Wrong Classification Research Articles

SSAT: Active Authorization Control and User’s Fingerprint Tracking Framework for DNN IP Protection

As training a high-performance deep neural network (DNN) model requires a large amount of data, powerful computing resources and expert knowledge, protecting well-trained DNN models from intellectual property (IP) infringement has raised serious concerns in recent years. Most existing methods using DNN watermarks to verify the ownership of the models after IP infringement occurs, which is reactive in the sense that they cannot prevent unauthorized users from using the model in the first place. Different from these methods, in this article, we propose an active authorization control and user’s fingerprint tracking method for the IP protection of DNN models by utilizing sample-specific backdoor attack. The proposed method inversely and multiplely exploits sample-specific trigger as the key to implement authorization control for DNN model, in which the generated triggers are imperceptible and sample-specific for clean images. Specifically, a U-Net model is used to generate backdoor instances. Then, the target model is trained on the clean images and backdoor instances, which are inversely labeled as wrong classes and correct classes, respectively. Only authorized users can use the target model normally by pre-processing the clean images through the U-Net model. Moreover, the images processed by the U-Net model will contain unique fingerprint that can be extracted to verify and track the corresponding user’s identity. This article is the first work that utilizes the sample-specific backdoor attack to implement active authorization control and user’s fingerprint management for DNN model under black-box scenarios. Extensive experimental results on ImageNet dataset and YouTube Aligned Face dataset demonstrate that the proposed method is effective in protecting the DNN model from unauthorized usage. Specifically, the protected model has a low inference accuracy (1.00%) for unauthorized users, while maintaining a normal inference accuracy (97.67%) for authorized users. Besides, the proposed method can achieve 100% fingerprint tracking success rates on both the ImageNet and YouTube Aligned Face datasets. Moreover, it is demonstrated that the proposed method is robust against fine-tuning attack, pruning attack, pruning attack with retraining, reverse-engineering attack, adaptive attack, and JPEG compression attack. The code is available at https://github.com/nuaaaisec/SSAT .

Intensive treatment course to identify pseudoresistant epilepsy and expedite surgery referrals - A prospective intervention study

IntroductionA significant proportion of patients do not achieve seizure freedom despite treatment attempts with two different anti-seizure medications (ASMs). A subset may not truly have drug-resistant epilepsy (“pseudoresistant”), while rapid referral of patients with genuine drug-resistant epilepsy to surgery is mandated. This study was designed to evaluate a structured and intensive treatment course with the objective of promptly identifying cases of pseudoresistance and accelerating the time to referral to epilepsy surgery. MethodsFrom May 2017 to February 2021, this prospective interventional study recruited consecutive adult patients with epilepsy treated at Odense University Hospital, Denmark, who had at least one seizure per month despite attempts with two or more ASMs. The predefined endpoint was improvement in seizure activity. Secondary endpoints were referral to epilepsy surgery, patients with pseudoresistance, and achievement of seizure freedom. ResultsOf the 41 patients enrolled, 39 completed the study. The intervention comprised a initial seizure documentation, specialist evaluation, EEG monitoring as required, and an individualized plan for intensive treatment. The plans included e.g., optimization of medical treatment, seizure classification, and improvement of medication adherence. The subsequent intensive treatment (1–4 contacts/month; 1–13 contacts in total) was led by epilepsy nurses that executed the treatment plan. The intervention significantly improved seizure control, with 41.1 % of patients achieving seizure freedom and an additional 17.8 % of patients experiencing reduced seizure frequency. One-third of the patients turned out to be “pseudoresistant” due to various reasons, including wrong classification of seizures and inadequate adherence to ASMs. Ten patients were offered a referral for epilepsy surgery at the end of the study after an average of 34.8 weeks. ConclusionThis study demonstrates the efficacy of a standardized, intensive treatment course involving epilepsy nurses in identifying and managing patients with persisting seizures despite treatment attempts with two ASMs. This approach led to favourable seizure outcomes and facilitated expedited referrals for epilepsy surgery where appropriate.

