Search

Machine learning has been widely used for solving challenging problems in diverse areas. However, to the best of our knowledge, seldom literature has discussed in-depth how machine learning approaches can be used effectively to “hunt” (identify) threats, especially advanced persistent threats (APTs) , in a monitored environment. In this study, we share our past experiences in building machine learning-based threat-hunting models. Several challenges must be considered when a security team attempts to build such models. These challenges include (1) weak signal, (2) imbalanced data sets, (3) lack of high-quality labels, and (4) no storyline. In this study, we propose Fuchikoma and APTEmu to demonstrate how we tackle the above-mentioned challenges. The former is a proof of concept system for demonstrating the ideas behind autonomous threat-hunting. It is a machine learning-based anomaly detection and threat hunting system which leverages natural language processing (NLP) and graph algorithms. The latter is an APT emulator, which emulates the behavior of a well-known APT called APT3, which is the target used in the first round of MITRE ATT&CK Evaluations. APTEmu generates attacks on Windows machines in a virtualized environment, and the captured system events can be further used to train and enhance Fuchikoma’s capabilities. We illustrate the steps and experiments we used to build the models, discuss each model’s effectiveness and limitations of each model, and propose countermeasures and solutions to improve the models. Our evaluation results show that machine learning algorithms can effectively assist threat hunting processes and significantly reduce security analysts’ efforts. Fuchikoma correctly identifies malicious commands and achieves high performance in terms of over 80% True Positive Rate and True Negative Rate and over 60% F3. We believe our proposed approaches provide valuable experiences in the area and shed light on automated threat-hunting research.

We assessed two electronic search tools that screen medical records for documented fractures. Both programs reliably identified patients with any fracture but missed individuals with minimal trauma fracture to different degrees. A hybrid tool combining the methodology of both tools is likely to improve the identification of those with osteoporosis. Most patients who suffer a minimal trauma fracture remain undiagnosed, placing them at high risk of refracture. Case finding can be improved by electronic search tools that screen medical records for documented fractures. Here, we assessed the efficacy of two new programs, AES and XRAIT, in identifying patients with minimal trauma fracture. Each tool was applied to search the electronic medical record and/or radiology reports at two tertiary hospitals in Sydney, Australia, from 1 July to 31 December 2018. Samples of the extracted reports were then manually reviewed to determine the sensitivity of each program in detecting minimal trauma fractures. At the two centers, AES detected 872 and 1364 cases, whereas XRAIT identified 1414 and 2180 patients with fractures, respectively. The true positive rate for “any fracture” was similar for both instruments (77–88%). However, the ability to detect “minimal trauma fractures” differed between programs and centers (53–75% accuracy), with each tool identifying separate subsets of patients. Concordance between both tools was less than half of the combined total number of minimal trauma fractures (43–45%). Considering the total number of minimal trauma fractures detected by both tools combined, AES correctly identified 52–55% of cases while XRAIT identified 88–93% of cases. Both programs reliably identified patients with any fracture but missed individuals with minimal trauma fracture to different degrees. Hybrid tools combining the methodology of XRAIT and AES are likely to improve the identification of patients who require investigation and treatment for osteoporosis.

One of the key roles of an orthopedic surgeon treating metastatic bone disease (MBD) is fracture risk prediction. Current widely used impending fracture risk tools such as Mirels scoring lack specificity. Two newer methods of fracture risk prediction, CT-based structural rigidity analysis (CTRA) and finite element analysis (FEA), have each been shown to be more accurate than Mirels. This case series illustrates comparative Mirels, CTRA, and FEA for 8 femurs in 7 subjects. These cases were selected from a much larger data set to portray examples of true positives, true negatives, false positives, and false negatives as defined by CTRA relative to the fracture outcome. Case illustrations demonstrate comparative Mirels and FEA. This series illustrates the use, efficacy, and limitations of these tools. As all current tools have limitations, further work is needed in refining and developing fracture risk prediction.

Objective To analyze the association between radiologists’ performance and image position within a batch in screen reading of mammograms in Norway. Method We described true and false positives and true and false negatives by groups of image positions and batch sizes for 2,937,312 screen readings performed from 2012 to 2018. Mixed-effects models were used to obtain adjusted proportions of true and false positive, true and false negative, sensitivity, and specificity for different image positions. We adjusted for time of day and weekday and included the individual variation between the radiologists as random effects. Time spent reading was included in an additional model to explore a possible mediation effect. Result True and false positives were negatively associated with image position within the batch, while the rates of true and false negatives were positively associated. In the adjusted analyses, the rate of true positives was 4.0 per 1000 (95% CI: 3.8–4.2) readings for image position 10 and 3.9 (95% CI: 3.7–4.1) for image position 60. The rate of true negatives was 94.4% (95% CI: 94.0–94.8) for image position 10 and 94.8% (95% CI: 94.4–95.2) for image position 60. Per 1000 readings, the rate of false negative was 0.60 (95% CI: 0.53–0.67) for image position 10 and 0.62 (95% CI: 0.55–0.69) for image position 60. Conclusion There was a decrease in the radiologists’ sensitivity throughout the batch, and although this effect was small, our results may be clinically relevant at a population level or when multiplying the differences with the number of screen readings for the individual radiologists. Key Points • True and false positive reading scores were negatively associated with image position within a batch. • A decreasing trend of positive scores indicated a beneficial effect of a certain number of screen readings within a batch. • False negative scores increased throughout the batch but the association was not statistically significant.

Objectives Alpha1-antitrypsin deficiency (AATD) is an inherited condition that predisposes individuals to an increased risk of developing lung and liver disease. Even though AATD is one of the most widespread inherited diseases in Caucasian populations, only a minority of affected individuals has been detected. Whereas methods have been validated for AATD testing, there is no universally-established algorithm for the detection and diagnosis of the disorder. In order to compare different methods for diagnosing AATD, we carried out a systematic review of the literature on AATD diagnostic algorithms. Methods Complete biochemical and molecular analyses of 5,352 samples processed in our laboratory were retrospectively studied using each of the selected algorithms. Results When applying the diagnostic algorithms to the same samples, the frequency of False Negatives varied from 1.94 to 12.9%, the frequency of True Negatives was 62.91% for each algorithm and the frequency of True Positives ranged from 24.19 to 35.15%. We, therefore, highlighted some differences among Negative Predictive Values, ranging from 0.83 to 0.97. Accordingly, the sensitivity of each algorithm ranged between 0.61 and 0.95. We also postulated 1.108 g/L as optimal AAT cut-off value, in absence of inflammatory status, which points to the possible presence of genetic AATD. Conclusions The choice of the diagnostic algorithm has a significant impact on the correct diagnosis of AATD, which is essential for appropriate treatment and medical care. The fairly large number of possible false negative diagnoses revealed by the present paper should also warn clinicians of negative results in patients with clinically-suspected AATD.

