Efficient penalized generalized linear mixed models for variable selection and genetic risk prediction in high-dimensional data.

MOTIVATION: Sparse regularized regression methods are now widely used in genome-wide association studies (GWAS) to address the multiple testing burden that limits discovery of potentially important predictors. Linear mixed models (LMMs) have become an attractive alternative to principal components (PCs) adjustment to account for population structure and relatedness in high-dimensional penalized models. However, their use in binary trait GWAS rely on the invalid assumption that the residual variance does not depend on the estimated regression coefficients. Moreover, LMMs use a single spectral decomposition of the covariance matrix of the responses, which is no longer possible in generalized linear mixed models (GLMMs). RESULTS: We introduce a new method called pglmm, a penalized GLMM that allows to simultaneously select genetic markers and estimate their effects, accounting for between-individual correlations and binary nature of the trait. We develop a computationally efficient algorithm based on penalized quasi-likelihood estimation that allows to scale regularized mixed models on high-dimensional binary trait GWAS. We show through simulations that when the dimensionality of the relatedness matrix is high, penalized LMM and logistic regression with PC adjustment fail to select important predictors, and have inferior prediction accuracy compared to pglmm. Further, we demonstrate through the analysis of two polygenic binary traits in a subset of 6731 related individuals from the UK Biobank data with 320K SNPs that our method can achieve higher predictive performance, while also selecting fewer predictors than a sparse regularized logistic lasso with PC adjustment. AVAILABILITY AND IMPLEMENTATION: Our Julia package PenalizedGLMM.jl is publicly available on github: https://github.com/julstpierre/PenalizedGLMM. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

CT-based Bosniak classification of cystic renal lesions: is version 2019 an improvement on version 2005?

OBJECTIVE: To compare the diagnostic performance and inter-reader agreement of the CT-based v2019 versus v2005 Bosniak classification systems for risk stratification of cystic renal lesions (CRL). METHODS: This retrospective study included adult patients with CRL identified on CT scan between 2005 and 2018. The reference standard was histopathology or a minimum 4-year imaging follow-up. The studies were reviewed independently by five readers (three senior, two junior), blinded to pathology results and imaging follow-up, who assigned Bosniak categories based on the 2005 and 2019 versions. Diagnostic performance of v2005 and v2019 Bosniak classifications for distinguishing benign from malignant lesions was calculated by dichotomizing CRL into the potential for ablative therapy (III-IV) or conservative management (I-IIF). Inter-reader agreement was calculated using Light's Kappa. RESULTS: One hundred thirty-nine patients with 149 CRL (33 malignant) were included. v2005 and v2019 Bosniak classifications achieved similar diagnostic performance with a sensitivity of 91% vs 91% and a specificity of 89% vs 88%, respectively. Inter-reader agreement for overall Bosniak category assignment was substantial for v2005 (κ = 0.78) and v2019 (κ = 0.75) between senior readers but decreased for v2019 when the Bosniak classification was dichotomized to conservative management (I-IIF) or ablative therapy (III-IV) (0.80 vs 0.71, respectively). For v2019, wall thickness was the morphological feature with the poorest inter-reader agreement (κ = 0.43 and 0.18 for senior and junior readers, respectively). CONCLUSION: No significant improvement in diagnostic performance and inter-reader agreement was shown between v2005 and v2019. The observed decrease in inter-reader agreement in v2019 when dichotomized according to management strategy may reflect the more stringent morphological criteria. KEY POINTS: • Versions 2005 and 2019 Bosniak classifications achieved similar diagnostic performance, but the specificity of higher risk categories (III and IV) was not increased while one malignant lesion was downgraded to v2019 Bosniak category II (i.e., not subjected to further follow-up). • Inter-reader agreement was similar between v2005 and v2019 but moderately decreased for v2019 when the Bosniak classification was dichotomized according to the potential need for ablative therapies (I-II-IIF vs III-IV).

Malignancy risk stratification of cystic renal lesions based on a contrast-enhanced CT-based machine learning model and a clinical decision algorithm.

OBJECTIVE: To distinguish benign from malignant cystic renal lesions (CRL) using a contrast-enhanced CT-based radiomics model and a clinical decision algorithm. METHODS: This dual-center retrospective study included patients over 18 years old with CRL between 2005 and 2018. The reference standard was histopathology or 4-year imaging follow-up. Training and testing datasets were acquired from two institutions. Quantitative 3D radiomics analyses were performed on nephrographic phase CT images. Ten-fold cross-validated LASSO regression was applied to the training dataset to identify the most discriminative features. A logistic regression model was trained to classify malignancy and tested on the independent dataset. Reported metrics included areas under the receiver operating characteristic curves (AUC) and balanced accuracy. Decision curve analysis for stratifying patients for surgery was performed in the testing dataset. A decision algorithm was built by combining consensus radiological readings of Bosniak categories and radiomics-based risks. RESULTS: A total of 149 CRL (139 patients; 65 years [56-72]) were included in the training dataset-35 Bosniak(B)-IIF (8.6% malignancy), 23 B-III (43.5%), and 23 B-IV (87.0%)-and 50 CRL (46 patients; 61 years [51-68]) in the testing dataset-12 B-IIF (8.3%), 10 B-III (60.0%), and 9 B-IV (100%). The machine learning model achieved high diagnostic performance in predicting malignancy in the testing dataset (AUC = 0.96; balanced accuracy = 94%). There was a net benefit across threshold probabilities in using the clinical decision algorithm over management guidelines based on Bosniak categories. CONCLUSION: CT-based radiomics modeling accurately distinguished benign from malignant CRL, outperforming the Bosniak classification. The decision algorithm best stratified lesions for surgery and active surveillance. KEY POINTS: • The radiomics model achieved excellent diagnostic performance in identifying malignant cystic renal lesions in an independent testing dataset (AUC = 0.96). • The machine learning-enhanced decision algorithm outperformed the management guidelines based on the Bosniak classification for stratifying patients to surgical ablation or active surveillance.