Moral reasoning in a digital age: blaming artificial intelligence for incorrect high-risk decisions

AbstractThe increasing involvement of Artificial Intelligence (AI) in moral decision situations raises the possibility of users attributing blame to AI-based systems for negative outcomes. In two experimental studies with a total of $$N = 911$$ participants, we explored the attribution of blame and underlying moral reasoning. Participants had to classify mushrooms in pictures as edible or poisonous with support of an AI-based app. Afterwards, participants read a fictitious scenario in which a misclassification due to an erroneous AI recommendation led to the poisoning of a person. In the first study, increased system transparency through explainable AI techniques reduced blaming of AI. A follow-up study showed that attribution of blame to each actor in the scenario depends on their perceived obligation and capacity to prevent such an event. Thus, blaming AI is indirectly associated with mind attribution and blaming oneself is associated with the capability to recognize a wrong classification. We discuss implications for future research on moral cognition in the context of human–AI interaction.

Prediction of Heart Disease Risk among Patients in Federal Medical Centre, Abeokuta Using Naïve Bayes

Heart disease is a silent killer, which can cause sudden death of individuals without obvious symptoms. The risk of heart disease is a sign in human beings that must not be neglected. Therefore, this study aimed to predict heart disease among patients in the Federal Medical Hospital Centre (FMC), Abeokuta. Descriptive statistics and data visualization techniques were used to gain insights into the distribution and relationships among the variables. Subsequently, a Naive Bayes classifier model was built using 80% of the data for training and 20% for testing. In addition, a Decision Tree Algorithm (DT) model was used to compare the performance of the Naive Bayes model. The performances of the two models were evaluated using accuracy, sensitivity, ROC-AUC, specificity, precision, and the F1-score. The Naive Bayes model achieved an overall accuracy of 83.61%, precision of 89.29%, recall of 78.12%, F1-Score of 83.33%, ROC-AUC of 90%, sensitivity of 78.12%, and specificity of 89.66. On the other hand, they were compared with the Decision Tree (DT) model which achieved an overall accuracy of 75.41%, precision of 77.42%, recall of 75%, F1-Score of 76.19%, ROC-AUC of 84.54%, sensitivity of 75%, and specificity of 75.86%. Similarly, the confusion matrix for both analyses gave the correct classification of 25 and 22 cases of patients who have heart disease while their wrong classification was 7 and 10 cases of patients who have no heart disease respectively. Furthermore, the importance of features carried out on both showed that the most significant features are the maximum heart rate achieved, fasting blood sugar, resting blood pressure, and chest pain respectively for Naive Bayes and Decision Tree. The findings of the analysis showed that the Naive Bayes model outperformed the Decision Tree in every aspect of the analyses in predicting the risk of heart disease based on the data used and, it suggested that medical health insurance should consider incorporating predictive modelling techniques like Naive Bayes into their risk assessment algorithms, which can be of great use in the medical line.

Classification and detection of noise in IoT based MQ gas sensors using artificial neural network

Abstract IoT-based Metal Oxide Semiconductor (MOS) gas sensors have potential applications in health, industrial, and agriculture sectors. MQ gas sensors are easy to use, have a large detection range, high sensitivity, and efficiency, and can be interfaced with Arduino for effective use. However, like other any other sensors, the MQ gas sensors are also not able to escape the effect of noise that affects the sensitivity and selectivity causing the unwanted wrong classification and identification of gases, etc finally leading to a wrong justification of results and data. Hence, the study of noise and its removal is very crucial for greater accuracy in its analysis. The main aim of our work is to check whether one can classify different types of noises or noisy signals accurately using even simple ANNs so that after testing the different filtration techniques for the signals we can obtain an independent system specifically for the MQ gas sensors, that can classify as well as filter the noisy signals based on the category to which they are classified. The best and most effective noise filtration method is then obtained. During the study Narrow Neural Network model and Medium Neural Network demonstrated high accuracy in validation and testing, with ROC curves indicating their efficiency and effectiveness.

Quantitative evaluation of internal cluster validation indices using binary data sets

AbstractAimsDifferent clustering methods often classify the same data set differently. Selecting the “best” clustering solution from alternatives is possible with cluster validation indices. Because of the large variety of cluster validation indices (CVIs), choosing the most suitable index concerning the data set and clustering algorithms is challenging. We aim to assess different internal clustering validation indices.MethodsArtificial binary data sets with equal‐ and unequal‐sized well‐separated a priori clusters were simulated and three levels of noise were then added. Twenty replications of each of the six types of data sets (two group sizes × three levels of noise) were created and analyzed by three clustering algorithms with Jaccard dissimilarity. Twenty‐seven clustering validation indices are evaluated including both geometric and non‐geometric indices.ResultsAlthough, in theory, all CVIs could differentiate between good and wrong classifications, only a few perform as expected with noisy data. Tau and silhouette widths proved to be the best geometric CVIs both for equal and unequal cluster sizes. Among non‐geometric indices, crispness and OptimClass performed best.ConclusionWe recommend using these best‐performing CVIs. We suggest plotting the CVI value against the number of clusters because the lack of a sharp peak means that the position of the maximum is uncertain.

