boundary data difference between anisotropic and isotropic forward models for a reference conductivity was 50%. Use of the correct anisotropic FEM in image reconstruction, as opposed to an isotropic one, corrected an error of 24 mm in imaging a 10% conductivity decrease located in the hippocampus, improved localisation for conductivity changes deep in the brain and due to epilepsy by 4-17 mm, and, overall, led to a substantial improvement on image quality. This suggests that incorporation of anisotropy in numerical models used for image reconstruction is likely to improve EIT image quality.

Prediction of hip fracture risk by quantitative ultrasound in more than 7000 Swiss women > or =70 years of age: comparison of three technologically different bone ultrasound devices in the SEMOF study.

To compare the prediction of hip fracture risk of several bone ultrasounds (QUS), 7062 Swiss women > or =70 years of age were measured with three QUSs (two of the heel, one of the phalanges). Heel QUSs were both predictive of hip fracture risk, whereas the phalanges QUS was not. INTRODUCTION: As the number of hip fracture is expected to increase during these next decades, it is important to develop strategies to detect subjects at risk. Quantitative bone ultrasound (QUS), an ionizing radiation-free method, which is transportable, could be interesting for this purpose. MATERIALS AND METHODS: The Swiss Evaluation of the Methods of Measurement of Osteoporotic Fracture Risk (SEMOF) study is a multicenter cohort study, which compared three QUSs for the assessment of hip fracture risk in a sample of 7609 elderly ambulatory women > or =70 years of age. Two QUSs measured the heel (Achilles+; GE-Lunar and Sahara; Hologic), and one measured the heel (DBM Sonic 1200; IGEA). The Cox proportional hazards regression was used to estimate the hazard of the first hip fracture, adjusted for age, BMI, and center, and the area under the ROC curves were calculated to compare the devices and their parameters. RESULTS: From the 7609 women who were included in the study, 7062 women 75.2 +/- 3.1 (SD) years of age were prospectively followed for 2.9 +/- 0.8 years. Eighty women reported a hip fracture. A decrease by 1 SD of the QUS variables corresponded to an increase of the hip fracture risk from 2.3 (95% CI, 1.7, 3.1) to 2.6 (95% CI, 1.9, 3.4) for the three variables of Achilles+ and from 2.2 (95% CI, 1.7, 3.0) to 2.4 (95% CI, 1.8, 3.2) for the three variables of Sahara. Risk gradients did not differ significantly among the variables of the two heel QUS devices. On the other hand, the phalanges QUS (DBM Sonic 1200) was not predictive of hip fracture risk, with an adjusted hazard risk of 1.2 (95% CI, 0.9, 1.5), even after reanalysis of the digitalized data and using different cut-off levels (1700 or 1570 m/s). CONCLUSIONS: In this elderly women population, heel QUS devices were both predictive of hip fracture risk, whereas the phalanges QUS device was not.

The Atherosclerosis Burden Score (ABS): a Convenient Ultrasound-Based Score of Peripheral Atherosclerosis for Coronary Artery Disease Prediction.

Ultrasonographic detection of subclinical atherosclerosis improves cardiovascular risk stratification, but uncertainty persists about the most discriminative method to apply. In this study, we found that the "atherosclerosis burden score (ABS)", a novel straightforward ultrasonographic score that sums the number of carotid and femoral arterial bifurcations with plaques, significantly outperformed common carotid intima-media thickness, carotid mean/maximal thickness, and carotid/femoral plaque scores for the detection of coronary artery disease (CAD) (receiver operating characteristic (ROC) curve area under the curve (AUC) = 0.79; P = 0.027 to <0.001 with the other five US endpoints) in 203 patients undergoing coronary angiography. ABS was also more correlated with CAD extension (R = 0.55; P < 0.001). Furthermore, in a second group of 1128 patients without cardiovascular disease, ABS was weakly correlated with the European Society of Cardiology chart risk categories (R (2) = 0.21), indicating that ABS provided information beyond usual cardiovascular risk factor-based risk stratification. Pending prospective studies on hard cardiovascular endpoints, ABS appears as a promising tool in primary prevention.

Functional error modeling for uncertainty quantification in hydrogeology

Approximate models (proxies) can be employed to reduce the computational costs of estimating uncertainty. The price to pay is that the approximations introduced by the proxy model can lead to a biased estimation. To avoid this problem and ensure a reliable uncertainty quantification, we propose to combine functional data analysis and machine learning to build error models that allow us to obtain an accurate prediction of the exact response without solving the exact model for all realizations. We build the relationship between proxy and exact model on a learning set of geostatistical realizations for which both exact and approximate solvers are run. Functional principal components analysis (FPCA) is used to investigate the variability in the two sets of curves and reduce the dimensionality of the problem while maximizing the retained information. Once obtained, the error model can be used to predict the exact response of any realization on the basis of the sole proxy response. This methodology is purpose-oriented as the error model is constructed directly for the quantity of interest, rather than for the state of the system. Also, the dimensionality reduction performed by FPCA allows a diagnostic of the quality of the error model to assess the informativeness of the learning set and the fidelity of the proxy to the exact model. The possibility of obtaining a prediction of the exact response for any newly generated realization suggests that the methodology can be effectively used beyond the context of uncertainty quantification, in particular for Bayesian inference and optimization.

Accuracy and Precision of Head Motion Information in Multi-Channel Free Induction Decay Navigators for Magnetic Resonance Imaging.

Free induction decay (FID) navigators were found to qualitatively detect rigid-body head movements, yet it is unknown to what extent they can provide quantitative motion estimates. Here, we acquired FID navigators at different sampling rates and simultaneously measured head movements using a highly accurate optical motion tracking system. This strategy allowed us to estimate the accuracy and precision of FID navigators for quantification of rigid-body head movements. Five subjects were scanned with a 32-channel head coil array on a clinical 3T MR scanner during several resting and guided head movement periods. For each subject we trained a linear regression model based on FID navigator and optical motion tracking signals. FID-based motion model accuracy and precision was evaluated using cross-validation. FID-based prediction of rigid-body head motion was found to be with a mean translational and rotational error of 0.14±0.21 mm and 0.08±0.13(°) , respectively. Robust model training with sub-millimeter and sub-degree accuracy could be achieved using 100 data points with motion magnitudes of ±2 mm and ±1(°) for translation and rotation. The obtained linear models appeared to be subject-specific as inter-subject application of a "universal" FID-based motion model resulted in poor prediction accuracy. The results show that substantial rigid-body motion information is encoded in FID navigator signal time courses. Although, the applied method currently requires the simultaneous acquisition of FID signals and optical tracking data, the findings suggest that multi-channel FID navigators have a potential to complement existing tracking technologies for accurate rigid-body motion detection and correction in MRI.