Interpretable Prediction Models Research Articles

Feature Ranking in Predictive Models for Hospital-Acquired Acute Kidney Injury

Acute Kidney Injury (AKI) is a common complication encountered among hospitalized patients, imposing significantly increased cost, morbidity, and mortality. Early prediction of AKI has profound clinical implications because currently no treatment exists for AKI once it develops. Feature selection (FS) is an essential process for building accurate and interpretable prediction models, but to our best knowledge no study has investigated the robustness and applicability of such selection process for AKI. In this study, we compared eight widely-applied FS methods for AKI prediction using nine-years of electronic medical records (EMR) and examined heterogeneity in feature rankings produced by the methods. FS methods were compared in terms of stability with respect to data sampling variation, similarity between selection results, and AKI prediction performance. Prediction accuracy did not intrinsically guarantee the feature ranking stability. Across different FS methods, the prediction performance did not change significantly, while the importance rankings of features were quite different. A positive correlation was observed between the complexity of suitable FS method and sample size. This study provides several practical implications, including recognizing the importance of feature stability as it is desirable for model reproducibility, identifying important AKI risk factors for further investigation, and facilitating early prediction of AKI.

A Clinical Interpretable Approach Applied to Cardiovascular Risk Assessment.

The effectiveness of predictive models in supporting the Clinical Decision is closely related with their clinical interpretability, i.e.the model should provide clear information on how to reach a specific classification/decision. In fact, the development of interpretable and accurate predictive models assumes a key importance as these tools can be very useful in Clinical Decision Support Systems (CDSS). The development of those models may comprise two main perspectives; existent clinical knowledge (clinical expert knowledge, clinical guidelines, current models, etc.) as well as data driven approaches able to extract (new) knowledge from recent clinical datasets. This work focuses in knowledge extraction from recent datasets (data driven) based on computational intelligence techniques. The main hypothesis that supports this work is that individuals with similar characteristics present a similar risk prof ile. Thus, this work addresses the development of stratification models able to learn distinct groups (or classes) of subjects assessing the similarity between characterizing variables. In particular, in the current study a data-driven supervised cluster approach is proposed aiming the derivation of meaningful rules directly from the dataset. The validation was performed based on the largest Portuguese coronary artery disease patient's dataset, provided by the Portuguese Society of Cardiology and comprising 13902 acute coronary syndrome patients. The goal was to assess the risk of death 30 days after admission. The models' performance was assessed through the sensitivity, specificity and geometric mean values. The obtained results show the potential of this approach, as they represent an acceptable performance (GM= 72%) while the clinical interpretability of the model is assured through the derived rules. Despite the achieved results, there are several research directions to be followed in order to enhance this work.

An analytic approach for interpretable predictive models in high-dimensional data in the presence of interactions with exposures.

ABSTRACTPredicting a phenotype and understanding which variables improve that prediction are two very challenging and overlapping problems in the analysis of high‐dimensional (HD) data such as those arising from genomic and brain imaging studies. It is often believed that the number of truly important predictors is small relative to the total number of variables, making computational approaches to variable selection and dimension reduction extremely important. To reduce dimensionality, commonly used two‐step methods first cluster the data in some way, and build models using cluster summaries to predict the phenotype. It is known that important exposure variables can alter correlation patterns between clusters of HD variables, that is, alter network properties of the variables. However, it is not well understood whether such altered clustering is informative in prediction. Here, assuming there is a binary exposure with such network‐altering effects, we explore whether the use of exposure‐dependent clustering relationships in dimension reduction can improve predictive modeling in a two‐step framework. Hence, we propose a modeling framework called ECLUST to test this hypothesis, and evaluate its performance through extensive simulations. With ECLUST, we found improved prediction and variable selection performance compared to methods that do not consider the environment in the clustering step, or to methods that use the original data as features. We further illustrate this modeling framework through the analysis of three data sets from very different fields, each with HD data, a binary exposure, and a phenotype of interest. Our method is available in the eclust CRAN package.