Interpretation of gastrointestinal PET/computed tomography (PET/CT) is often complicated by anatomy including bowel folds, flexures, variant redundancy, decompressed bowel segments and physiological uptake. This makes it very difficult to identify both true positives and true negatives, compromising both sensitivity and specificity. CT enterography is increasingly being integrated into the field of nuclear medicine to address these issues. This technique uses the combination of negative/neutral contrast to distend the lumen of the gut and iodinated contrast to enhance the gastrointestinal wall and pathological findings. Apart from augmentation in the diagnostic performance, the technique also improves the quality of the imaging, confidence of the reporting physician and inter-rater agreement. Therefore, this technique has found favor among nuclear medicine physicians, in the imaging of chronic inflammatory disorders and malignancies in and of the gut. It is a feasible and easily executable procedure with minimal and manageable side-effects and should be routinely recommended in cases where interference from physiologic findings is expected.

In recent years, ensemble classification methods have been widely investigated in both industry and literature in the field of machine learning and artificial intelligence. The main advantage of this approach is to benefit from a set of classifiers instead of using a single classifier with the aim of improving the prediction performance, such as accuracy. Selecting the base classifiers and the method for combining them are the most challenging issues in the ensemble classifiers. In this paper, we propose a heterogeneous dynamic ensemble classifier (HDEC) which uses multiple classification algorithms. The main advantage of using heterogeneous algorithms is increasing the diversity among the base classifiers as it is a key point for an ensemble system to be successful. In this method, we first train many classifiers with the original data. Then, they are separated based on their strength in recognizing either positive or negative instances. For doing this, we consider the true positive rate and true negative rate, respectively. In the next step, the classifiers are categorized into two groups according to their efficiency in the mentioned measures. Finally, the outputs of the two groups are compared with each other to generate the final prediction. For evaluating the proposed approach, it has been applied to 12 datasets from the UCI and LIBSVM repositories and calculated two popular prediction performance metrics, including accuracy and geometric mean. The experimental results show the superiority of the proposed approach in comparison to other state-of-the-art methods.

Objective To investigate the performance of Dual-AI Deep Learning Platform in detecting unreported pulmonary nodules that are 6 mm or greater, comprising computer-vision (CV) algorithm to detect pulmonary nodules, with positive results filtered by natural language processing (NLP) analysis of the dictated report. Methods Retrospective analysis of 5047 chest CT scans and corresponding reports. Cases which were both CV algorithm positive (nodule ≥ 6 mm) and NLP negative (nodule not reported), were outputted for review by 2 chest radiologists. Results The CV algorithm detected nodules that are 6 mm or greater in 1830 (36.3%) of 5047 cases. Three hundred fifty-five (19.4%) were unreported by the radiologist, as per NLP algorithm. Expert review determined that 139 (39.2%) of 355 cases were true positives (2.8% of all cases). One hundred thirty (36.7%) of 355 cases were unnecessary alerts-vague language in the report confounded the NLP algorithm. Eighty-six (24.2%) of 355 cases were false positives. Conclusions Dual-AI platform detected actionable unreported nodules in 2.8% of chest CT scans, yet minimized intrusion to radiologist's workflow by avoiding alerts for most already-reported nodules.

The human right ventricle is barely monitored during open-chest surgery due to the absence of intraoperative imaging techniques capable of elaborating its complex function. Accordingly, artificial intelligence could not be adopted for this specific task. We recently proposed a video-based approach for the real-time evaluation of the epicardial kinematics to support medical decisions. Here, we employed two supervised machine learning algorithms based on our technique to predict the patients’ outcomes before chest closure. Videos of the beating hearts were acquired before and after pulmonary valve replacement in twelve Tetralogy of Fallot patients and recordings were properly labeled as the “unhealthy” and “healthy” classes. We extracted frequency-domain-related features to train different supervised machine learning models and selected their best characteristics via 10-fold cross-validation and optimization processes. Decision surfaces were built to classify two additional patients having good and unfavorable clinical outcomes. The k-nearest neighbors and support vector machine showed the highest prediction accuracy, the patients’ class was identified with a true positive rate ≥95% and the decision surfaces correctly classified the additional patients in the “healthy” (good outcome) or “unhealthy” (unfavorable outcome) classes. We demonstrated that classifiers employed with our video-based technique may aid cardiac surgeons in decision making before chest closure.

AimsThe application of artificial intelligence (AI) for automated diagnosis of electrocardiograms (ECGs) can improve access to high quality diagnostic care in remote settings but is limited by the reliance on signal-based data that are not routinely available. We sought to develop a multilabel automated diagnosis model for electrocardiographic images, more suitable for broader use.Methods and ResultsA total of 2,228,236 12-lead ECGs from 811 municipalities in Minas Gerais, Brazil were sampled into 90% (training):5%(validation):5%(held-out test), and were transformed to ECG images with varying lead locations and formats. We trained a convolutional neural network (CNN) using an EfficientNet-based architecture on ECG images to identify 6 physician-defined clinical labels spanning rhythm and conduction disorders, as well as a hidden label for gender. We trained another CNN for signal-based classification. The image-based model performed well on the held-out test set (average AUROC 0.99, AUPRC 0.68). This was replicated in a distinct test set from Brazil validated by at least two cardiologists (average AUROC 0.99, AUPRC 0.86) as well as an external validation set of 21,785 ECGs from Germany (average AUROC 0.97, AUPRC 0.73), with performance superior to signal-based models. Expert review of 120 out of 120 high confidence false positive predictions on the held-out and external validation sets were confirmed to be true positives with incorrect labels. The model learned generalizable features, confirmed using Gradient-weighted Class Activation Mapping (Grad-CAM).ConclusionWe developed an externally validated model that extends the automated diagnosis of key rhythm and conduction disorders to printed ECGs as well as to the identification of hidden features, allowing the application of AI to ECGs captured across broad settings.

The automated plate assessment system (APAS Independence; Clever Culture System, Bäch, Switzerland) is an automated imaging station linked with interpretive software that detects target colonies of interest on chromogenic media and sorts samples as negative or presumptive positive. We evaluated the accuracy of the APAS to triage methicillin-resistant Staphylococcus aureus (MRSA) and S. aureus cultures using chromogenic medium compared to that by human interpretation. Patient samples collected from the nares on ESwabs were plated onto BD BBL CHROMagar MRSA II and BD BBL CHROMagar Staph aureus and allowed to incubate for 20 to 24 h at 37°C in a non-CO2 incubator. Mauve colonies are suggestive of S. aureus and were confirmed with latex agglutination. Following incubation, samples were first interrogated by APAS before being read by a trained technologist blinded to the APAS interpretation. The triaging by both APAS and the technologists was compared for accuracy. Any discordant results required further analysis by a third reader. Over a 5-month period, 5,913 CHROMagar MRSA cultures were evaluated. Of those, 236 were read as concordantly positive, 5,525 were read as concordantly negative, and 152 required discordant analysis. Positive and negative percent agreements (PPA and NPA, respectively) were 100% and 97.3%, respectively. The APAS detected 5 positive cultures that were missed by manual reading and determined to be true positives. In a separate analysis, 744 CHROMagar Staph aureus plates were read in parallel. Of these, 133 were concordantly positive, 585 were concordantly negative, and 26 required discordant analysis. PPA and NPA were 95.7% and 96.7%, respectively. This study confirmed the high sensitivity of digital image analysis by the APAS Independence such that negative cultures can be reliably reported without technologist intervention (negative predictive values [NPVs] of 100% for CHROMagar MRSA and 99.0% for CHROMagar Staph aureus). Triaging using the APAS Independence may provide great efficiency in a laboratory with high throughput of MRSA and S. aureus surveillance cultures.