Improved medical image inpainting using automatic multi-task learning driven deep learning approach

Distorted medical images can drastically reduce diagnosis accuracy using computer-aided diagnostic (CAD) systems. The objective of medical image classification is to improve diagnostic imaging precision and restore regions degraded by image inpainting. Deep Learning (DL) methods like Generative Adversarial Networks (GANs) are being used to solve these medical image inpainting issues. Recently, various improved GAN models were proposed for improved image inpainting, but challenges like computational inefficiency, unreliable restoration quality, and lack of correct analysis of discriminator block. To overcome these problems, a novel GAN model is proposed with modified Generator and Discriminator blocks for improved medical image inpainting. To improve image inpainting quality with minimum computational complexity, the Multi-Task Learning (MTL)-based generator block is designed. The MTL encoder is connected with three decoders for three tasks edge detection, organ boundary generation, and image completion. The MTL-based encoder and decoders are modified with an optimum number of layers and regularization functions to improve the outcomes. The output of the generator block is input into the discriminator block, where we designed the relevant feedback strategy to decrease generator loss by automatically changing the weights on wrong classifications. The simulation results revealed the efficiency of the proposed medical image inpainting approach and discriminator block performances over the state-of-the-art. Due to the modified MTL-based generator and discriminator block in the proposed approach, the PSNR is improved by 14.31 %, SSIM is improved by 8.56%, UQI is improved by 9.45%, and MSE is reduced by 22.78% using the proposed model compared to the state-of-the-arts.

Convolutional neural networks to classify human stress that occurs during in‐field sugarcane harvesting: A case study

AbstractAssessing human stress in agriculture proves to be a complex and time‐intensive endeavor within the field of ergonomics, particularly for the development of agricultural systems. This methodology involves the utilization of instrumentation and the establishment of a dedicated laboratory setup. The complexity arises from the need to capture and analyze various physiological and psychological indicators, such as heart rate (HR), muscle activity, and subjective feedback to comprehensively assess the impact of farm operations on subjects. The instrumentation typically includes wearable devices, sensors, and monitoring equipment to gather real‐time data of subject during the performance of farm operations. Deep learning (DL) models currently achieve human performance levels on real‐world face recognition tasks. In this study, we went beyond face recognition and experimented with the recognition of human stress based on facial features during the drudgery‐prone agricultural operation of sugarcane harvesting. This is the first research study for deploying artificial intelligence‐driven DL techniques to identify human stress in agriculture instead of monitoring several ergonomic characteristics. A total of 20 (10 each for male and female) subjects comprising 4300 augmented RGB images (215 per subject) were acquired during sugarcane harvesting seasons and then these images were deployed for training (80%) and validation (20%). Human stress and nonstress states were determined based on four ergonomic physiological parameters: heart rate (ΔHR), oxygen consumption rate (OCR), energy expenditure rate (EER), and acceptable workload (AWL). Stress was defined when ΔHR, OCR, EER, and AWL reached or exceeded certain standard threshold values. Four convolutional neural network‐based DL models (1) DarkNet53, (2) InceptionV3, (3) MobileNetV2 and (4) ResNet50 were selected due to their remarkable feature extraction abilities, simple and effective implementation to edge computation devices. In all four DL models, training performance results delivered training accuracy ranging from 73.8% to 99.1% at combinations of two mini‐batch sizes and four levels of epochs. The maximum training accuracies were 99.1%, 99.0%, 97.7%, and 95.4% at the combination of mini‐batch size 16 and 25 epochs for DarkNet53, InceptionV3, ResNet50, and MobileNetV2, respectively. Due to the best performance, DarkNet53 was tested further on an independent data set of 100 images and found 89.8%–93.3% confident to classify stressed images for female subjects while 92.2%–94.5% for male subjects, though it was trained on the integrated data set. The comparative classification of the developed model and ergonomic measurements for stress classification was carried out with a net accuracy of 88% where there were few instances of wrong classifications.

