Identifying predictors of antimicrobial exposure in hospitalized patients using a machine learning approach.

AIMS: Analysis and tracking of antimicrobial utilization (AU) are crucial in antimicrobial stewardship efforts which are used to find effective interventions for controlling antimicrobial resistance. In antimicrobial stewardship, standard risk adjustment models are needed for benchmarking appropriate AU and for fair inter-facility comparison. In this study we identify patient- and facility-level predictors of antimicrobial usage in hospitalized patients using a machine learning approach, which can be used to inform a risk adjustment model to facilitate assessment of AU. To our knowledge, this is the first time machine learning has been applied for this purpose. METHODS AND RESULTS: Patient admission records were retrieved from the Duke Antimicrobial Stewardship Outreach Network which include clinical data for 27 community hospitals in the southeastern United States. Candidate features (predictors) were then generated from these records. The number of features was reduced using a statistical approach, and missing values of the reduced feature set were imputed using bootstrapping and expectation-maximization algorithm. Finally, support vector regression (SVR) and cubist regression (CB) models were applied to find root-mean-square error values which were used to evaluate the selected feature set. The performance of the SVR and CB models was found to be better than that of linear null and negative binomial null models, thereby demonstrating the effectiveness of our selected features. CONCLUSIONS: Relevant patient- and facility-level predictors of antimicrobial usage in days of therapy were obtained and evaluated. The potential predictor set can be used in risk adjustment strategies for benchmarking antimicrobial use. SIGNIFICANCE AND IMPACT OF THE STUDY: One reason for the rapid emergence of antimicrobial resistance is inappropriate use of antibiotics in hospitalized patients. Identifying predictors of antimicrobial exposure using a machine learning technique can improve the use of AU, enhance patient health outcomes, and reduce the infection spread caused by antimicrobial-resistant organisms.

Capreomycin resistance prediction in two species of Mycobacterium using a stacked ensemble method.

AIMS: Predicting bacterial resistance provides valuable information that can assist in clinical decisions. With recent advances in whole genome sequencing technology, the detection of antibiotic resistance (AR) proteins directly from genomic data is becoming feasible. AR genes/proteins can be identified using best-hit methods that work by comparing candidate sequences with known AR genes in public databases. However, these approaches may fail to detect resistance genes with sequences that differ significantly from known sequences. Our goal is to develop a machine learning technique to accurately predict capreomycin resistance in Mycobacteria with low false discovery rates. METHODS AND RESULTS: We present a stacked ensemble learning model as an alternative to traditional DNA sequence alignment-based methods using optimal features generated from the physicochemical, evolutionary and secondary structure properties of protein sequences. We train logistic regression, C5.0 and support vector machine (SVM) algorithms as our base classifiers, and our stacked ensemble predictors combine the results from the base classifiers to achieve higher accuracy. Compared with our most accurate base classifier (SVM), our most accurate stacked ensemble predictor increases training accuracy by 2·43%. Our stacked ensemble predictors achieve test accuracy up to 81·25%. CONCLUSIONS: We developed a stacked ensemble model to predict capreomycin resistance for Mycobacteria with an accuracy >80% using protein sequences with sequence similarity ranging between 10% and 70%. This performance cannot be achieved with best-hit methods due to differences in sequence similarity. SIGNIFICANCE AND IMPACT OF THE STUDY: Today an estimated one-half million cases of multidrug-resistant (MDR) and extensively drug-resistant (XDR) tuberculosis (TB) occur annually worldwide at a great cost. Because capreomycin is a second-line drug used to treat drug-resistant TB, the ability to use a machine learning approach to classify capreomycin-resistant TB in a timely manner is crucial for the successful treatment of MDR or XDR TB.

Improved identification of epidemiologically related strains of Salmonella enterica by use of a fusion algorithm based on pulsed-field gel electrophoresis and multiple-locus variable-number tandem-repeat analysis.

Pulsed-field gel electrophoresis (PFGE) and multiple-locus variable-number tandem-repeat analysis (MLVA) are used to assess genetic similarity between bacterial strains. There are cases, however, when neither of these methods quantifies genetic variation at a level of resolution that is well suited for studying the molecular epidemiology of bacterial pathogens. To improve estimates based on these methods, we propose a fusion algorithm that combines the information obtained from both PFGE and MLVA assays to assess epidemiological relationships. This involves generating distance matrices for PFGE data (Dice coefficients) and MLVA data (single-step stepwise-mutation model) and modifying the relative distances using the two different data types. We applied the algorithm to a set of Salmonella enterica serovar Typhimurium isolates collected from a wide range of sampling dates, locations, and host species. All three classification methods (PFGE only, MLVA only, and fusion) produced a similar pattern of clustering relative to groupings of common phage types, with the fusion results being slightly better. We then examined a group of serovar Newport isolates collected over a limited geographic and temporal scale and showed that the fusion of PFGE and MLVA data produced the best discrimination of isolates relative to a collection site (farm). Our analysis shows that the fusion of PFGE and MLVA data provides an improved ability to discriminate epidemiologically related isolates but provides only minor improvement in the discrimination of less related isolates.