Due to its relevance in a variety of aerial surveillance and countermeasure systems, detecting point targets against a cluttered backdrop in infrared images has always remained an important topic of research. We evaluate the performance of six point target detection algorithms- Top hat morphology, Modified Top Hat Morphology, Contour Morphology, Method of Directional derivative, Max-Mean and Max-Median- for a synthetic IR video dataset comprising of point targets following predefined trajectories amid cluttered background. True positive rate (Probability of detection) and false alarm rate averaged over all the video frames have been considered as performance measure. It is found that for all the algorithms considered there is a trade off between the True Positive and False Alarm Rate.

Intelligent lameness detection is important for improving cow welfare. A method based on YOLOv3 deep learning algorithm and relative step size characteristic vector is proposed to classify lame and non-lame cows. Videos were decomposed into sequence frames, and leg targets of cows in each frame were detected by YOLOv3 algorithm. Relative step sizes of cow's front and rear legs were calculated based on leg coordinates, and the relative step size characteristic vector was constructed. Finally, a trained Long Short-Term Memory (LSTM) classification model was used to classify lame and non-lame cows based on the characteristic vector. A total of 210 videos were selected for verification using LSTM, support vector machine (SVM), K-Nearest Neighbour (KNN) and decision tree classifier (DTC) algorithms. Results showed that accuracy of lameness detection based on LSTM was 98.57%, which was 2.93%, 3.88%, and 9.25% higher than SVM, KNN, and DTC, respectively. True positive rate of the LSTM was 0.97, which was 0.03, 0.04 and 0.06 higher than SVM, KNN and DTC, respectively. False positive rate of LSTM was 0.03, which was 0.03, 0.06 and 0.11 lower than SVM, KNN and DTC, respectively. A bidirectional LSTM performed slightly better than LSTM but would be more demanding on hardware. Comparison of LSTM with a purely deep learning method showed the latter performed slightly better, but was less conducive to interpretation and diagnosis of lameness. The relative step size characteristic vector proposed was effective for classification of lame and non-lame cows, and could lead to intelligent detection of lameness.

Rapid test (RT) is the principal screening method in the HIV control practice. However, this method may lead to inaccurate detection, primarily due to the more than 4 weeks of window phase. In the present study, we performed a HIV DNA screening method to show its application prospects in men who have sex with men (MSM). From July 2017 to April 2018, we recruited 1,301 MSM from Beijing who were not previously diagnosed as HIV positive. Both HIV DNA detection and RT were performed. In total, 141 and 135 HIV-positive results were detected by DNA detection and RT, respectively. By repetitive and confirmative tests (Western blot), we verified that DNA detection detected 10 more true positives than RT and 4 false positives were corrected from RT. This represents 14 inaccurate RT results that were corrected by DNA measurement. Therefore, DNA measurement should be fully considered as a screening method in the detection of HIV among MSM in the future.

A Bloom filter is a special case of an artificial neural network with two layers. Traditionally, it is seen as a simple data structure supporting membership queries on a set. The standard Bloom filter does not support the delete operation, and therefore, many applications use a counting Bloom filter to enable deletion. This paper proposes a generalization of the counting Bloom filter approach, called “autoscaling Bloom filters”, which allows adjustment of its capacity with probabilistic bounds on false positives and true positives. Thus, by relaxing the requirement on perfect true positive rate, the proposed autoscaling Bloom filter addresses the major difficulty of Bloom filters with respect to their scalability. In essence, the autoscaling Bloom filter is a binarized counting Bloom filter with an adjustable binarization threshold. We present the mathematical analysis of its performance and provide a procedure for minimizing its false positive rate., Neural Computing and Applications, 32 (8), ISSN:1433-3058, ISSN:0941-0643

ObjectiveIBD prevalence is estimated to be rising, but no detailed, recent UK data are available. The last reported prevalence estimate in the UK was 0.40% in 2003. We aimed to establish the current, and project future, prevalence in Lothian, Scotland.DesignWe conducted an all-age multiparameter search strategy using inpatient IBD international classification of disease (ICD-10) coding (K50/51)(1997–2018), IBD pathology coding (1990–2018), primary and secondary care prescribing data (2009–2018) and a paediatric registry, (1997–2018) to identify ‘possible’ IBD cases up to 31/08/2018. Diagnoses were manually confirmed through electronic health record review as per Lennard-Jones/Porto criteria. Autoregressive integrated moving average (ARIMA) regression was applied to forecast prevalence to 01/08/2028.ResultsIn total, 24 601 possible IBD cases were identified of which 10 499 were true positives. The point prevalence for IBD in Lothian on 31/08/2018 was 784/100 000 (UC 432/100 000, Crohn’s disease 284/100 000 and IBD unclassified (IBDU) 68/100 000). Capture–recapture methods identified an additional 427 ‘missed’ cases (95% CI 383 to 477) resulting in a ‘true’ prevalence of 832/100 000 (95% CI 827 to 837).Prevalence increased by 4.3% per year between 2008 and 2018 (95% CI +3.7 to +4.9%, p80 years of age.ConclusionsWe report a rigorously validated IBD cohort with all-age point prevalence on 31/08/2018 of 1 in 125, one of the highest worldwide.

Mobile ads are plagued with fraudulent clicks which is a major challenge for the advertising community. Although popular ad networks use many techniques to detect click fraud, they do not protect the client from possible collusion between publishers and ad networks. In addition, ad networks are not able to monitor the user’s activity for click fraud detection once they are redirected to the advertising site after clicking the ad. We propose a new crowdsource-based system called Click Fraud Crowdsourcing (CFC) that collaborates with both advertisers and ad networks in order to protect both parties from any possible click fraudulent acts. The system benefits from both a global view, where it gathers multiple ad requests corresponding to different ad network-publisher-advertiser combinations, and a local view, where it is able to track the users’ engagement in each advertising website. The results demonstrated that our approach offers a lower false positive rate (0.1) when detecting click fraud as opposed to proposed solutions in the literature, while maintaining a high true positive rate (0.9). Furthermore, we propose a new mobile ad charging model that benefits from our system to charge advertisers based on the duration spent in the advertiser’s website.