Simplified Convolutional Neural Network Model for Automatic Classification of Retinal Diseases from Optical Coherence Tomography Images

Optical coherence tomography (OCT) allows for direct and immediate imaging of the morphology of retinal tissue. It has become a crucial imaging modality for diagnosing eye problems in ophthalmology. One of the most significant morphological characteristics of the retina is the structure of the retinal layers, which provides important evidence for diagnostic purposes and is related to a variety of retinal diseases. In this paper, a convolutional neural network (CNN) model is proposed that can identify the difference between a normal retina and three common macular diseases: Diabetic macular edema (DME), Drusen, and Choroidal neovascularization (CNV). This proposed model was trained and tested on an open source dataset of OCT images also with professional disease classifications such as DME, CNV, Drusen, and Normal. The suggested model has achieved 98.3% overall classification accuracy, with only 7 wrong classifications out of 368 test samples. The suggested model significantly outperforms other models that made use of the identical dataset. The final results show that the suggested model is particularly adapted to the detection of retinal disorders in ophthalmology centers.

The consequences of the new European reclassification of non-invasive brain stimulation devices and the medical device regulations pose an existential threat to research and treatment: An invited opinion paper

A significant amount of European basic and clinical neuroscience research includes the use of transcranial magnetic stimulation (TMS) and low intensity transcranial electrical stimulation (tES), mainly transcranial direct current stimulation (tDCS). Two recent changes in the EU regulations, the introduction of the Medical Device Regulation (MDR) (2017/745) and the Annex XVI have caused significant problems and confusions in the brain stimulation field. The negative consequences of the MDR for non-invasive brain stimulation (NIBS) have been largely overlooked and until today, have not been consequently addressed by National Competent Authorities, local ethical committees, politicians and by the scientific communities. In addition, a rushed bureaucratic decision led to seemingly wrong classification of NIBS products without an intended medical purpose into the same risk group III as invasive stimulators.Overregulation is detrimental for any research and for future developments, therefore researchers, clinicians, industry, patient representatives and an ethicist were invited to contribute to this document with the aim of starting a constructive dialogue and enacting positive changes in the regulatory environment.

A Bayesian deep learning framework for reliable fault diagnosis in wind turbine gearboxes under various operating conditions

Vibration-based fault diagnostics combined with deep learning approaches has promising applications in detecting and diagnosing faults in wind turbine gearboxes. Specifically when time series vibration data is transformed to a 2-dimensional cyclic spectral coherence maps, the accuracy of deep neural networks in classifying faults increases. Nevertheless, standard deep learning techniques are vulnerable to inaccurate predictions when tested with new data originating from unseen faults or unusual operating conditions. To address some of these shortcomings in the context of wind turbine gearboxes, this paper investigates fault diagnostics using Bayesian convolutional neural network which provide accurate results with uncertainty bounds reducing wrong overconfident classifications. The performance of Bayesian and standard neural networks is compared using a simulation-based dataset of acceleration signals generated from a multibody dynamic model of a 5 MW wind turbine. The framework proposed in this paper has relevance to fault detection and diagnosis in other rotating machinery applications.

Determination of the Diagnostic Performance of Laboratory Tests in the Absence of a Perfect Reference Standard: The Case of SARS-CoV-2 Tests.

Currently, assessing the diagnostic performance of new laboratory tests assumes a perfect reference standard, which is rarely the case. Wrong classifications of the true disease status will inevitably lead to biased estimates of sensitivity and specificity. Using Bayesian' latent class models (BLCMs), an approach that does not assume a perfect reference standard, we re-analyzed data of a large prospective observational study assessing the diagnostic accuracy of an antigen test for the diagnosis of SARS-CoV-2 infection in clinical practice. A cohort of consecutive patients presenting to a COVID-19 testing facility affiliated with a Swiss University Hospital were recruited (n = 1465). Two real-time PCR tests were conducted in parallel with the Roche/SD Biosensor rapid antigen test on nasopharyngeal swabs. A two-test (PCR and antigen test), three-population BLCM was fitted to the frequencies of paired test results. Based on the BLCM, the sensitivities of the RT-PCR and the Roche/SD Biosensor rapid antigen test were 98.5% [95% CRI 94.8;100] and 82.7% [95% CRI 66.8;100]. The specificities were 97.7% [96.1;99.7] and 99.9% [95% CRI 99.6;100]. Applying the BLCM, the diagnostic accuracy of RT-PCR was high but not perfect. In contrast to previous results, the sensitivity of the antigen test was higher. Our results suggest that BLCMs are valuable tools for investigating the diagnostic performance of laboratory tests in the absence of perfect reference standard.

