Gilles Cohen

Learning from imbalanced data in surveillance of nosocomial infection

OBJECTIVE An important problem that arises in hospitals is the monitoring and detection of nosocomial or hospital acquired infections (NIs). This paper describes a retrospective analysis of a prevalence survey of NIs done in the Geneva University Hospital. Our goal is to identify patients with one or more NIs on the basis of clinical and other data collected during the survey. METHODS AND MATERIAL Standard surveillance strategies are time-consuming and cannot be applied hospital-wide; alternative methods are required. In NI detection viewed as a classification task, the main difficulty resides in the significant imbalance between positive or infected (11%) and negative (89%) cases. To remedy class imbalance, we explore two distinct avenues: (1) a new re-sampling approach in which both over-sampling of rare positives and under-sampling of the noninfected majority rely on synthetic cases (prototypes) generated via class-specific sub-clustering, and (2) a support vector algorithm in which asymmetrical margins are tuned to improve recognition of rare positive cases. RESULTS AND CONCLUSION Experiments have shown both approaches to be effective for the NI detection problem. Our novel re-sampling strategies perform remarkably better than classical random re-sampling. However, they are outperformed by asymmetrical soft margin support vector machines which attained a sensitivity rate of 92%, significantly better than the highest sensitivity (87%) obtained via prototype-based re-sampling.

Fusing visual and clinical information for lung tissue classification in high-resolution computed tomography

OBJECTIVE We investigate the influence of the clinical context of high-resolution computed tomography (HRCT) images of the chest on tissue classification. METHODS AND MATERIALS 2D regions of interest in HRCT axial slices from patients affected with an interstitial lung disease are automatically classified into five classes of lung tissue. Relevance of the clinical parameters is studied before fusing them with visual attributes. Two multimedia fusion techniques are compared: early versus late fusion. Early fusion concatenates features in one single vector, yielding a true multimedia feature space. Late fusion consisting of the combination of the probability outputs of two support vector machines. RESULTS AND CONCLUSION The late fusion scheme allowed a maximum of 84% correct predictions of testing instances among the five classes of lung tissue. This represents a significant improvement of 10% compared to a pure visual-based classification. Moreover, the late fusion scheme showed high robustness to the number of clinical parameters used, which suggests that it is appropriate for mining clinical attributes with missing values in clinical routine.

Comparative Performance Analysis of State-of-the-Art Classification Algorithms Applied to Lung Tissue Categorization

In this paper, we compare five common classifier families in their ability to categorize six lung tissue patterns in high-resolution computed tomography (HRCT) images of patients affected with interstitial lung diseases (ILD) and with healthy tissue. The evaluated classifiers are naive Bayes, k-nearest neighbor, J48 decision trees, multilayer perceptron, and support vector machines (SVM). The dataset used contains 843 regions of interest (ROI) of healthy and five pathologic lung tissue patterns identified by two radiologists at the University Hospitals of Geneva. Correlation of the feature space composed of 39 texture attributes is studied. A grid search for optimal parameters is carried out for each classifier family. Two complementary metrics are used to characterize the performances of classification. These are based on McNemar’s statistical tests and global accuracy. SVM reached best values for each metric and allowed a mean correct prediction rate of 88.3% with high class-specific precision on testing sets of 423 ROIs.

A classification framework for lung tissue categorization

We compare five common classifier families in their ability to categorize six lung tissue patterns in high-resolution computed tomography (HRCT) images of patients affected with interstitial lung diseases (ILD) but also normal tissue. The evaluated classifiers are Naive Bayes, k-Nearest Neighbor (k-NN), J48 decision trees, Multi-Layer Perceptron (MLP) and Support Vector Machines (SVM). The dataset used contains 843 regions of interest (ROI) of healthy and five pathologic lung tissue patterns identified by two radiologists at the University Hospitals of Geneva. Correlation of the feature space composed of 39 texture attributes is studied. A grid search for optimal parameters is carried out for each classifier family. Two complementary metrics are used to characterize the performances of classification. Those are based on McNemars statistical tests and global accuracy. SVM reached best values for each metric and allowed a mean correct prediction rate of 87.9% with high class-specific precision on testing sets of 423 ROIs.

Lung Tissue Classification in HRCT Data Integrating the Clinical Context

In this paper, we investigate the influence of the clinical context of high-resolution computed tomography (HRCT) images of the chest on tissue classification. Evaluation of the classification performance is based on high-quality visual data extracted from clinical routine. The clinical attributes with highest information gain ratio show to be relevant and consistent for the classification of lung tissue patterns. A combination of visual and clinical attributes allowed a mean of 93% correct predictions of testing instances among the five classes of lung tissue with optimized support vector machines (SVM), which represents a significant benefit of 8% compared to a pure visually-based classification.

Data Imbalance in Surveillance of Nosocomial Infections

An important problem that arises in hospitals is the monitoring and detection of nosocomial or hospital acquired infections (NIs). This paper describes a retrospective analysis of a prevalence survey of NIs done in the Geneva University Hospital. Our goal is to identify patients with one or more NIs on the basis of clinical and other data collected during the survey. In this classification task, the main difficulty resides in the significant imbalance between positive or infected (11%) and negative (89%) cases. To remedy class imbalance, we propose a novel approach in which both oversampling of rare positives and undersampling of the non infected majority rely on synthetic cases generated via class-specific subclustering. Experiments have shown this approach to be remarkably more effective than classical random resampling methods.