Hamstring strain injury (HSI) is one of the most prevalent and severe injury in professional soccer. The purpose was to analyze and compare the predictive ability of a range of machine learning techniques to select the best performing injury risk factor model to identify professional soccer players at high risk of HSIs. A total of 96 male professional soccer players underwent a pre-season screening evaluation that included a large number of individual, psychological and neuromuscular measurements. Injury surveillance was prospectively employed to capture all the HSI occurring in the 2013/2014 season. There were 18 HSIs. Injury distribution was 55.6% dominant leg and 44.4% non-dominant leg. The model generated by the SmooteBoostM1 technique with a cost-sensitive ADTree as the base classifier reported the best evaluation criteria (area under the receiver operating characteristic curve score=0.837, true positive rate=77.8%, true negative rate=83.8%) and hence was considered the best for predicting HSI. The prediction model showed moderate to high accuracy for identifying professional soccer players at risk of HSI during pre-season screenings. Therefore, the model developed might help coaches, physical trainers and medical practitioners in the decision-making process for injury prevention.

Orthogonal confirmation of next-generation sequencing (NGS)-detected germline variants is standard practice, although published studies have suggested that confirmation of the highest-quality calls may not always be necessary. The key question is how laboratories can establish criteria that consistently identify those NGS calls that require confirmation. Most prior studies addressing this question have had limitations: they have been generally of small scale, omitted statistical justification, and explored limited aspects of underlying data. The rigorous definition of criteria that separate high-accuracy NGS calls from those that may or may not be true remains a crucial issue. We analyzed five reference samples and over 80.000 patient specimens from two laboratories. Quality metrics were examined for approximately 200.000 NGS calls with orthogonal data, including 1662 false positives. A classification algorithm used these data to identify a battery of criteria that flag 100% of false positives as requiring confirmation (CI lower bound, 98.5% to 99.8%, depending on variant type) while minimizing the number of flagged true positives. These criteria identify false positives that the previously published criteria miss. Sampling analysis showed that smaller data sets resulted in less effective criteria. Our methodology for determining test- and laboratory-specific criteria can be generalized into a practical approach that can be used by laboratories to reduce the cost and time burdens of confirmation without affecting clinical accuracy.

The Iowa Newborn Screening (NBS) Program began screening for very long-chain acyl-CoA dehydrogenase deficiency (VLCAD) in 2003. Untreated VLCAD can lead to liver failure, heart failure, and death. Current confirmatory testing recommendations by the American College of Medical Genetics (ACMG) for VLCAD list molecular and functional analysis (i.e., fibroblast fatty acid oxidation probe) as optional. This can lead to misclassification of VLCAD carriers as false positives. Iowa implemented a comprehensive VLCAD confirmatory testing algorithm at the beginning of 2016 that included both molecular and fibroblast analysis. Here, we compare the historic multi-algorithmic confirmatory testing protocol (2005–2016) to this comprehensive protocol (2016–2017). A metabolic specialist reviewed all medical records and NBS data for each out-of-range VLCAD that fell in each testing period. During the comprehensive testing period, 48,651 specimens were screened. Thirteen individuals with out-of-range C14:1 results were classified as follows after review: ten carriers, zero true positives, zero false positives, zero lost to follow-up, and four unable to assess carrier status. During the variable testing period, a total of 486,566 specimens were screened. Eighty-five individuals with out-of-range C14:1 were classified as follows: 45 carriers, two true positives, four false positives, four lost to follow-up, and 30 unable to assess carrier status. Our findings suggest that many out-of-range VLCAD cases that do not receive molecular confirmatory testing could be carriers mistakenly classified as false positives. We recommend comprehensive molecular and functional testing for all children with out-of-range VLCAD NBS results. ELECTRONIC SUPPLEMENTARY MATERIAL: The online version of this article (10.1007/s12687-019-00409-8) contains supplementary material, which is available to authorized users.

Although recently developed Electromyography-based (EMG) prosthetic hands could classify a significant amount of wrist motions, classifying 5-6 grasping patterns in real-time is a challenging task. The collaboration of EMG and vision has addressed this problem to a certain extent but could not achieve significant performance in real-time. In this paper, we propose a fusion method that can improve the real-time prediction accuracy of the EMG system by merging a probability matrix that represents the usage of the six grasping patterns for the targeted object. The YOLO object detection system retrieves a probability matrix of the identified object, and it is used to correct the error in the EMG classification system. The experiments revealed that the optimized ANN model outperformed the KNN, LDA, NB, and DT by achieving the highest mean True Positive Rate (mTPR) of 69.34%(21.54) in real-time for all the six grasping patterns. Furthermore, the proposed feature set (Age, Gender, and Handedness of the user) showed that their influence increases the mTPR of ANN by 16.05%(2.70). The proposed system takes 393.89ms(178.23ms) to produce a prediction. Therefore, the user does not feel a delay between intention and execution. Furthermore, the system facilitates users to use multiple-grasping patterns for an object.

Biomarkers abound in many areas of clinical research, and often investigators are interested in combining them for diagnosis, prognosis, or screening. In many applications, the true positive rate for a biomarker combination at a prespecified, clinically acceptable false positive rate is the most relevant measure of predictive capacity. We propose a distribution-free method for constructing biomarker combinations by maximizing the true positive rate while constraining the false positive rate. Theoretical results demonstrate desirable properties of biomarker combinations produced by the new method. In simulations, the biomarker combination provided by our method demonstrated improved operating characteristics in a variety of scenarios when compared with alternative methods for constructing biomarker combinations., 37 pages (including appendices)

To develop and investigate deep learning–based detectors for brain metastases detection on non-enhanced (NE) CT. The study included 116 NECTs from 116 patients (81 men, age 66.5 ± 10.6 years) to train and test single-shot detector (SSD) models using 89 and 27 cases, respectively. The annotation was performed by three radiologists using bounding-boxes defined on contrast-enhanced CT (CECT) images. NECTs were coregistered and resliced to CECTs. The detection performance was evaluated at the SSD’s 50% confidence threshold using sensitivity, positive-predictive value (PPV), and the false-positive rate per scan (FPR). For false negatives and true positives, binary logistic regression was used to examine the possible contributing factors. For lesions 6 mm or larger, the SSD achieved a sensitivity of 35.4% (95% confidence interval (CI): [32.3%, 33.5%]); 51/144) with an FPR of 14.9 (95% CI [12.4, 13.9]). The overall sensitivity was 23.8% (95% CI: [21.3%, 22.8%]; 55/231) and PPV was 19.1% (95% CI: [18.5%, 20.4%]; 98/ of 513), with an FPR of 15.4 (95% CI [12.9, 14.5]). Ninety-five percent of the lesions that SSD failed to detect were also undetectable to radiologists (168/176). Twenty-four percent of the lesions (13/50) detected by the SSD were undetectable to radiologists. Logistic regression analysis indicated that density, necrosis, and size contributed to the lesions’ visibility for radiologists, while for the SSD, the surrounding edema also enhanced the detection performance. The SSD model we developed could detect brain metastases larger than 6 mm to some extent, a quarter of which were even retrospectively unrecognizable to radiologists.