Using adaptive learning rate to generate adversarial images

Abstract Convolutional neural networks (CNNs) have proved their efficiency in performing image classification tasks, as they can automatically extract the image features and make the corresponding prediction. Meanwhile, the CNNs application is highly challenged by their vulnerability to adversarial samples. These samples are slightly different from the legitimate samples, but the CNN gives wrong classification. There are various ways to find the adversarial samples. The most common method is using backpropagation to generate gradients as the directed perturbation. Contrarily to set a constrained limitation, in this paper, we use iterative fast gradient sign method to generate adversarial images with the minimum perturbation. The CNNs were trained to perform surgical tool recognition as a configuration for the modern operation room. The coefficient or the learning rate which influenced the modification per iteration, was set to be adaptive instead of a fixed number. A few functions were utilized to perform the learning rate decay to compare the performance. Especially, we propose a new adaptive learning rate algorithm that consider the loss as a part of influence factor constitute the learning rate for the rest iterations. According to the experiments, our loss adaptive learning rate method was proved to be efficient to get the minimal perturbations for adversarial attack.

Improved Two Stage Generative Adversarial Networks for Adversarial Example Generation with Real Exposure

Introduction: Deep neural networks due to their linear nature are sensitive to adversarial examples. They can easily be broken just by a small disturbance to the input data. Some of the existing methods to perform these kinds of attacks are pixel-level perturbation and spatial transformation of images. Method: These methods generate adversarial examples that can be fed to the network for wrong predictions. The drawback that comes with these methods is that they are really slow and computationally expensive. This research work performed a black box attack on the target model classifier by using the generative adversarial networks (GAN) to generate adversarial examples that can fool a classifier model to classify the images as wrong classes. The proposed method used a biased dataset that does not contain any data of the target label to train the first generator Gnorm of the first stage GAN, and after the first training has finished, the second stage generator Gadv, which is a new generator model that does not take random noise as input but the output of the first generator Gnorm. Result: The generated examples have been superimposed with the Gnorm output with a small constant, and then the superimposed data have been fed to the target model classifier to calculate the loss. Some additional losses have been included to constrain the generation from generating target examples. Conclusion: The proposed model has shown a better fidelity score, as evaluated using Fretchet inception distance score (FID), which was up to 42.43 in the first stage and up to 105.65 in the second stage with the attack success rate of up to 99.13%.

A two-stage Kriging estimation variance reduction method for efficient time-variant reliability-based design optimization

Adaptive Kriging modelling technique is an effective tool to reduce the computational burden for time-variant reliability-based design optimization (TRBDO) significantly. However, the existing methods may allocate sampling resources to unimportant regions and do not consider the correlations between time trajectories, both of which will waste some computationally expensive samples. In this paper, we propose a two-stage Kriging estimation variance reduction method (KEVAR2) to address these challenges. First, the expression of Kriging estimation variance is derived, which can quantify the contribution of correlations between time trajectories to the Kriging prediction uncertainty. Second, the global and local adaptive Kriging approaches are proposed based on the Kriging estimation variance reduction strategy. Meanwhile, two metrics, i.e., the expectation of wrong classification rate and the error of safe rate are proposed to quantify global and local errors of Kriging models respectively. After that, we propose a two-stage framework that integrates the global and local adaptive Kriging approaches together, where the first stage performs global adaptive sampling to quickly reduce the Kriging prediction error, and then the second stage performs local adaptive sampling in the region near the current optimal solution to achieve high efficiency. Finally, five case studies demonstrate the effectiveness of the proposed KEVAR2.