Machine Learning of the Progression of Intermediate Age-Related Macular Degeneration Based on OCT Imaging.

To develop a data-driven interpretable predictive model of incoming drusen regression as a sign of disease activity and identify optical coherence tomography (OCT) biomarkers associated with its risk in intermediate age-related macular degeneration (AMD). Patients with AMD were observed every 3 months, using Spectralis OCT imaging, for a minimum duration of 12 months and up to a period of 60 months. Segmentation of drusen and the overlying layers was obtained using a graph-theoretic method, and the hyperreflective foci were segmented using a voxel classification method. Automated image analysis steps were then applied to identify and characterize individual drusen at baseline, and their development was monitored at every follow-up visit. Finally, a machine learning method based on a sparse Cox proportional hazard regression was developed to estimate a risk score and predict the incoming regression of individual drusen. The predictive model was trained and evaluated on a longitudinal dataset of 61 eyes from 38 patients using cross-validation. The mean follow-up time was 37.8 ± 13.8 months. A total of 944 drusen were identified at baseline, out of which 249 (26%) regressed during follow-up. The prediction performance was evaluated as area under the curve (AUC) for different time periods. Prediction within the first 2 years achieved an AUC of 0.75. The predictive model proposed in this study represents a promising step toward image-guided prediction of AMD progression. Machine learning is expected to accelerate and contribute to the development of new therapeutics that delay the progression of AMD.

Predictive computational phenotyping and biomarker discovery using reference-free genome comparisons.

BackgroundThe identification of genomic biomarkers is a key step towards improving diagnostic tests and therapies. We present a reference-free method for this task that relies on a k-mer representation of genomes and a machine learning algorithm that produces intelligible models. The method is computationally scalable and well-suited for whole genome sequencing studies.ResultsThe method was validated by generating models that predict the antibiotic resistance of C. difficile, M. tuberculosis, P. aeruginosa, and S. pneumoniae for 17 antibiotics. The obtained models are accurate, faithful to the biological pathways targeted by the antibiotics, and they provide insight into the process of resistance acquisition. Moreover, a theoretical analysis of the method revealed tight statistical guarantees on the accuracy of the obtained models, supporting its relevance for genomic biomarker discovery.ConclusionsOur method allows the generation of accurate and interpretable predictive models of phenotypes, which rely on a small set of genomic variations. The method is not limited to predicting antibiotic resistance in bacteria and is applicable to a variety of organisms and phenotypes. Kover, an efficient implementation of our method, is open-source and should guide biological efforts to understand a plethora of phenotypes (http://github.com/aldro61/kover/).Electronic supplementary materialThe online version of this article (doi:10.1186/s12864-016-2889-6) contains supplementary material, which is available to authorized users.

Prognostic value of different biomarkers for cardiovascular death in unselected older patients in the emergency department.