Blood culture tests have a major role in rapid and accurate diagnosis and selection of optimal antibiotic treatment for infectious diseases. The true-positive and contamination rates of blood cultures are performance indicators for the practice of blood culture. However, the optimal number of blood cultures has not been adequately elucidated. We aimed to assess the validity of the number of blood cultures obtained in the hospital. We conducted a retrospective study of blood cultures collected from hospitalized patients between 2010 and 2014 at a tertiary care hospital in Japan and assessed the validity of the number of blood cultures per 1,000 admissions as an indicator for the optimal practice of obtaining blood culture, considering the average length of stay. We also investigated the validity of true-positive and contamination rates of blood cultures in the hospital. The number of blood cultures in this study ranged from 28.6 to 56 per 1,000 patient-days and from 404.4 to 894.6 per 1,000 admissions. The median true-positive and contamination rates of blood cultures per year was 10.1% and 2.8%, respectively. The median contamination rate of blood cultures in a community-acquired group was significantly higher than that in a hospital-acquired group (4.1% vs. 1.9%, p = 0.029). We conclude that there is a possibility that the number of blood cultures per 1,000 admissions is a more appropriate factor to evaluate than per patient-days. Moreover, the true-positive and contamination rates of blood cultures in our hospital are valid.

The purpose of this research is to develop a mobile robot which detects wood splinters automatically. It is hard to find the wood splinters from the floor images because the size of the wood splinters is small and the color is similar. In this research, we develop a mobile robot equipped with two rollers to attach cotton to the wood splinters and object detection to find the location of the wood splinters using the attached cotton. We also build a system to generate a route from a range selected by the inspector and to show results based on running records and detection results. As a result, the accuracy attaching cotton to the wood splinters was 76%. As the precision detecting cotton, the true positive rate was 99%, but there was false positive in the detected result. From the above, we could detect 75% wood splinters. As the result of the running test through the gymnasium, the total inspection time was 87 minutes, and the operation took one person when setting and no person when running, and inspection omission was at most 2.4%. By increasing the number of the dataset of the floor image including the cotton, we will improve this accuracy and reduce false positive when detecting cotton from now.

Reject option is introduced in classification tasks to prevent potential misclassifications. Although optimization of error-reject trade-off has been widely investigated, it is shown that error rate itself is not an appropriate performance measure, when misclassification costs are unequal or class distributions are imbalanced. ROC analysis is proposed as an alternative approach to performance evaluation in terms of true positives (TP) and false positives (FP). Considering classification with reject option, we need to represent the tradeoff between TP, FP and rejection rates. In this paper, we propose 3D ROC analysis to determine the optimal rejection threshold as an analogy to decision threshold optimization in 2D ROC curves. We have demonstrated our proposed method with Naive Bayes classifier on Heart Disease dataset and validated the efficiency of the method on multiple datasets from UCI Machine Learning Repository. Our experiments reveal that classification with optimized rejection threshold significantly improves true positive rates in biomedical datasets. Furthermore, false positive rates remain the same with rejection rates below 10% on average. © 2021, Ismail Saritas. All rights reserved.

AIMS: The aim of this study is to assess the value of ultrasound in the differential diagnosis of orbital lesions. METHODS: One hundred and thirty-eight patients with clinical features of an orbital mass were examined by orbital ultrasound prior to undergoing surgery, from January 2000 to January 2017. All patients underwent excisional or incisional orbital biopsy. The results of orbital ultrasonography were compared with the final histological diagnosis. RESULTS: Orbital lesions were localized by ultrasonography in 133/138 cases (96.4%); the false-negative rate of orbital echography was 3.6% (5/138). The nature of the orbital lesions was correctly determined by ultrasonography in 54.3% of the cases (75/138) preoperatively (true positives). In 58/138 (42%) patients, there was no correspondence between the ultrasound diagnosis and the histological diagnosis (false positives). The sensitivity of orbital ultrasonography for the detection of an orbital mass was 93.75% (CI 87.87–99.63%), while the specificity yielded no meaningful result (CI 0.00–60.24%). Moreover, the specificity of orbital ultrasonography to identify a malignant tumor falls within a CI of 0–62.72%. CONCLUSIONS: Orbital ultrasonography is a rapid and noninvasive test that is highly sensitive in displaying an orbital mass; however, the specificity in the differential diagnosis of orbital lesions is not meaningful, particularly if malignancy is suspected. The assessment of orbital diseases requires multiple diagnostic approaches to balance the strengths and weaknesses of each method.

Handling unbalanced intrusion detection data sets are difficult as minority intrusion classes may not be easy to detect. One of the possible causes of the problem is the characteristic of learning algorithms that usually favour majority classes in data sets. The contribution of this study is to improve the detection rate for intrusions in the unbalanced CICIDS2017 data set by using sampling techniques. We evaluated Random Under-Sampling (RUS), Synthetic Minority Over-sampling Technique (SMOTE) and the combination of RUS and SMOTE. After applying the sampling techniques, we performed intrusion detection and used the accuracy plus True Positive Rate (TPR) as the evaluation metrics for the detection results. The results showed that RUS gave the best detection performance overall. Besides, 12 out of the 15 classes, including some hard-to-detect minority classes, were detected with result improvement.

The main goal of this article is experimental research on machine learning methodology for cyberattacks diagnosing. This study based on two publicly available datasets UNSW-NB15 and NSL-KDD. Its scope included the features reduction problem and calculations of the classification efficiency. We applied Machine Learning (ML) and Deep Learning (DL) methods to classify traffic. The methods of supervised k-nearest neighbours (k-NN) and artificial neural networks (ANN) were used. Accuracy, Precision, True Positive Rate (TPR), False Positive Rate (FPR) were calculated based on a series of numerical experiments for all types of attacks and for DoS (Deny of Service) attacks only. The features number reduction for observation is achieved, and the effect of this reduction on classification accuracy was investigated. We made some conclusions about the possibility of implementing ML/DL methods in intrusion detection systems (IDS).

