Predicting Bacteremia in Children With Cancer and Fever in Chemotherapy-induced Neutropenia: Results of the Prospective Multicenter SPOG 2003 FN Study.

STUDY AIM:: To develop a score predicting the risk of bacteremia in cancer patients with fever and neutropenia (FN), and to evaluate its performance. METHODS:: Pediatric patients with cancer presenting with FN induced by nonmyeloablative chemotherapy were observed in a prospective multicenter study. A score predicting the risk of bacteremia was developed from a multivariate mixed logistic regression model. Its cross-validated predictive performance was compared with that of published risk prediction rules. RESULTS:: Bacteremia was reported in 67 (16%) of 423 FN episodes. In 34 episodes (8%), bacteremia became known only after reassessment after 8 to 24 hours of inpatient management. Predicting bacteremia at reassessment was better than prediction at presentation with FN. A differential leukocyte count did not increase the predictive performance. The reassessment score predicting future bacteremia in 390 episodes without known bacteremia used the following 4 variables: hemoglobin ≥90 g/L at presentation (weight 3), platelet count <50 G/L (3), shaking chills (5), and other need for inpatient treatment or observation according to the treating physician (3). Applying a threshold ≥3, the score-simplified into a low-risk checklist-predicted bacteremia with 100% sensitivity, with 54 episodes (13%) classified as low-risk, and a specificity of 15%. CONCLUSIONS:: This reassessment score, simplified into a low-risk checklist of 4 routinely accessible characteristics, identifies pediatric patients with FN at risk for bacteremia. It has the potential to contribute to the reduction of use of antimicrobials in, and to shorten the length of hospital stays of pediatric patients with cancer and FN.

Predicting adverse events in children with fever and chemotherapy-induced neutropenia: the prospective multicenter SPOG 2003 FN study.

PURPOSE To develop a score predicting the risk of adverse events (AEs) in pediatric patients with cancer who experience fever and neutropenia (FN) and to evaluate its performance. PATIENTS AND METHODS Pediatric patients with cancer presenting with FN induced by nonmyeloablative chemotherapy were observed in a prospective multicenter study. A score predicting the risk of future AEs (ie, serious medical complication, microbiologically defined infection, radiologically confirmed pneumonia) was developed from a multivariate mixed logistic regression model. Its cross-validated predictive performance was compared with that of published risk prediction rules. Results An AE was reported in 122 (29%) of 423 FN episodes. In 57 episodes (13%), the first AE was known only after reassessment after 8 to 24 hours of inpatient management. Predicting AE at reassessment was better than prediction at presentation with FN. A differential leukocyte count did not increase the predictive performance. The score predicting future AE in 358 episodes without known AE at reassessment used the following four variables: preceding chemotherapy more intensive than acute lymphoblastic leukemia maintenance (weight = 4), hemoglobin > or = 90 g/L (weight = 5), leukocyte count less than 0.3 G/L (weight = 3), and platelet count less than 50 G/L (weight = 3). A score (sum of weights) > or = 9 predicted future AEs. The cross-validated performance of this score exceeded the performance of published risk prediction rules. At an overall sensitivity of 92%, 35% of the episodes were classified as low risk, with a specificity of 45% and a negative predictive value of 93%. CONCLUSION This score, based on four routinely accessible characteristics, accurately identifies pediatric patients with cancer with FN at risk for AEs after reassessment.

Automated quantitative pupillometry for the prognostication of coma after cardiac arrest.

BACKGROUND: Sedation and therapeutic hypothermia (TH) delay neurological responses and might reduce the accuracy of clinical examination to predict outcome after cardiac arrest (CA). We examined the accuracy of quantitative pupillary light reactivity (PLR), using an automated infrared pupillometry, to predict outcome of post-CA coma in comparison to standard PLR, EEG, and somato-sensory evoked potentials (SSEP). METHODS: We prospectively studied over a 1-year period (June 2012-June 2013) 50 consecutive comatose CA patients treated with TH (33 °C, 24 h). Quantitative PLR (expressed as the % of pupillary response to a calibrated light stimulus) and standard PLR were measured at day 1 (TH and sedation; on average 16 h after CA) and day 2 (normothermia, off sedation: on average 46 h after CA). Neurological outcome was assessed at 90 days with Cerebral Performance Categories (CPC), dichotomized as good (CPC 1-2) versus poor (CPC 3-5). Predictive performance was analyzed using area under the ROC curves (AUC). RESULTS: Patients with good outcome [n = 23 (46 %)] had higher quantitative PLR than those with poor outcome [n = 27; 16 (range 9-23) vs. 10 (1-30) % at day 1, and 20 (13-39) vs. 11 (1-55) % at day 2, both p < 0.001]. Best cut-off for outcome prediction of quantitative PLR was <13 %. The AUC to predict poor outcome was higher for quantitative than for standard PLR at both time points (day 1, 0.79 vs. 0.56, p = 0.005; day 2, 0.81 vs. 0.64, p = 0.006). Prognostic accuracy of quantitative PLR was comparable to that of EEG and SSEP (0.81 vs. 0.80 and 0.73, respectively, both p > 0.20). CONCLUSIONS: Quantitative PLR is more accurate than standard PLR in predicting outcome of post-anoxic coma, irrespective of temperature and sedation, and has comparable prognostic accuracy than EEG and SSEP.