Model Selection for Support Vector Classifiers via Genetic Algorithms. An Application to Medical Decision Support

This paper addresses the problem of tuning hyperparameters in support vector machine modeling. A Genetic Algorithm-based wrapper, which seeks to evolve hyperparameter values using an empirical error estimate as a fitness function, is proposed and experimentally evaluated on a medical dataset. Model selection is then fully automated. Unlike other hyperparameters tuning techniques, genetic algorithms do not require supplementary information making them well suited for practical purposes. This approach was motivated by an application where the number of parameters to adjust is greater than one. This method produces satisfactory results.

One-Class Support Vector Machines with a Conformal Kernel. A Case Study in Handling Class Imbalance

Class imbalance is a widespread problem in many classification tasks such as medical diagnosis and text categorization. To overcome this problem, we investigate one-class SVMs which can be trained to differentiate two classes on the basis of examples from a single class. We propose an improvement of one-class SVMs via a conformal kernel transformation as described in the context of binary SVM classifiers by [2,3]. We tested this improved one-class SVM on a health care problem that involves discriminating 11% nosocomially infected patients from 89% non infected patients. The results obtained are encouraging: compared with three other SVM-based approaches to coping with class imbalance, one-class SVMs achieved the highest sensitivity recorded so far on the nosocomial infection dataset. However, the price to pay is a concomitant decrease specificity, and it is for domain experts to decide the proportion of false positive cases they are willing to accept in order to ensure treatment of all infected patients.

Does colonization with methicillin-susceptible Staphylococcus aureus protect against nosocomial acquisition of methicillin-resistant S. aureus?

OBJECTIVE To test the hypothesis that methicillin-susceptible Staphylococcus aureus (MSSA) carriage may protect against nosocomial methicillin-resistant S. aureus (MRSA) acquisition by competing for colonization of the anterior nares. DESIGN Prospective cohort and nested case-control study. SETTING Swiss university hospital. PATIENTS All adult patients admitted to 14 wards of the general medicine division between April 1 and October 31, 2007. METHODS Patients were screened for MRSA and MSSA carriage at admission to and discharge from the division. Associations between nosocomial MRSA acquisition and MSSA colonization at admission and other confounders were analyzed by univariable and multivariable analysis. RESULTS Of 898 patients included, 183 (20%) were treated with antibiotics. Nosocomial MRSA acquisition occurred in 70 (8%) of the patients (case patients); 828 (92%) of the patients (control subjects) were free of MRSA colonization at discharge. MSSA carriage at admission was 20% and 21% for case patients and control subjects, respectively. After adjustment by multivariate logistic regression, no association was observed between MSSA colonization at admission and nosocomial MRSA acquisition (adjusted odds ratio [aOR], 1.2 [95% confidence interval (CI), 0.6-2.3]). By contrast, 4 independent predictors of nosocomial MRSA acquisition were identified: older age (aOR per 1-year increment, 1.05 [95% CI, 1.02-1.08]); increased length of stay (aOR per 1-day increment, 1.05 [95% CI, 1.02-1.09]); increased nursing workload index (aOR per 1-point increment, 1.02 [95% CI, 1.01-1.04]); and previous treatment with macrolides (aOR, 5.6 [95% CI, 1.8-17.7]). CONCLUSIONS Endogenous MSSA colonization does not appear to protect against nosocomial MRSA acquisition in a population of medical patients without frequent antibiotic exposure.

Methicillin-resistant Staphylococcus aureus risk profiling: who are we missing?

BackgroundTargeted screening of patients at high risk for methicillin-resistant Staphylococcus aureus (MRSA) carriage is an important component of MRSA control programs, which rely on prediction tools to identify those high-risk patients. Most previous risk studies reported a substantial rate of patients who are eligible for screening, but failed to be enrolled. The characteristics of these missed patients are seldom described. We aimed to determine the rate and characteristics of patients who were missed by a MRSA screening programme at our institution to see how the failure to include these patients might impact the accuracy of clinical prediction tools.FindingsFrom March-June 2010 all patients admitted to 13 internal medicine wards at the University of Geneva Hospital (HUG) were prospectively screened for MRSA carriage. Of 1968 patients admitted to the ward, 267 patients (13.6%) failed to undergo appropriate MRSA screening. Forty-one (2.4%) screened patients were MRSA carriers at admission. On multivariate regression, patients who were missed by screening were more likely to be aged < 50 years (OR 2.4 [1.4-3.9]), transferred to internal medicine from another ward in the hospital (OR 2.8 [1.1-7.1]), and have a history of malignancy (OR 3.2[2.1-5.1]). There was no significant difference in the rate of previous MRSA carriage between screened and unscreened patients.ConclusionsOur findings highlight the potential bias that “missed” patients may introduce into MRSA risk scores. Reporting on the proportions and characteristics of missed patients is essential for accurate interpretation of MRSA prediction tools.