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

Predicting 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.

Predicting histopathology markers of endometrial carcinoma with a quantitative image analysis approach based on spherical harmonics in multiparametric MRI.

PURPOSE: Identifying optimal machine learning pipelines for computer-aided diagnosis is key for the development of robust, reproducible, and clinically relevant imaging biomarkers for endometrial carcinoma. The purpose of this study was to introduce the mathematical development of image descriptors computed from spherical harmonics (SPHARM) decompositions as well as the associated machine learning pipeline, and to evaluate their performance in predicting deep myometrial invasion (MI) and histopathological high-grade in preoperative multiparametric magnetic resonance imaging (MRI). PATIENTS AND METHODS: This retrospective study included 128 women with histopathology-confirmed endometrial carcinomas who underwent 1.5-T MRI before hysterectomy between January 2011 and July 2015. SPHARM descriptors of each tumor were computed on multiparametric MRI images (T2-weighted, diffusion-weighted, dynamic contrast-enhanced-MRI and apparent diffusion coefficient maps). Tensor-based logistic regression was used to classify two-dimensional SPHARM rotationally-invariant descriptors. Head-to-head comparisons with radiomics analyses were performed with DeLong tests with Bonferroni-Holm correction to compare diagnostic performances. RESULTS: With all MRI contrasts, SPHARM analysis resulted in area under the curve, sensitivity, specificity, and balanced accuracy values of 0.94 (95% confidence interval [CI]: 0.85, 1.00), 100% (95% CI: 100, 100), 74% (95% CI: 51, 92), 87% (95% CI: 78, 98), respectively, for predicting deep MI. For predicting high-grade tumor histology, the corresponding values for the same diagnostic metrics were 0.81 (95% CI: 0.64, 0.90), 93% (95% CI: 67, 100), 63% (95% CI: 45, 79) and 78% (95% CI: 64, 86). The corresponding values achieved via radiomics were 0.92 (95% CI: 0.82, 0.95), 82% (95% CI: 65, 93), 80% (95% CI: 51, 94), 81% (95% CI: 70, 91) for deep MI and 0.72 (95% CI: 0.58, 0.83), 93% (95% CI: 65, 100), 55% (95% CI: 41, 69), 74% (95% CI: 52, 88) for high-grade histology. The diagnostic performance of the SPHARM analysis was not significantly different (P = 0.62) from that of radiomics for predicting deep MI but was significantly higher (P = 0.044) for predicting high-grade histology. CONCLUSION: The proposed SPHARM analysis yields similar or higher diagnostic performance than radiomics in identifying deep MI and high-grade status in histology-proven endometrial carcinoma.

Site-Specific Variation in Radiomic Features of Head and Neck Squamous Cell Carcinoma and Its Impact on Machine Learning Models.

Current radiomic studies of head and neck squamous cell carcinomas (HNSCC) are typically based on datasets combining tumors from different locations, assuming that the radiomic features are similar based on histopathologic characteristics. However, molecular pathogenesis and treatment in HNSCC substantially vary across different tumor sites. It is not known if a statistical difference exists between radiomic features from different tumor sites and how they affect machine learning model performance in endpoint prediction. To answer these questions, we extracted radiomic features from contrast-enhanced neck computed tomography scans (CTs) of 605 patients with HNSCC originating from the oral cavity, oropharynx, and hypopharynx/larynx. The difference in radiomic features of tumors from these sites was assessed using statistical analyses and Random Forest classifiers on the radiomic features with 10-fold cross-validation to predict tumor sites, nodal metastasis, and HPV status. We found statistically significant differences (p-value ≤ 0.05) between the radiomic features of HNSCC depending on tumor location. We also observed that differences in quantitative features among HNSCC from different locations impact the performance of machine learning models. This suggests that radiomic features may reveal biologic heterogeneity complementary to current gold standard histopathologic evaluation. We recommend considering tumor site in radiomic studies of HNSCC.

Development and validation of a clinical risk calculator for mortality after colectomy for fulminant Clostridium difficile colitis.

BACKGROUND: Clostridium difficile colitis is an increasingly important cause of morbidity and mortality. Fulminant C. difficile colitis (FCDC) is a severe form of the colitis driven by a significant systemic inflammatory response, and managed with a total abdominal colectomy. Despite surgery, postoperative mortality rates remain high. The aim of this study was to develop a bedside calculator to predict the risk of 30-day postoperative mortality for patients with FCDC. METHODS: After institutional review board approval, the American College of Surgeons National Surgical Quality Improvement Program database (2005-2015) was used to include adult patients who underwent emergency surgery for FCDC. A priori preoperative predictors of mortality were selected from the literature: age, immunosuppression, preoperative shock, intubation, and laboratory values. The predictive accuracy of different logistic regression models was measured by calculating the area under the receiver-operating characteristic curve. A cohort of 124 patients from Québec was used to validate the developed mortality calculator. RESULTS: A total of 557 patients met the inclusion criteria, and the overall mortality was 44%. After developing the calculator, no statistically significant differences were found in comparison with the American College of Surgeons National Surgical Quality Improvement Program probability of mortality available in the database (area under the receiver operating curve, 75.61 vs. 75.14; p = 0.79). External validation with the cohort of patients from Quebec showed an area under the curve of 74.0% (95% confidence interval, 65.0-82.9). CONCLUSION: A clinically applicable calculator using preoperative variables to predict postoperative mortality for patients with FCDC was developed and externally validated. This calculator may help guide preoperative decision making. LEVEL OF EVIDENCE: Prognostic and epidemiological study, level III.