Risk stratification of elderly patients presenting with heart failure (HF) to an emergency department (ED) is an unmet challenge. We prospectively investigated the prognostic performance of different biomarkers in unselected older patients in the ED. We consecutively enrolled 302 non-surgical patients ⩾70 years presenting to the ED with a wide range of cardiovascular and non-cardiovascular comorbid conditions. N-terminal pro-B-type natriuretic peptide (NT-proBNP), mid-regional pro-adrenomedullin (MR-proADM), mid-regional pro-atrial natriuretic peptide (MR-proANP), C-terminal pro-endothelin-1 (CT-proET-1), ultrasensitive C-terminal pro-arginine-vasopressin (Copeptin-us) and high-sensitivity cardiac troponin T (hs-cTnT) were measured at admission. Two cardiologists independently adjudicated the final diagnosis of HF after reviewing all available baseline data using circulating NT-proBNP levels. A final diagnosis of HF was found in 120 (40%) of the 302 patients. All patients were followed up for cardiovascular death within the following 12 months. In order to test the prognostic performance of the investigated biomarkers we used boosting models with age and sex as mandatory covariates. Boosting is a statistical learning technique with built-in variable selection developed to obtain sparse and interpretable prediction models. Follow-up was 100% complete. During a median follow-up time of 225 days (interquartile range (IQR) 156-319 days), 30 (9.9%) of 302 patients (aged 81±6 years) had cardiovascular deaths. Of these 30 patients, 21 had HF and nine had no HF diagnosed prior to admission. The boosting model selected MR-proADM and hs-cTNT as predictors of cardiovascular deaths. The median values of MR-proADM and hs-cTnT at presentation were significantly higher in patients with cardiovascular deaths compared to surviving patients during follow-up (2.56 nmol/L (IQR 1.62-4.48) vs. 1.11 nmol/L (IQR 0.83-1.80), P<0.001 and 81 ng/L (IQR 38-340) vs. 17 ng/L (IQR 0.9-38), P=0.004). One unit increase in the log-transformed MR-proADM levels was associated with a 1.99-fold risk of death (95% confidence interval (CI) 1.61-2.45, P<0.001). The second marker, hs-cTnT, showed an increased predicted risk but was not significantly correlated to event-free survival (hazard ratio 3.22, 95% CI 0.97-10.68, P=0.056). Within different biomarkers, MR-proADM was the only predictor of cardiovascular deaths in unselected older patients presenting to the ED.

Toward interpretable predictive models in B2B recommender systems

Recommender systems are becoming increasingly important for businesses because they can help companies offer personalized product recommendations to customers. There have been many acknowledged recognized successes of consumer-oriented recommender systems, particularly in e-commerce. In this work, we describe our experiences building a business-to-business (B2B) recommendation engine that matches company clients to internal company products. The underlying pairing of clients and products is based on co-clustering principles and helps reveal potential future purchases. Unlike most consumer-oriented recommendation systems, our approach takes into account the need for interpretability. We do not only provide a cursory explanation, as offered in most traditional recommender systems. In our approach, the recipient of the generated recommendations are sales and marketing teams; thus, we offer a detailed reasoning in straightforward English that considers multiple aspects regarding why a client may be a suitable match for the particular offering. Finally, we analyze the outcome of a country-wide deployment of the proposed methodology for selected IBM sales teams.

Supervised item response models for informative prediction

Supporting human decision-making is a major goal of data mining. The more decision-making is critical, the more interpretability is required in the predictive model. This paper proposes a new framework to build a fully interpretable predictive model for questionnaire data, while maintaining a reasonable prediction accuracy with regard to the final outcome. Such a model has applications in project risk assessment, in healthcare, in social studies, and, presumably, in any real-world application that relies on questionnaire data for informative and accurate prediction. Our framework is inspired by models in item response theory (IRT), which were originally developed in psychometrics with applications to standardized academic tests. We extend these models, which are essentially unsupervised, to the supervised setting. For model estimation, we introduce a new iterative algorithm by combining Gauss---Hermite quadrature with an expectation---maximization algorithm. The learned probabilistic model is linked to the metric learning framework for informative and accurate prediction. The model is validated by three real-world data sets: Two are from information technology project failure prediction and the other is an international social survey about people's happiness. To the best of our knowledge, this is the first work that leverages the IRT framework to provide informative and accurate prediction on ordinal questionnaire data.

Building interpretable predictive models for pediatric hospital readmission using Tree-Lasso logistic regression.