Criminalistic Aspects of Classifying Investigatory Actions as One of the Key Means of Proof

The paper contains an analysis of the concept of means of proof. Based on the etymology of the term “means” as a technique and mode of action in order to achieve something or as a tool in some activity, it is concluded that means of proof should include, fist, evidence and any information that the court, the prosecutor, the investigator or the inquirer use to establish the circumstances that need to be proven in criminal proceedings, as well as other circumstances relevant to the criminal case, second, samples for comparative research that do not have any independent significance, but are used for comparison with physical evidence and other objects, third, investigatory and other procedural actions that are the key method of collecting evidence for the investigator (the inquirer), fourth, tactics that are usually determined as most rational, effective modes of action or most expedient behavior of the person performing investigatory actions, fifth, operative search activities whose results may pose grounds for initiating a criminal case, be presented to the investigative body, to the investigator or the court responsible for the criminal case or materials of verifying a report on crime, sixth, tactical (operative-tactical) combinations (operations). Further the authors examine the concept and classification of investigative actions according to the subject authorizing investigative actions, to the methods they are based on, to the subjects performing such actions, depending on the necessity of involving witnesses, of involving experts, depending on the use of technical equipment, based on the criteria of the speed of fixing the traces of crime, and according to the stages of an investigation. It is concluded that classifications are a methodological basis for solving both theoretical and practical problems, and even wrong classifications play a positive role in the progress of this science because they allow us to reject the flawed option and quicker determine the way to the right solution of the problems that the science faces.

Impact of synthetic noise signature and physiologic ECG signal on designing ML-based ECG noise detection framework.

Automatic signal analysis using artificial intelligence is getting popular in digital healthcare, such as ECG rhythm analysis, where ECG signals are collected from traditional ECG machines or wearable ECG sensors. However, the risk of using an automated system for ECG analysis when noise is present can lead to incorrect diagnosis or treatment decisions. A noise detector is crucial to minimise the risk of incorrect diagnosis. Machine learning (ML) models are used in ECG noise detection before clinical decision-making systems to mitigate false alarms. However, it is essential to prove the generalisation capability of the ML model in different situations. ML models performance is 50% lesser when the model is trained with synthetic and tested with physiologic ECG datasets compared to trained and tested with physiologic ECG datasets. This suggests that the ML model must be trained with physiologic ECG datasets rather than synthetic ones or add more various types of noise in synthetic ECG datasets that can mimic physiologic ECG.Clinical relevance- ML model trained with synthetic noisy ECG can increase the 50% misclassification rate in ECG noise detection compared to training with physiologic ECG datasets. The wrong classification of noise-free and noisy ECG will lead to misdiagnosis regarding the patient's condition, which could be a cause of death.

우리나라 교육과정 통해 살펴본 무용 교육

The purpose of this study is to examine the characteristics of the art/PE curriculum from the perspective of dance, from Syllabus Period to the current revised curriculum in 2015. In order to achieve this purpose, the curriculum document was used as an analysis target to find the context through the overall process of change and to find the historical change and subject identity of the dance in the curriculum.
 First, during the introduction of the dance curriculum (Syllabus Period to the 1st curriculum), the dance department exposed its limitations in taking root in the curriculum due to the natural background of not originally starting as an independent subject like art and music. Second, during the settlement of the dance curriculum (the 2nd to the 4th curriculum), dance was presented as one of the contents of the PE curriculum, developing as a discipline, and many universities established dance departments. Third, the transition period of the dance curriculum (the 5th to the 7th curriculum) was a time when it was organized and oriented as an integrated subject, and as it became possible to select between music and art, it was actually divided into art and PE subjects. Fourth, during the transformation period of the dance curriculum (Revised Curriculum in 2009 to Revised Curriculum in 2015), as the boundary between art and physical education subjects was established, the term ‘expression’ is presented in the domain instead of the term ‘dance’.
 Moreover, as plays are additionally included in the arts subject, dance is adding to the identity crisis and seriousness as a subject. Through the upcoming 2022 revised curriculum and 2025 high school credit system, therefore, I hope that dance subjects can improve the wrong curriculum classification standards so far.

Boosting Semi-Supervised Semantic Segmentation with Probabilistic Representations

Recent breakthroughs in semi-supervised semantic segmentation have been developed through contrastive learning. In prevalent pixel-wise contrastive learning solutions, the model maps pixels to deterministic representations and regularizes them in the latent space. However, there exist inaccurate pseudo-labels which map the ambiguous representations of pixels to the wrong classes due to the limited cognitive ability of the model. In this paper, we define pixel-wise representations from a new perspective of probability theory and propose a Probabilistic Representation Contrastive Learning (PRCL) framework that improves representation quality by taking its probability into consideration. Through modelling the mapping from pixels to representations as the probability via multivariate Gaussian distributions, we can tune the contribution of the ambiguous representations to tolerate the risk of inaccurate pseudo-labels. Furthermore, we define prototypes in the form of distributions, which indicates the confidence of a class, while the point prototype cannot. More- over, we propose to regularize the distribution variance to enhance the reliability of representations. Taking advantage of these benefits, high-quality feature representations can be derived in the latent space, thereby the performance of se- mantic segmentation can be further improved. We conduct sufficient experiment to evaluate PRCL on Pascal VOC and CityScapes to demonstrate its superiority. The code is available at https://github.com/Haoyu-Xie/PRCL.