Early multimodal outcome prediction after cardiac arrest in patients treated with hypothermia.

OBJECTIVES: Therapeutic hypothermia and pharmacological sedation may influence outcome prediction after cardiac arrest. The use of a multimodal approach, including clinical examination, electroencephalography, somatosensory-evoked potentials, and serum neuron-specific enolase, is recommended; however, no study examined the comparative performance of these predictors or addressed their optimal combination. DESIGN: Prospective cohort study. SETTING: Adult ICU of an academic hospital. PATIENTS: One hundred thirty-four consecutive adults treated with therapeutic hypothermia after cardiac arrest. MEASUREMENTS AND MAIN RESULTS: Variables related to the cardiac arrest (cardiac rhythm, time to return of spontaneous circulation), clinical examination (brainstem reflexes and myoclonus), electroencephalography reactivity during therapeutic hypothermia, somatosensory-evoked potentials, and serum neuron-specific enolase. Models to predict clinical outcome at 3 months (assessed using the Cerebral Performance Categories: 5 = death; 3-5 = poor recovery) were evaluated using ordinal logistic regressions and receiving operator characteristic curves. Seventy-two patients (54%) had a poor outcome (of whom, 62 died), and 62 had a good outcome. Multivariable ordinal logistic regression identified absence of electroencephalography reactivity (p < 0.001), incomplete recovery of brainstem reflexes in normothermia (p = 0.013), and neuron-specific enolase higher than 33 μg/L (p = 0.029), but not somatosensory-evoked potentials, as independent predictors of poor outcome. The combination of clinical examination, electroencephalography reactivity, and neuron-specific enolase yielded the best predictive performance (receiving operator characteristic areas: 0.89 for mortality and 0.88 for poor outcome), with 100% positive predictive value. Addition of somatosensory-evoked potentials to this model did not improve prognostic accuracy. CONCLUSIONS: Combination of clinical examination, electroencephalography reactivity, and serum neuron-specific enolase offers the best outcome predictive performance for prognostication of early postanoxic coma, whereas somatosensory-evoked potentials do not add any complementary information. Although prognostication of poor outcome seems excellent, future studies are needed to further improve prediction of good prognosis, which still remains inaccurate.

Predicting Adverse Events in Children with Fever and Chemotherapy-Induced Neutropenia. Results of the Prospective Multicenter SPOG 2003 FN Study

PURPOSE To develop a score predicting the risk of adverse events (AEs) in pediatric patients with cancer who experience fever and neutropenia (FN) and to evaluate its performance. PATIENTS AND METHODS Pediatric patients with cancer presenting with FN induced by nonmyeloablative chemotherapy were observed in a prospective multicenter study. A score predicting the risk of future AEs (ie, serious medical complication, microbiologically defined infection, radiologically confirmed pneumonia) was developed from a multivariate mixed logistic regression model. Its cross-validated predictive performance was compared with that of published risk prediction rules. Results An AE was reported in 122 (29%) of 423 FN episodes. In 57 episodes (13%), the first AE was known only after reassessment after 8 to 24 hours of inpatient management. Predicting AE at reassessment was better than prediction at presentation with FN. A differential leukocyte count did not increase the predictive performance. The score predicting future AE in 358 episodes without known AE at reassessment used the following four variables: preceding chemotherapy more intensive than acute lymphoblastic leukemia maintenance (weight = 4), hemoglobin > or = 90 g/L (weight = 5), leukocyte count less than 0.3 G/L (weight = 3), and platelet count less than 50 G/L (weight = 3). A score (sum of weights) > or = 9 predicted future AEs. The cross-validated performance of this score exceeded the performance of published risk prediction rules. At an overall sensitivity of 92%, 35% of the episodes were classified as low risk, with a specificity of 45% and a negative predictive value of 93%. CONCLUSION This score, based on four routinely accessible characteristics, accurately identifies pediatric patients with cancer with FN at risk for AEs after reassessment.