Quantification and early identification of unplanned readmission risk have the potential to improve the quality of care during hospitalization and after discharge. However, high dimensionality, sparsity, and class imbalance of electronic health data and the complexity of risk quantification, challenge the development of accurate predictive models. Predictive models require a certain level of interpretability in order to be applicable in real settings and create actionable insights. This paper aims to develop accurate and interpretable predictive models for readmission in a general pediatric patient population, by integrating a data-driven model (sparse logistic regression) and domain knowledge based on the international classification of diseases 9th-revision clinical modification (ICD-9-CM) hierarchy of diseases. Additionally, we propose a way to quantify the interpretability of a model and inspect the stability of alternative solutions. The analysis was conducted on >66,000 pediatric hospital discharge records from California, State Inpatient Databases, Healthcare Cost and Utilization Project between 2009 and 2011. We incorporated domain knowledge based on the ICD-9-CM hierarchy in a data driven, Tree-Lasso regularized logistic regression model, providing the framework for model interpretation. This approach was compared with traditional Lasso logistic regression resulting in models that are easier to interpret by fewer high-level diagnoses, with comparable prediction accuracy. The results revealed that the use of a Tree-Lasso model was as competitive in terms of accuracy (measured by area under the receiver operating characteristic curve-AUC) as the traditional Lasso logistic regression, but integration with the ICD-9-CM hierarchy of diseases provided more interpretable models in terms of high-level diagnoses. Additionally, interpretations of models are in accordance with existing medical understanding of pediatric readmission. Best performing models have similar performances reaching AUC values 0.783 and 0.779 for traditional Lasso and Tree-Lasso, respectfully. However, information loss of Lasso models is 0.35 bits higher compared to Tree-Lasso model. We propose a method for building predictive models applicable for the detection of readmission risk based on Electronic Health records. Integration of domain knowledge (in the form of ICD-9-CM taxonomy) and a data-driven, sparse predictive algorithm (Tree-Lasso Logistic Regression) resulted in an increase of interpretability of the resulting model. The models are interpreted for the readmission prediction problem in general pediatric population in California, as well as several important subpopulations, and the interpretations of models comply with existing medical understanding of pediatric readmission. Finally, quantitative assessment of the interpretability of the models is given, that is beyond simple counts of selected low-level features.

Pattern-Aided Regression Modeling and Prediction Model Analysis

This paper first introduces pattern aided regression (PXR) models, a new type of regression models designed to represent accurate and interpretable prediction models. This was motivated by two observations: (1) Regression modeling applications often involve complex diverse predictor-response relationships , which occur when the optimal regression models (of given regression model type) fitting two or more distinct logical groups of data are highly different. (2) State-of-the-art regression methods are often unable to adequately model such relationships. This paper defines PXR models using several patterns and local regression models, which respectively serve as logical and behavioral characterizations of distinct predictor-response relationships. The paper also introduces a contrast pattern aided regression (CPXR) method, to build accurate PXR models. In experiments, the PXR models built by CPXR are very accurate in general, often outperforming state-of-the-art regression methods by big margins. Usually using (a) around seven simple patterns and (b) linear local regression models, those PXR models are easy to interpret; in fact, their complexity is just a bit higher than that of (piecewise) linear regression models and is significantly lower than that of traditional ensemble based regression models. CPXR is especially effective for high-dimensional data. The paper also discusses how to use CPXR methodology for analyzing prediction models and correcting their prediction errors.

Accounting for genetic architecture improves sequence based genomic prediction for a Drosophila fitness trait.

The ability to predict quantitative trait phenotypes from molecular polymorphism data will revolutionize evolutionary biology, medicine and human biology, and animal and plant breeding. Efforts to map quantitative trait loci have yielded novel insights into the biology of quantitative traits, but the combination of individually significant quantitative trait loci typically has low predictive ability. Utilizing all segregating variants can give good predictive ability in plant and animal breeding populations, but gives little insight into trait biology. Here, we used the Drosophila Genetic Reference Panel to perform both a genome wide association analysis and genomic prediction for the fitness-related trait chill coma recovery time. We found substantial total genetic variation for chill coma recovery time, with a genetic architecture that differs between males and females, a small number of molecular variants with large main effects, and evidence for epistasis. Although the top additive variants explained 36% (17%) of the genetic variance among lines in females (males), the predictive ability using genomic best linear unbiased prediction and a relationship matrix using all common segregating variants was very low for females and zero for males. We hypothesized that the low predictive ability was due to the mismatch between the infinitesimal genetic architecture assumed by the genomic best linear unbiased prediction model and the true genetic architecture of chill coma recovery time. Indeed, we found that the predictive ability of the genomic best linear unbiased prediction model is markedly improved when we combine quantitative trait locus mapping with genomic prediction by only including the top variants associated with main and epistatic effects in the relationship matrix. This trait-associated prediction approach has the advantage that it yields biologically interpretable prediction models.