Ultrasound imaging using variations of the iterative Born technique [biomedical diagnosis].

The iterative Born method is an inverse technique that has been used successfully in ultrasound imaging. However, the calculation cost of the standard iterative Born method is high, and parallel computation is limited to the forward problem. In this work, two methods are introduced to increase the rate of convergence of the iterative Born algorithm. These methods are tested on three different objects. The results are promising, with both algorithms giving accurate results at lower computational cost. The first method, referred to as the coarse resolution initial value (CRIV) method, uses the iterative Born algorithm for a coarse grid to quickly estimate the initial value of the object to be reconstructed. From this initial value, the final image is obtained for a finer grid with additional iterations. The cost of this method is 40% less than that of the iterative Born technique. The second method, the quadriphase source (QS) method, simultaneously uses four single sources, and object reconstruction for each is performed in parallel; the reconstruction results for all four sources then are averaged to obtain the final image. The cost of this method is 20% less than that of the standard iterative Born method. When the object to be reconstructed is of low contrast and/or has a small phase shift, the QS method is very promising because parallel computation can be used to solve both the forward and inverse problems. However, the QS method fails for high contrast objects.

Inverse imaging of the breast with a material classification technique.

In recent publications [Chew et al., IEEE Trans. Blomed. Eng. BME-9, 218-225 (1990); Borup et al., Ultrason. Imaging 14, 69-85 (1992)] the inverse imaging problem has been solved by means of a two-step iterative method. In this paper, a third step is introduced for ultrasound imaging of the breast. In this step, which is based on statistical pattern recognition, classification of tissue types and a priori knowledge of the anatomy of the breast are integrated into the iterative method. Use of this material classification technique results in more rapid convergence to the inverse solution--approximately 40% fewer iterations are required--as well as greater accuracy. In addition, tumors are detected early in the reconstruction process. Results for reconstructions of a simple two-dimensional model of the human breast are presented. These reconstructions are extremely accurate when system noise and variations in tissue parameters are not too great. However, for the algorithm used, degradation of the reconstructions and divergence from the correct solution occur when system noise and variations in parameters exceed threshold values. Even in this case, however, tumors are still identified within a few iterations.

Phase insensitive homomorphic image processing for speckle reduction.

Speckle appears in all conventional ultrasound images and is caused by the use of a phase-sensitive transducer. Speckle is an undesirable property as it can mask small but perhaps diagnostically significant image features. In this paper a homomorphic, hybrid nonlinear processing method, based on cancellation of scattering interference, is developed and examined. Experiments with synthetic and real ultrasound imagery show that the proposed method improves the contrast-to-noise ratio in both lesion and cyst areas and preserves edge clarity.

FDTD simulations for ultrasound propagation in a 2-D breast model.

To increase the survival rates of patients with breast cancer, an ultrasound imaging system must detect tumors when they are small, with a diameter of 5 mm or less. This requires an understanding of how propagation of ultrasound energy is affected by the complex structure of the breast. In this paper, a Finite-Difference Time-Domain (FDTD) method is developed to simulate ultrasound propagation in a two-dimensional model of the human breast. The FDTD simulations make it possible to better understand the behavior of an ultrasound signal in the breast. For example, here the simulations are used to investigate the effect of fat lobes adjacent to the skin layer in a simple breast model. Experimental work performed at the University of Pennsylvania has shown that strong refraction caused by the fat lobes results in nulls in the forward transmitted field. This result was duplicated with the FDTD simulations, and it was shown that the effect of refraction is clearly evident for energy exiting the breast. The existence of strong refraction has a significant impact on ultrasound imaging since it implies that an imaging method based on a weak scattering assumption is unlikely to work well.

An eccentrically coated asymmetric antenna applicator for intracavitary hyperthermia treatment of cancer.

In this paper a model based on transmission line theory is used to predict the behavior of an eccentrically coated asymmetric antenna applicator for use in intracavitary hyperthermia. Theoretical results for the heating rate (HR) of the applicators are compared to experimental results. The experimental results were obtained at City of Hope National Medical Center using four different 915-MHz applicators, each with a different antenna size and eccentricity of the coating. A parameter delta is defined where delta << 1.0 is a thin wire approximation; delta is primarily a function of the eccentricity of the coating, the antenna diameter, and the coating diameter. It is found that when delta approximately less than 0.5, the theoretical model works well. In particular, it predicts the directivity due to the eccentricity of the coating. However, as this eccentricity is increased or as the antenna diameter is increased (delta approximately greater than 0.6), the model no longer accurately predicts directivity. Thus, the model that can be used to predict the HR profiles for an eccentrically coated asymmetric antenna only when delta approximately less than 0.5.