Digital soil mapping: strategy for data pre-processing

The region of greatest variability on soil maps is along the edge of their polygons, causing disagreement among pedologists about the appropriate description of soil classes at these locations. The objective of this work was to propose a strategy for data pre-processing applied to digital soil mapping (DSM). Soil polygons on a training map were shrunk by 100 and 160 m. This strategy prevented the use of covariates located near the edge of the soil classes for the Decision Tree (DT) models. Three DT models derived from eight predictive covariates, related to relief and organism factors sampled on the original polygons of a soil map and on polygons shrunk by 100 and 160 m were used to predict soil classes. The DT model derived from observations 160 m away from the edge of the polygons on the original map is less complex and has a better predictive performance.

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).

Measuring potentially avoidable hospital readmissions.

The objectives of this study were to develop a computerized method to screen for potentially avoidable hospital readmissions using routinely collected data and a prediction model to adjust rates for case mix. We studied hospital information system data of a random sample of 3,474 inpatients discharged alive in 1997 from a university hospital and medical records of those (1,115) readmitted within 1 year. The gold standard was set on the basis of the hospital data and medical records: all readmissions were classified as foreseen readmissions, unforeseen readmissions for a new affection, or unforeseen readmissions for a previously known affection. The latter category was submitted to a systematic medical record review to identify the main cause of readmission. Potentially avoidable readmissions were defined as a subgroup of unforeseen readmissions for a previously known affection occurring within an appropriate interval, set to maximize the chance of detecting avoidable readmissions. The computerized screening algorithm was strictly based on routine statistics: diagnosis and procedures coding and admission mode. The prediction was based on a Poisson regression model. There were 454 (13.1%) unforeseen readmissions for a previously known affection within 1 year. Fifty-nine readmissions (1.7%) were judged avoidable, most of them occurring within 1 month, which was the interval used to define potentially avoidable readmissions (n = 174, 5.0%). The intra-sample sensitivity and specificity of the screening algorithm both reached approximately 96%. Higher risk for potentially avoidable readmission was associated with previous hospitalizations, high comorbidity index, and long length of stay; lower risk was associated with surgery and delivery. The model offers satisfactory predictive performance and a good medical plausibility. The proposed measure could be used as an indicator of inpatient care outcome. However, the instrument should be validated using other sets of data from various hospitals.

What matters for predicting spatial distributions of trees: techniques, data, or species' characteristics?

Data characteristics and species traits are expected to influence the accuracy with which species' distributions can be modeled and predicted. We compare 10 modeling techniques in terms of predictive power and sensitivity to location error, change in map resolution, and sample size, and assess whether some species traits can explain variation in model performance. We focused on 30 native tree species in Switzerland and used presence-only data to model current distribution, which we evaluated against independent presence-absence data. While there are important differences between the predictive performance of modeling methods, the variance in model performance is greater among species than among techniques. Within the range of data perturbations in this study, some extrinsic parameters of data affect model performance more than others: location error and sample size reduced performance of many techniques, whereas grain had little effect on most techniques. No technique can rescue species that are difficult to predict. The predictive power of species-distribution models can partly be predicted from a series of species characteristics and traits based on growth rate, elevational distribution range, and maximum elevation. Slow-growing species or species with narrow and specialized niches tend to be better modeled. The Swiss presence-only tree data produce models that are reliable enough to be useful in planning and management applications.

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.

Comparing potentially avoidable hospitalization rates related to ambulatory care sensitive conditions in Switzerland: the need to refine the definition of health conditions and to adjust for population health status.