Development of Interpretable Predictive Models for BPH and Prostate Cancer.

BACKGROUNDTraditional methods for deciding whether to recommend a patient for a prostate biopsy are based on cut-off levels of stand-alone markers such as prostate-specific antigen (PSA) or any of its derivatives. However, in the last decade we have seen the increasing use of predictive models that combine, in a non-linear manner, several predictives that are better able to predict prostate cancer (PC), but these fail to help the clinician to distinguish between PC and benign prostate hyperplasia (BPH) patients. We construct two new models that are capable of predicting both PC and BPH.METHODSAn observational study was performed on 150 patients with PSA ≥3 ng/mL and age >50 years. We built a decision tree and a logistic regression model, validated with the leave-one-out methodology, in order to predict PC or BPH, or reject both.RESULTSStatistical dependence with PC and BPH was found for prostate volume (P-value < 0.001), PSA (P-value < 0.001), international prostate symptom score (IPSS; P-value < 0.001), digital rectal examination (DRE; P-value < 0.001), age (P-value < 0.002), antecedents (P-value < 0.006), and meat consumption (P-value < 0.08). The two predictive models that were constructed selected a subset of these, namely, volume, PSA, DRE, and IPSS, obtaining an area under the ROC curve (AUC) between 72% and 80% for both PC and BPH prediction.CONCLUSIONPSA and volume together help to build predictive models that accurately distinguish among PC, BPH, and patients without any of these pathologies. Our decision tree and logistic regression models outperform the AUC obtained in the compared studies. Using these models as decision support, the number of unnecessary biopsies might be significantly reduced.

Computer-aided diagnostics in digital pathology: automated evaluation of early-phase pancreatic cancer in mice.

Digital pathology diagnostics are often based on subjective qualitative measures. A murine model of early-phase pancreatic ductal adenocarcinoma provides a controlled environment with a priori knowledge of the genetic mutation and stage of the disease. Use of this model enables the application of supervised learning methods to digital pathology. A computerized diagnostics system for histological detection of pancreatic adenocarcinoma was developed and tested. METHODS : Pathological H&E-stained specimens with early pancreatic lesions were identified and evaluated with a system that models cancer detection using a top-down object learning paradigm, mimicking the way a pathologist learns. First, the dominant primitives were identified and segmented in the images, i.e., the ducts, nuclei and tumor stroma. A boost-based machine learning technique was used for duct segmentation, classification and outlier pruning. Second, a set of morphological features traditionally used for cancer diagnosis which provides quantitative image features was employed to quantify subtle findings such as duct deformation and nuclei malformations. Finally, a visually interpretable predictive model was trained to distinguish between normal tissue and premalignant cancer lesions, given ground truth samples. RESULTS : A predictive success rate of 92% was achieved using tenfold cross-validation and 93% on an independent test set. Comparison was made with state-of-the-art classification algorithms that are not interpretable as visible features yielded the contribution of individual primitive features to the prediction outcome. Quantitative image analysis and classification were successful in preclinical histology diagnosis for early-stage pancreatic adenocarcinoma. The usage of annotated contours coupled with interpretable supervised learning methods and outlier pruning can be adapted to other medical imaging tasks. The usage of interpretable supervised learning techniques may improve the success of CAD in histopathological diagnosis.