Simple Summary Collar-mounted activity monitors using battery-powered accelerometers can continuously and accurately analyze specific canine behaviors and activity levels. These include normal behaviors and those that are indicators of disease conditions such as scratching, inappetence, excessive weight, or osteoarthritis. Algorithms used to analyze activity data are validated by video recordings of specific canine behaviors, which were used to label accelerometer data. The study described here was noteworthy for the large volume of data collected from more than 2500 dogs in clinical and real-world home settings. The accelerometer data were analyzed by a machine learning methodology, whereby algorithms were continually updated as additional data were acquired. The study determined that algorithms from the accelerometer data detected eating and drinking behaviors with a high degree of accuracy. Accurate detection of other behaviors such as licking, petting, rubbing, scratching, and sniffing was also demonstrated. The study confirmed that activity monitors using validated algorithms can accurately detect important health-related canine behaviors via a collar-mounted accelerometer. The validated algorithms have widespread practical benefits when used in commercially available canine activity monitors. Abstract Collar-mounted canine activity monitors can use accelerometer data to estimate dog activity levels, step counts, and distance traveled. With recent advances in machine learning and embedded computing, much more nuanced and accurate behavior classification has become possible, giving these affordable consumer devices the potential to improve the efficiency and effectiveness of pet healthcare. Here, we describe a novel deep learning algorithm that classifies dog behavior at sub-second resolution using commercial pet activity monitors. We built machine learning training databases from more than 5000 videos of more than 2500 dogs and ran the algorithms in production on more than 11 million days of device data. We then surveyed project participants representing 10,550 dogs, which provided 163,110 event responses to validate real-world detection of eating and drinking behavior. The resultant algorithm displayed a sensitivity and specificity for detecting drinking behavior (0.949 and 0.999, respectively) and eating behavior (0.988, 0.983). We also demonstrated detection of licking (0.772, 0.990), petting (0.305, 0.991), rubbing (0.729, 0.996), scratching (0.870, 0.997), and sniffing (0.610, 0.968). We show that the devices’ position on the collar had no measurable impact on performance. In production, users reported a true positive rate of 95.3% for eating (among 1514 users), and of 94.9% for drinking (among 1491 users). The study demonstrates the accurate detection of important health-related canine behaviors using a collar-mounted accelerometer. We trained and validated our algorithms on a large and realistic training dataset, and we assessed and confirmed accuracy in production via user validation.

Recently it has been shown, that it is possible to detect malware based on the memory access patterns produced before executions reaches its Entry Point. In this paper, we investigate the usefulness of memory access patterns over time, i.e to what extent can machine learning algorithm trained on “old” data, detect new malware samples, that was not part of the training set and how does this performance change over time. During our experiments, we found that machine learning models trained on memory access patterns of older samples can provide both high accuracy and a high true positive rate for the period from several months to almost a year from the update of the model. We also perform a substantial analysis of our findings that may aid researchers who work with malware and Big Data. © 2020 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.

Summary Background Carbapenems are a family of end line antibiotics with increasing levels of resistance that are a cause for concern. Aim To ascertain whether the CPE screening programme employed in an acute tertiary hospital is fit for purpose. Method We outlined the current working algorithm employed using a universal screening programme over a 26-month screening period. Rectal swabs are cultured on arrival. Those with suspicious growth are further investigated using NG-Carba 5 lateral flow tests and Vitek 2.0 sensitivity cards. These practices were compared with NHS guidelines. Findings & Conclusions In all, 53 true positives were detected from 45 patients since the screening was implemented in early 2018 (46 OXA-48, 6 KPC, 1 NDM). As the rate of screening increased, the number of positive screens decreased over time. There were a lot of similarities between the HSE guidelines and the published NHS CPE toolkit. It was evident that there is no standard practice being employed across all hospitals. Comparing the MUH to national guidelines it appears to be quicker and more effective with universal screening in place at reducing the potential contacts and identifying carriers. Cost analysis indicates that the need to confirm all positive strains in a reference lab is costly, unnecessary and time consuming. There are adequate confirmatory tests available in-house for routine positive screens. It was concluded that infection prevention and control are key to identifying and controlling possible outbreaks in a hospital setting.

IntroductionThe condition onset flag (COF) variable was introduced into hospitalization coding practice in 2008 to distinguish ICD coded diagnoses as either preexisting, new or status unknown. For datasets with multiple hospitalizations per participant, the COF enables identification of diagnoses that may have arisen in between admissions thereby enriching datasets considerably. However, datasets preceding 2008 lack equivalent depth, potentially leading to data waste. Objectives and Approach The aim of this study was to determine if a lookback period could accurately identify new onset conditions in datasets coded prior to 2008. Outcomes identified in an index admission and not present in the lookback period were classified as a new condition. Three different lookback periods entailing the preceding 3, 2, and immediately previous admission were investigated. Accuracy was determined by applying this procedure to datasets in which COF were assigned by coders and comparing agreement between the lookback procedure and the COF. Codes examined were myocardial infarction, pulmonary embolism and pneumonia. Sensitivity analyses excluded indexed admissions where a lookback period was not possible. True positives were those conditions correctly identified by the lookback periods when compared to that classified using the COF. The crude incidence rates were also determined. Results248,042 hospital records from 12,355 patients on chronic kidney replacement therapy were analysed. Lookback Period 1 provided the highest true positive rates for all 3 complications, ranging from 83% to 92%. Sensitivity analyses excluding patients with no previous separations, demonstrated a true positive percentage from 98% to 100%. Lookback period 3 true positive rates were the lowest (77.2-78.3%). The crude incidence calculated for first-time myocardial infarction using Lookback Period 1 was 25 events per 1000 patient years, consistent with current literature. Rates calculated using the COF variable was more than 3-fold lower. This finding was similar when evaluating other clinical outcomes. Conclusion / ImplicationsThe Lookback Period 1 method accurately identifies in-hospital complications compared with the current method using the COF variable and is a means of identifying new hospital complications in health administrative datasets predating the availability of COF, thereby enriching existing datasets and minimising data waste.

The main objective of this paper is the monitoring of daily activity patterns of fattening pigs at different locations in the housing environment using UHF-RFID. Four hundred fattening pigs were equipped with UHF-RFID ear tags and monitored during the fattening period for about four months. The RFID antennas were installed at the feeding troughs, playing devices and drinkers. A validation phase for each of these locations was carried out prior to the first data collection. The sensitivity (true positive rate) of the UHF-RFID system was about 80% at the feeding trough and the playing device and about 60% at the drinkers. The mean of the daily visiting time of all pigs at the trough was about 55 min. The mean visiting duration at the playing device was about 38 min and at the drinkers about 9 min. The visiting times of the pigs showed a high intra- and inter-variability. It was observed that the average visit duration at the feeding trough decreases over the course of a fattening period but increases at the playing device. A documentation of visiting times of animals is possible utilizing RFID systems, allowing a higher data density than video or direct observations.

Non-maximum suppression (NMS) is essential as a post-processing step to eliminate the false positives caused by redundant detections for a single object. However, NMS cannot eliminate all the redundant detections which have different classification results. In order to address this problem, this paper proposes a novel post-processing step, non-reliability suppression (NRS). On PASCAL VOC, the proposed NRS reduces 28.4% false positives with only a 0.3% decrease of true positives for a YOLOv3 detector when compared with the traditional greedy-NMS. Because the computational complexity of NRS is low, the addition of NRS does not require much computational burden.

Nowadays the usage of digital technology has been increasing exponentially. At the same time, the rate of malicious users has been increasing. Online social sites like Facebook and Twitter attract millions of people globally. This interest in online networking has opened to various issues including the risk of exposing false data by creating fake accounts resulting in the spread of malicious content. Fake accounts are a popular way to forward spam, commit fraud and abuse through an online social network. These problems need to be tackled in order to give the user a reliable online social network. In this paper, we are using different ML algorithms like Support Vector Machine (SVM), Logistic Regression (LR), Random Forest (RF) and K-Nearest Neighbours (KNN). Along with these algorithms we have used two different normalization techniques such as Z-Score and Min-Max to improve accuracy. We have implemented it to detect fake Twitter accounts and bots. Our approach achieved high accuracy and true positive rate for Random Forest and KNN.