Does the Stopping Opioids After Surgery Score Perform Well Among Racial and Socioeconomic Subgroups?

The Stopping Opioids After Surgery (SOS) score is a validated tool that was developed to determine the risk of sustained opioid use after surgical interventions, including orthopaedic procedures. Despite prior investigations validating the SOS score in diverse contexts, its performance across racial, ethnic, and socioeconomic subgroups has not been assessed. In a large, urban, academic health network, did the performance of the SOS score differ depending on (1) race and ethnicity or (2) socioeconomic status? This retrospective investigation was conducted using data from an internal, longitudinally maintained registry of a large, urban, academic health system in the Northeastern United States. Between January 1, 2018, and March 31, 2022, we treated 26,732 adult patients via rotator cuff repair, lumbar discectomy, lumbar fusion, TKA, THA, ankle or distal radius open reduction and internal fixation, or ACL reconstruction. We excluded 1% of patients (274 of 26,732) because of missing length of stay information, 0.06% (15) for missing discharge information, 1% (310) for missing medication information related to loss to follow-up, and 0.07% (19) who died during their hospital stay. Based on these inclusion and exclusion criteria, 26,114 adult patients were left for analysis. The median age in our cohort was 63 years (IQR 52 to 71), and most patients were women (52% [13,462 of 26,114]). Most patients self-reported their race and ethnicity as non-Hispanic White (78% [20,408 of 26,114]), but the cohort also included non-Hispanic Black (4% [939]), non-Hispanic Asian (2% [638]), and Hispanic (1% [365]) patients. Five percent (1295) of patients were of low socioeconomic status, defined by prior SOS score investigations as patients with Medicaid insurance. Components of the SOS score and the observed frequency of sustained postoperative opioid prescriptions were abstracted. The performance of the SOS score was compared across racial, ethnic, and socioeconomic subgroups using the c-statistic, which measures the capacity of the model to differentiate between patients with and without sustained opioid use. This measure should be interpreted on a scale between 0 and 1, where 0 represents a model that perfectly predicts the wrong classification, 0.5 represents performance no better than chance, and 1.0 represents perfect discrimination. Scores less than 0.7 are generally considered poor. The baseline performance of the SOS score in past investigations has ranged from 0.76 to 0.80. The c-statistic for non-Hispanic White patients was 0.79 (95% CI 0.78 to 0.81), which fell within the range of past investigations. The SOS score performed worse for Hispanic patients (c-statistic 0.66 [95% CI 0.52 to 0.79]; p < 0.001), where it tended to overestimate patients' risks of sustained opioid use. The SOS score for non-Hispanic Asian patients did not perform worse than in the White patient population (c-statistic 0.79 [95% CI 0.67 to 0.90]; p = 0.65). Similarly, the degree of overlapping CIs suggests that the SOS score did not perform worse in the non-Hispanic Black population (c-statistic 0.75 [95% CI 0.69 to 0.81]; p = 0.003). There was no difference in score performance among socioeconomic groups (c-statistic 0.79 [95% CI 0.74 to 0.83] for socioeconomically disadvantaged patients; 0.78 [95% CI 0.77 to 0.80] for patients who were not socioeconomically disadvantaged; p = 0.92). The SOS score performed adequately for non-Hispanic White patients but performed worse for Hispanic patients, where the 95% CI nearly included an area under the curve value of 0.5, suggesting that the tool is no better than chance at predicting sustained opioid use for Hispanic patients. In the Hispanic population, it commonly overestimated the risk of opioid dependence. Its performance did not differ among patients of different sociodemographic backgrounds. Future studies might seek to contextualize why the SOS score overestimates expected opioid prescriptions for Hispanic patients and how the utility performs among more specific Hispanic subgroups. The SOS score is a valuable tool in ongoing efforts to combat the opioid epidemic; however, disparities exist in terms of its clinical applicability. Based on this analysis, the SOS score should not be used for Hispanic patients. Additionally, we provide a framework for how other predictive models should be tested in various lesser-represented populations before implementation.