Tree-Structured Mixed-Effects Regression Modeling for Longitudinal Data

Tree-structured models have been widely used because they function as interpretable prediction models that offer easy data visualization. A number of tree algorithms have been developed for univariate response data and can be extended to analyze multivariate response data. We propose a tree algorithm by combining the merits of a tree-based model and a mixed-effects model for longitudinal data. We alleviate variable selection bias through residual analysis, which is used to solve problems that exhaustive search approaches suffer from, such as undue preference to split variables with more possible splits, expensive computational cost, and end-cut preference. Most importantly, our tree algorithm discovers trends over time on each of the subspaces from recursive partitioning, while other tree algorithms predict responses. We investigate the performance of our algorithm with both simulation and real data studies. We also develop an R package melt that can be used conveniently and freely. Additional results are provided as online supplementary material.

An iterative approach to build relevant ontology-aware data-driven models

In many fields involving complex environments or living organisms, data-driven models are useful to make simulations in order to extrapolate costly experiments and to design decision-support tools. Learning methods can be used to build interpretable models from data. However, to be really useful, such models must be trusted by their users. From this perspective, the domain expert knowledge can be collected and modeled to help guiding the learning process and to increase the confidence in the resulting models, as well as their relevance. Another issue is to design relevant ontologies to formalize complex knowledge. Interpretable predictive models can help in this matter. In this paper, we propose a generic iterative approach to design ontology-aware and relevant data-driven models. It is based upon an ontology to model the domain knowledge and a learning method to build the interpretable models (decision trees in this paper). Subjective and objective evaluations are both involved in the process. A case study in the domain of Food Industry demonstrates the interest of this approach.

Abstract 4760: Molecular classification and risk-stratification in medulloblastoma using an integrative genomic approach

Abstract Purpose. Medulloblastomas are heterogeneous tumors that collectively represent the most common malignant brain tumor in children. To understand the molecular characteristics underlying their heterogeneity and to identify whether such characteristics represent risk factors for patients with this disease, we performed an integrated genomic analysis of a large series of primary tumors. Patients and Methods. We profiled the mRNA transcriptome of 194 medulloblastomas and performed high-density SNP array and miRNA analysis on 115 and 98 of these, respectively. Unsupervised clustering analysis of mRNA expression data was used to identify molecular subgroups of medulloblastoma and DNA copy number, miRNA profiles, and clinical outcomes were analyzed for each. We then developed a Bayesian cumulative log odds model for predicting outcome, starting with ‘evidence’ provided by clinical features (metastasis and histology), then incrementally adding genomic evidence representing disease-subtype independent and dependent features. We then validated our results on independent test cohorts. Results. We identify 6 molecular subgroups of medulloblastoma, each with a unique combination of numerical and structural chromosomal aberrations that globally influence mRNA and miRNA expression. We reveal the relative contribution of each subgroup to clinical outcome as a whole and show that a previously unidentified molecular subgroup, characterized genetically by c-MYC copy number gains and transcriptionally by enrichment of photoreceptor pathways and increased miR-183∼96∼182 expression, is associated with significantly lower rates of event free and overall survival. Moreover, we show that integration of high-level clinical and genomic risk factors, including molecular subtype, can yield more comprehensive, accurate and biologically interpretable prediction models for outcome in medulloblastoma. We introduce a novel Bayesian nomogram indicating the amount of evidence that each feature contributes on a patient-by-patient basis. Conclusion. Our results detail the complex genomic heterogeneity of medulloblastomas, identify a previously unrecognized molecular subgroup with particularly poor clinical outcome and propose a Bayesian outcomes prediction model that outperforms the current clinical classification schema and previously published molecular markers proposed for medulloblastoma risk stratification. Citation Format: {Authors}. {Abstract title} [abstract]. In: Proceedings of the 102nd Annual Meeting of the American Association for Cancer Research; 2011 Apr 2-6; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2011;71(8 Suppl):Abstract nr 4760. doi:10.1158/1538-7445.AM2011-4760

Interpretable Predictive Models for Knowledge Discovery from Home-Care Electronic Health Records