To demonstrate how artificial intelligence (AI) can expand radiologists’ capacity, we visualized the features of invasive ductal carcinomas (IDCs) that our algorithm, developed and validated for basic pathological classification on mammograms, had focused on. IDC datasets were built using mammograms from patients diagnosed with IDCs from January 2006 to December 2017. The developing dataset was used to train and validate a VGG-16 deep learning (DL) network. The true positives (TPs) and accuracy of the algorithm were externally evaluated using the test dataset. A visualization technique was applied to the algorithm to determine which malignant findings on mammograms were revealed. The datasets were split into a developing dataset (988 images) and a test dataset (131 images). The proposed algorithm diagnosed 62 TPs with an accuracy of 0.61–0.70. The visualization of features on the mammograms revealed that the tubule forming, solid, and scirrhous types of IDCs exhibited visible features on the surroundings, corners of the masses, and architectural distortions, respectively. We successfully showed that features isolated by a DL-based algorithm trained to classify IDCs were indeed those known to be associated with each pathology. Thus, using AI can expand the capacity of radiologists through the discovery of previously unknown findings.

Advanced persistent threats (APT) have increased in recent times as a result of the rise in interest by nation-states and sophisticated corporations to obtain high profile information. Typically, APT attacks are more challenging to detect since they leverage zero-day attacks and common benign tools. Furthermore, these attack campaigns are often prolonged to evade detection. We leverage an approach that uses a provenance graph to obtain execution traces of host nodes in order to detect anomalous behavior. By using the provenance graph, we extract features that are then used to train an online adaptive metric learning. Online metric learning is a deep learning method that learns a function to minimize the separation between similar classes and maximizes the separation between dis-similar instances. We compare our approach with baseline models and we show our method outperforms the baseline models by increasing detection accuracy on average by 11.3 % and increases True positive rate (TPR) on average by 18.3 %.

We present development and analysis of deep learning architectures for real life and real-time native Indian monkey (langur) detection. Our monkey detection work is motivated with the idea of developing an application which will assist farmers in protecting their crops from attack of native wild animals. Protecting destruction of crops from wild animals is very important for agrarian societies. We engineered pre-trained VGG19 convolution neural network (CNN) to detect live monkeys in images streamed in an image sequence generated from a USB webcam. We used transfer learning approach to custom train, YOLO, a much faster CNN architecture to detect monkeys. For training and analysis, we used the Google Open Images V4 dataset. We give details of how two significantly different architectures were customized for native monkey detection. The performance of the detectors developed is analysed in terms of time taken for detection and accuracy of detection. Custom trained architecture achieved high true positive rate of 97.19% and true negative rate of 95.32%, which are significantly higher and better than the pre-trained architecture. The developed system was successfully tried out on a 4GB RAM and Intel quadcore i3 laptop at Indroda National park.

Two tests for covid-19 antibodies developed by the drug companies Roche and Abbott are “highly specific” but one was found to have lower sensitivity than was previously reported by the company, evaluation by Public Health England has shown. Abbott had reported that its assays had a sensitivity of 100% (the true positive rate) 14 days after symptom onset. However, PHE found that it had a sensitivity of 93.4% at 14 days, respectively. PHE published the reports after experts called for transparency in relation to the evaluation results, so that any claims could be fully scrutinised.1 Now that the evaluations have been made public, experts have said it is not clear whether these tests are any better than other available antibody tests and emphasised that the evaluations did not meet the standard needed for accreditation. Roche’s Anti-SARS-CoV-2 serology assay was tested between 5 and 7 May by the …

Background The diagnosis of Helicobacter pylori infection can be made by PCR on gastric biopsies. The objective of this study was to evaluate retrospectively the performance of the Allplex™ H pylori and ClariR PCR Assay (Seegene). Material and methods A collection of 180 DNA samples extracted from gastric biopsies was used in this study: 90 DNAs from H pylori-negative patients and 90 from H pylori-positive patients. The Allplex™ H pylori and ClariR Assay was performed on a CFX96™ real-time PCR System and analyzed using the Seegene Viewer software. The real-time PCR used as the reference was our in-house H pylori PCR, and discrepant results were tested by the Amplidiag® H pylori + ClariR PCR (Mobidiag). Results The performance of the Allplex™ H pylori and ClariR Assay showed 100% sensitivity, 97.6% specificity, 98% PPV, and 100% NPV. Regarding the detection of H pylori in the 90 expected negative samples, eight late amplifications were obtained (Ct > 39). Six of these eight samples were also positive using the Amplidiag® H pylori + ClariR kit and were therefore considered as true positives. For the two remaining cases, non-pathological evidence of H pylori infection was found. H pylori was detected in all 90 positive samples. Compared with our in-house H pylori PCR, all H pylori WT cases or mutated cases were correctly detected. Conclusions The Allplex™ H pylori and ClariR Assay showed an excellent performance and can be integrated into the armamentarium of diagnostic tests for H pylori infection. This kit has the advantage of differentiating the main mutations associated with macrolide resistance.

Social media is a platform where many young people are getting bullied. As social networking sites are increasing, cyberbullying is increasing day by day. To identify word similarities in the tweets made by bullies and make use of machine learning and can develop an ML model automatically detect social media bullying actions. However, many social media bullying detection techniques have been implemented, but many of them were textual based. The goal of this paper is to show the implementation of software that will detect bullied tweets, posts, etc. A machine learning model is proposed to detect and prevent bullying on Twitter. Two classifiers i.e. SVM and Naive Bayes are used for training and testing the social media bullying content. Both Naive Bayes and SVM (Support Vector Machine) were able to detect the true positives with 71.25% and 52.70% accuracy respectively. But SVM outperforms Naive Bayes of similar work on the same dataset. Also, Twitter API is used to fetch tweets and tweets are passed to the model to detect whether the tweets are bullying or not.