BACKGROUND: Regional rates of hospitalization for ambulatory care sensitive conditions (ACSC) are used to compare the availability and quality of ambulatory care but the risk adjustment for population health status is often minimal. The objectives of the study was to examine the impact of more extensive risk adjustment on regional comparisons and to investigate the relationship between various area-level factors and the properly adjusted rates. METHODS: Our study is an observational study based on routine data of 2 million anonymous insured in 26 Swiss cantons followed over one or two years. A binomial negative regression was modeled with increasingly detailed information on health status (age and gender only, inpatient diagnoses, outpatient conditions inferred from dispensed drugs and frequency of physician visits). Hospitalizations for ACSC were identified from principal diagnoses detecting 19 conditions, with an updated list of ICD-10 diagnostic codes. Co-morbidities and surgical procedures were used as exclusion criteria to improve the specificity of the detection of potentially avoidable hospitalizations. The impact of the adjustment approaches was measured by changes in the standardized ratios calculated with and without other data besides age and gender. RESULTS: 25% of cases identified by inpatient main diagnoses were removed by applying exclusion criteria. Cantonal ACSC hospitalizations rates varied from to 1.4 to 8.9 per 1,000 insured, per year. Morbidity inferred from diagnoses and drugs dramatically increased the predictive performance, the greatest effect found for conditions linked to an ACSC. More visits were associated with fewer PAH although very high users were at greater risk and subjects who had not consulted at negligible risk. By maximizing health status adjustment, two thirds of the cantons changed their adjusted ratio by more than 10 percent. Cantonal variations remained substantial but unexplained by supply or demand. CONCLUSION: Additional adjustment for health status is required when using ACSC to monitor ambulatory care. Drug-inferred morbidities are a promising approach.

THE PREVALENCE AND PREDICTION OF PULMONARY FIBROSIS IN AN

This thesis describes an ancillary project to the Early Diagnosis of Mesothelioma and Lung Cancer in Prior Asbestos Workers study and was conducted to determine the effects of asbestos exposure, pulmonary function and cigarette smoking in the prediction of pulmonary fibrosis. 613 workers who were occupationally exposed to asbestos for an average of 25.9 (SD=14.69) years were sampled from Sarnia, Ontario. A structured questionnaire was administered during a face-to-face interview along with a low-dose computed tomography (LDCT) of the thorax. Of them, 65 workers (10.7%, 95%CI 8.12—12.24) had LDCT-detected pulmonary fibrosis. The model predicting fibrosis included the variables age, smoking (dichotomized), post FVC % splines and post- FEV1% splines. This model had a receiver operator characteristic area under the curve of 0.738. The calibration of the model was evaluated with R statistical program and the bootstrap optimism-corrected calibration slope was 0.692. Thus, our model demonstrated moderate predictive performance.

Limited sampling strategies for estimation of cyclosporine exposure in pediatric hematopoietic stem cell transplant recipients : methodological improvement and introduction of sampling time deviation analysis.

Le suivi thérapeutique est recommandé pour l’ajustement de la dose des agents immunosuppresseurs. La pertinence de l’utilisation de la surface sous la courbe (SSC) comme biomarqueur dans l’exercice du suivi thérapeutique de la cyclosporine (CsA) dans la transplantation des cellules souches hématopoïétiques est soutenue par un nombre croissant d’études. Cependant, pour des raisons intrinsèques à la méthode de calcul de la SSC, son utilisation en milieu clinique n’est pas pratique. Les stratégies d’échantillonnage limitées, basées sur des approches de régression (R-LSS) ou des approches Bayésiennes (B-LSS), représentent des alternatives pratiques pour une estimation satisfaisante de la SSC. Cependant, pour une application efficace de ces méthodologies, leur conception doit accommoder la réalité clinique, notamment en requérant un nombre minimal de concentrations échelonnées sur une courte durée d’échantillonnage. De plus, une attention particulière devrait être accordée à assurer leur développement et validation adéquates. Il est aussi important de mentionner que l’irrégularité dans le temps de la collecte des échantillons sanguins peut avoir un impact non-négligeable sur la performance prédictive des R-LSS. Or, à ce jour, cet impact n’a fait l’objet d’aucune étude. Cette thèse de doctorat se penche sur ces problématiques afin de permettre une estimation précise et pratique de la SSC. Ces études ont été effectuées dans le cadre de l’utilisation de la CsA chez des patients pédiatriques ayant subi une greffe de cellules souches hématopoïétiques. D’abord, des approches de régression multiple ainsi que d’analyse pharmacocinétique de population (Pop-PK) ont été utilisées de façon constructive afin de développer et de valider adéquatement des LSS. Ensuite, plusieurs modèles Pop-PK ont été évalués, tout en gardant à l’esprit leur utilisation prévue dans le contexte de l’estimation de la SSC. Aussi, la performance des B-LSS ciblant différentes versions de SSC a également été étudiée. Enfin, l’impact des écarts entre les temps d’échantillonnage sanguins réels et les temps nominaux planifiés, sur la performance de prédiction des R-LSS a été quantifié en utilisant une approche de simulation qui considère des scénarios diversifiés et réalistes représentant des erreurs potentielles dans la cédule des échantillons sanguins. Ainsi, cette étude a d’abord conduit au développement de R-LSS et B-LSS ayant une performance clinique satisfaisante, et qui sont pratiques puisqu’elles impliquent 4 points d’échantillonnage ou moins obtenus dans les 4 heures post-dose. Une fois l’analyse Pop-PK effectuée, un modèle structural à deux compartiments avec un temps de délai a été retenu. Cependant, le modèle final - notamment avec covariables - n’a pas amélioré la performance des B-LSS comparativement aux modèles structuraux (sans covariables). En outre, nous avons démontré que les B-LSS exhibent une meilleure performance pour la SSC dérivée des concentrations simulées qui excluent les erreurs résiduelles, que nous avons nommée « underlying AUC », comparée à la SSC observée qui est directement calculée à partir des concentrations mesurées. Enfin, nos résultats ont prouvé que l’irrégularité des temps de la collecte des échantillons sanguins a un impact important sur la performance prédictive des R-LSS; cet impact est en fonction du nombre des échantillons requis, mais encore davantage en fonction de la durée du processus d’échantillonnage impliqué. Nous avons aussi mis en évidence que les erreurs d’échantillonnage commises aux moments où la concentration change rapidement sont celles qui affectent le plus le pouvoir prédictif des R-LSS. Plus intéressant, nous avons mis en exergue que même si différentes R-LSS peuvent avoir des performances similaires lorsque basées sur des temps nominaux, leurs tolérances aux erreurs des temps d’échantillonnage peuvent largement différer. En fait, une considération adéquate de l'impact de ces erreurs peut conduire à une sélection et une utilisation plus fiables des R-LSS. Par une investigation approfondie de différents aspects sous-jacents aux stratégies d’échantillonnages limités, cette thèse a pu fournir des améliorations méthodologiques notables, et proposer de nouvelles voies pour assurer leur utilisation de façon fiable et informée, tout en favorisant leur adéquation à la pratique clinique.