The purpose of this methodological study was to compare methods of developing predictive rules that are parsimonious and clinically interpretable from electronic health record (EHR) home visit data, contrasting logistic regression with three data mining classification models. We address three problems commonly encountered in EHRs: the value of including clinically important variables with little variance, handling imbalanced datasets, and ease of interpretation of the resulting predictive models. Logistic regression and three classification models using Ripper, decision trees, and Support Vector Machines were applied to a case study for one outcome of improvement in oral medication management. Predictive rules for logistic regression, Ripper, and decision trees are reported and results compared using F-measures for data mining models and area under the receiver-operating characteristic curve for all models. The rules generated by the three classification models provide potentially novel insights into mining EHRs beyond those provided by standard logistic regression, and suggest steps for further study.

Predicting Relapse in Patients With Medulloblastoma by Integrating Evidence From Clinical and Genomic Features

Despite significant progress in the molecular understanding of medulloblastoma, stratification of risk in patients remains a challenge. Focus has shifted from clinical parameters to molecular markers, such as expression of specific genes and selected genomic abnormalities, to improve accuracy of treatment outcome prediction. Here, we show how integration of high-level clinical and genomic features or risk factors, including disease subtype, can yield more comprehensive, accurate, and biologically interpretable prediction models for relapse versus no-relapse classification. We also introduce a novel Bayesian nomogram indicating the amount of evidence that each feature contributes on a patient-by-patient basis. A Bayesian cumulative log-odds model of outcome was developed from a training cohort of 96 children treated for medulloblastoma, starting with the evidence provided by clinical features of metastasis and histology (model A) and incrementally adding the evidence from gene-expression-derived features representing disease subtype-independent (model B) and disease subtype-dependent (model C) pathways, and finally high-level copy-number genomic abnormalities (model D). The models were validated on an independent test cohort (n = 78). On an independent multi-institutional test data set, models A to D attain an area under receiver operating characteristic (au-ROC) curve of 0.73 (95% CI, 0.60 to 0.84), 0.75 (95% CI, 0.64 to 0.86), 0.80 (95% CI, 0.70 to 0.90), and 0.78 (95% CI, 0.68 to 0.88), respectively, for predicting relapse versus no relapse. The proposed models C and D outperform the current clinical classification schema (au-ROC, 0.68), our previously published eight-gene outcome signature (au-ROC, 0.71), and several new schemas recently proposed in the literature for medulloblastoma risk stratification.

Modelling species distribution in complex environments: an evaluation of predictive ability and reliability in five shorebird species

ABSTRACTAim Many wader populations around the world are declining as a consequence of habitat degradation or loss. It is therefore important to identify species‐specific habitat demands accurately and to define the important factors explaining species distribution, in order to develop tools that can be used in conservation planning. The aim of this study is to create reliable, functional and ecologically interpretable predictive distribution models for five breeding wader species.Location The archipelago of SW Finland in the Baltic Sea.Methods We used multivariate adaptive regression splines (MARS) to create single‐species and multiresponse distribution models based on 525 study islands and 12 abiotic and biotic environmental variables. Model evaluation was carried out on independent data not used in model building (100 + 116 islands). The models were tested for discrimination with receiver‐operating characteristic statistics and for calibration with Millers calibration statistics (MCS).Results The single‐species models for the turnstone (Arenaria interpres), redshank (Tringa totanus) and oystercatcher (Haematopus ostralegus) showed good predictive abilities, regarding both discrimination and calibration, when evaluated on independent data. The multiresponse models for the less prevalent species, common sandpiper (Actitis hypoleucos) and the common ringed plover (Charadrius hiaticula) had better discriminative abilities than the single‐species models. The most influential predictor overall was occurrence of small larids. Exposure, area of forest and low and flat areas were also important, as well as shore habitats.Main conclusions We found that the ability of MARS to fit non‐linear and multiresponse models makes it a useful method to quantitatively relate species occurrence to environmental characteristics of a complex environment.