We compared the ability of 2 commercial molecular amplification assays (RealTime SARS-CoV-2 on the m2000 [abbreviated ACOV; Abbott] and ID Now COVID-19 [abbreviated IDNOW; Abbott]) and a laboratory-developed test (modified CDC 2019-nCoV reverse transcriptase PCR [RT-PCR] assay with RNA extraction by eMag [bioMérieux] and amplification on QuantStudio 6 or ABI 7500 real-time PCR system [abbreviated CDC COV]) to detect severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) RNA in upper respiratory tract specimens., We compared the ability of 2 commercial molecular amplification assays (RealTime SARS-CoV-2 on the m2000 [abbreviated ACOV; Abbott] and ID Now COVID-19 [abbreviated IDNOW; Abbott]) and a laboratory-developed test (modified CDC 2019-nCoV reverse transcriptase PCR [RT-PCR] assay with RNA extraction by eMag [bioMérieux] and amplification on QuantStudio 6 or ABI 7500 real-time PCR system [abbreviated CDC COV]) to detect severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) RNA in upper respiratory tract specimens. Discrepant results were adjudicated by medical record review. A total of 200 nasopharyngeal swab specimens in viral transport medium (VTM) were collected from symptomatic patients between 27 March and 9 April 2020. Results were concordant for 167 specimens (83.5% overall agreement), including 94 positive and 73 negative specimens. The ACOV assay yielded 33 additional positive results, 25 of which were also positive by the CDC COV assay but not by the IDNOW assay. In a follow-up evaluation, 97 patients for whom a dry nasal swab specimen yielded negative results by IDNOW had a paired nasopharyngeal swab specimen collected in VTM and tested by the ACOV assay; SARS-CoV-2 RNA was detected in 13 (13.4%) of these specimens. Medical record review deemed all discrepant results to be true positives. The IDNOW test was the easiest to perform and provided a result in the shortest time but detected fewer cases of COVID-19. The ACOV assay detected more cases of COVID-19 than the CDC COV or IDNOW assays.

An enhanced multimodal stacking scheme is proposed for quick and accurate online detection of harmful pornographic contents on the Internet. To accurately detect harmful contents, the implicative visual features (auditory features) are extracted using a bi-directional RNN (recurrent neural network) with VGG-16 (a multilayered dilated convolutional network) to implicitly express the signal change patterns over time within each input. Using only the implicative visual and auditory features, a video classifier and an audio classifier are trained, respectively. By using both features together, one fusion classifier is also trained. Then, these three component classifiers are stacked in the enhanced ensemble scheme to reduce the false negative errors in a serial order of the fusion classifier, video classifier, and audio classifier for a quick online detection. The proposed multimodal stacking scheme yields an improved true positive rate of 95.40% and a false negative rate of 4.60%, which are superior values to previous studies. In addition, the proposed stacking scheme can accurately detect harmful contents up to 74.58% and an average rate of 62.16% faster than the previous stacking scheme. Therefore, the proposed enhanced multimodal stacking scheme can be used to quickly and accurately filter out harmful contents in the online environments.

Lung cancer stands as the deadliest cancer worldwide, and early detection of pulmonary nodules is the focus of many studies to enhance the survival rate. As with many diseases, deep learning is becoming a commonly used technique for computer-aided diagnosis (CAD) in detecting lung nodules. Most lung CAD systems rely on a detection module followed by a false positive (FP) reduction module (FPR); however, FPR removes FPs as well as true positives (TPs). Thus, as a tradeoff, in order to retain high sensitivity, a large number of FPs remain. In our experience, small pulmonary vessels have been the primary source of FPs. Hence, we propose an additional module cascaded on normal FPR module to specifically reduce the number of FPs due to pulmonary vessel. Utilizing a 3D deep learning architecture, we find that the inclusion of various fields of view (FOVs) improves the accuracy of the chosen model. We explore the impact of the selection of the FOVs, the method used to integrate the features from each FOV, and using the FOV as a data augmentation method. We show that this vessel specific FPR module significantly improves the CAD system’s FP rate while only sacrificing 5% of the previously achieved sensitivity.

Rationale and Objectives To examine the effect of fatigue on diagnostic performance and interpretation of complex CT trauma studies by radiology residents. Materials and Methods Ten 2nd, 3rd, and 4th year radiology residents were given 1 hour to review as many complex CT cases as they could, once when they were fresh and once when they were fatigued (after a full day or night of diagnostic reading activity). Number of cases completed, interpretation accuracy and time per case were analyzed. Results On average readers completed 7.5 cases when fresh and 7.3 when fatigued, with 4th and 3rd year residents spending more time per case when fresh than fatigued and 2nd year residents spending more time when fatigued. The total number of true positives for major (fresh mean = 29.4 vs fatigued mean = 24.1) and minor (fresh mean = 23.30 vs fatigued mean = 18.90) findings decreased for fresh vs fatigued; and there was a significant difference as a function of year of residency (F = 4.72, p = 0.027). The 4th year residents had the most TPS in both conditions, followed by 3rd then 2nd year residents. There were more false positives when fatigued than fresh. Conclusion Fatigue significantly impacts radiologic diagnostic efficiency and efficacy, and differs as a function of year of residency.

The introduction of dedicated phlebotomy teams certified for blood collection has been reported to be highly cost-effective by reducing contamination rates. However, data on their effects on blood volume and true positive rate are limited. Therefore, we investigated the effect of replacing interns with a phlebotomy team on blood culture results. We performed a 24-month retrospective, quasi-experimental study before and after the introduction of a phlebotomy team dedicated to collecting blood cultures in a 2700-bed tertiary-care hospital. The microbiology laboratory database was used to identify adult patients with positive blood culture results. During the study period, there were no changes in blood collection method, blood culture tubes, and the application of antiseptic measures. Blood volume was measured by the BACTEC™ FX system based on red blood cell metabolism. A total of 162,207 blood cultures from 23,563 patients were analyzed, comprising 78,673 blood cultures during the intern period and 83,534 during the phlebotomy team period. Blood volume increased from a mean of 2.1 ml in the intern period to a mean of 5.6 ml in the phlebotomy team period (p

Prediction of skin sensitisation potential and potency by non-animal methods is the target of many active research programmes. Although the aim is to predict sensitisation potential and potency in humans, data from the murine local lymph node assay (LLNA) constitute much the largest source of quantitative data on in vivo skin sensitisation. The LLNA has been the preferred in vivo method for identification of skin sensitising chemicals and as such is potentially valuable as a benchmark for assessment of non-animal approaches. However, in common with all predictive test methods, the LLNA is subject to false positives and false negatives with an overall level of accuracy said variously to be approximately 80% or 90%. It is also necessary to consider the extent to which, for true positives, LLNA potency correlates with human potency. In this paper LLNA potency and human potency are compared so as to express quantitatively the correlation between them, and reasons for non-agreement between LLNA and human potency are analysed. This leads to a better definition of the applicability domain of the LLNA, within which LLNA data can be used confidently to predict human potency and as a benchmark to assess the performance of non-animal approaches.

We used the random forest classifier to predict Gross Motor Function Classification System (GMFCS) levels I-IV from patient reported abilities recorded on the Gillette Functional Assessment Questionnaire (FAQ). The classifier exhibited outstanding accuracy across GMFCS levels I-IV, with 83%-91% true positive rate (TPR), area under the receiver operation characteristic (ROC) curve greater than 0.96 for all levels, and misclassification by more than one level only occurring 1.2% of the time. This new approach to GMFCS level assignment overcomes several difficulties with the current method: (i) it is based on a broad spectrum of functional abilities, (ii) it resolves functional ability profiles that conflict with existing GMFCS level definitions, (iii) it is based entirely on self-reported abilities, and (iv) it removes complex age dependence. Further work is needed to examine inter-center differences in classifier performance-which would most likely reflect interpretive differences in GMFCS level definitions between centers.