Models to predict the distribution and abundance of breeding ducks in Canada

Detailed knowledge of waterfowl abundance and distribution across Canada is lacking, which limits our ability to effectively conserve and manage their populations. We used 15 years of data from an aerial transect survey to model the abundance of 17 species or species groups of ducks within southern and boreal Canada. We included 78 climatic, hydrological, and landscape variables in Boosted Regression Tree models, allowing flexible response curves and multiway interactions among variables. We assessed predictive performance of the models using four metrics and calculated uncertainty as the coefficient of variation of predictions across 20 replicate models. Maps of predicted relative abundance were generated from resulting models, and they largely match spatial patterns evident in the transect data. We observed two main distribution patterns: a concentrated prairie-parkland distribution and a more dispersed pan-Canadian distribution. These patterns were congruent with the relative importance of predictor variables and model evaluation statistics among the two groups of distributions. Most species had a hydrological variable as the most important predictor, although the specific hydrological variable differed somewhat among species. In some cases, important variables had clear ecological interpretations, but in some instances, e.g., topographic roughness, they may simply reflect chance correlations between species distributions and environmental variables identified by the model-building process. Given the performance of our models, we suggest that the resulting prediction maps can be used in future research and to guide conservation activities, particularly within the bounds of the survey area.

A hierarchical Bayesian model for predicting the functional consequences of amino-acid polymorphisms

Genetic polymorphisms in deoxyribonucleic acid coding regions may have a phenotypic effect on the carrier, e.g. by influencing susceptibility to disease. Detection of deleterious mutations via association studies is hampered by the large number of candidate sites; therefore methods are needed to narrow down the search to the most promising sites. For this, a possible approach is to use structural and sequence-based information of the encoded protein to predict whether a mutation at a particular site is likely to disrupt the functionality of the protein itself. We propose a hierarchical Bayesian multivariate adaptive regression spline (BMARS) model for supervised learning in this context and assess its predictive performance by using data from mutagenesis experiments on lac repressor and lysozyme proteins. In these experiments, about 12 amino-acid substitutions were performed at each native amino-acid position and the effect on protein functionality was assessed. The training data thus consist of repeated observations at each position, which the hierarchical framework is needed to account for. The model is trained on the lac repressor data and tested on the lysozyme mutations and vice versa. In particular, we show that the hierarchical BMARS model, by allowing for the clustered nature of the data, yields lower out-of-sample misclassification rates compared with both a BMARS and a frequen-tist MARS model, a support vector machine classifier and an optimally pruned classification tree.