Erez Shmueli’s research while affiliated with Tel Aviv University and other places

Background Public reluctance to receive COVID-19 vaccination is associated with safety concerns. By contrast, the seasonal influenza vaccine has been administered for decades with a solid safety record and a high level of public acceptance. We compare the safety profile of the BNT162b2 COVID-19 booster vaccine to that of the seasonal influenza vaccine. Methods We study a prospective cohort of 5079 participants in Israel and a retrospective cohort of 250,000 members of MHS selected randomly. We examine reactions to BNT162b2 mRNA COVID-19 booster and to influenza vaccinations. All prospective cohort participants wore a smartwatch and completed a daily digital questionnaire. We compare pre-vaccination and post-vaccination smartwatch heart-rate data, and a stress measure based on heart-rate variability. We also examine adverse events from electronic health records. Results In the prospective cohort, 1905 participants receive the COVID-19 booster vaccine; 899 receive influenza vaccination. Focusing on those who receive both vaccines yields a total of 689 participants in the prospective cohort and 31,297 members in the retrospective cohort. Individuals reporting a more severe reaction after influenza vaccination tend to likewise report a more severe reaction after COVID-19 vaccination. In paired analysis, the increase in both heart rate and stress measure for each participant is higher for COVID-19 than for influenza in the first 2 days after vaccination. No elevated risk of hospitalization due to adverse events is found following either vaccine. Except for Bell’s palsy after influenza vaccination, no elevated risk of adverse events is found. Conclusions The more pronounced side effects after COVID-19 vaccination may explain the greater concern associated with it. Nevertheless, our comprehensive analysis supports the safety profile of both vaccines.

Deep learning has become a very popular method for text classification in recent years, due to its ability to improve the accuracy of previous state-of-the-art methods on several benchmarks. However, these improvements required hundreds of thousands to millions labeled training examples, which in many cases can be very time consuming and/or expensive to acquire. This problem is especially significant in domain specific text classification tasks where pretrained embeddings and models are not optimal. In order to cope with this problem, we propose a novel learning framework, Ensembled Transferred Embeddings (ETE), which relies on two key ideas: (1) Labeling a relatively small sample of the target dataset, in a semi-automatic process (2) Leveraging other datasets from related domains or related tasks that are large-scale and labeled, to extract “transferable embeddings” Evaluation of ETE on a large-scale real-world item categorization dataset provided to us by PayPal, shows that it significantly outperforms traditional as well as state-of-the-art item categorization methods.KeywordsText classificationDeep neural networksEmbeddingsTransfer learningMachine Learning

Item categorization is a machine learning task which aims at classifying e-commerce items, typically represented by textual attributes, to their most suitable category from a predefined set of categories. An accurate item categorization system is essential for improving both the user experience and the operational processes of the company. In this work, we focus on item categorization settings in which the textual attributes representing items are noisy and short, and labels (i.e., accurate classification of items into categories) are not available. In order to cope with such settings, we propose a novel learning framework, Ensembled Transferred Embeddings (ETE), which relies on two key ideas: 1) labeling a relatively small sample of the target dataset, in a semi-automatic process, and 2) leveraging other datasets from related domains or related tasks that are large-scale and labeled, to extract "transferable embeddings". Evaluation of ETE on a large-scale real-world dataset provided to us by PayPal, shows that it significantly outperforms traditional as well as state-of-the-art item categorization methods.

Background: Clinical trial guidelines for assessing the safety of vaccines, including the FDA criteria, are primarily based on subjective, self-reported questionnaires. Despite the tremendous technological advances in recent years, objective, continuous assessment of physiological measures post-vaccination is rarely performed. Methods: To evaluate the short-term effects of the BNT162b2 COVID-19 vaccine on physiological measures, we conducted a prospective observational study during the mass vaccination campaign in Israel. 160 individuals >18 years who were not previously found to be COVID-19 positive and who received the second dose of the COVID-19 vaccine between 1 January 2021, and 13 March 2021 were equipped with a chest-patch sensor and a dedicated mobile application. The chest-patch sensor continuously measured 13 physiological vitals one day before the inoculation (baseline), for four days: heart rate, blood oxygen saturation, respiratory rate, systolic and diastolic blood pressure, pulse pressure, mean arterial pressure, heart rate variability, stroke volume, cardiac output, cardiac index, systemic vascular resistance, and body temperature. The mobile application collected daily self-reported questionnaires starting one day before the inoculation, for 15 days on local and systemic reactions, sleep quality, stress levels, physical activity, and mood levels. Findings: Within the first 48 hours post-vaccination, we identified significant changes (p-value <0.05) in nearly all 13 chest-patch indicators compared to their baseline levels. 48.5% (n=78) reported no local or systemic reaction. Nevertheless, we identified considerable changes in chest-patch indicators during the first 48 hours post-vaccination also in this group of presumably asymptomatic participants. Within three days from vaccination, these measures returned to baseline levels in both groups, further supporting the safety of the vaccine. Interpretation: Our work underscores the importance of obtaining objective physiological data in addition to self-reported questionnaires when performing clinical trials, particularly in ones conducted in very short time frames. Funding: The European Research Council (ERC) project #949850.

Current efforts for COVID-19 screening mainly rely on reported symptoms and potential exposure to infected individuals. Here, we developed a machine-learning model for COVID-19 detection that utilizes four layers of information: 1) sociodemographic characteristics of the tested individual, 2) spatiotemporal patterns of the disease observed near the testing episode, 3) medical condition and general health consumption of the tested individual over the past five years, and 4) information reported by the tested individual during the testing episode. We evaluated our model on 140,682 members of Maccabi Health Services, tested for COVID-19 at least once between February and October 2020. These individuals had 264,516 COVID-19 PCR-tests, out of which 16,512 were found positive. Our multilayer model obtained an area under the curve (AUC) of 81.6% when tested over all individuals, and of 72.8% when tested over individuals who did not report any symptom. Furthermore, considering only information collected before the testing episode – that is, before the individual may had the chance to report on any symptom – our model could reach a considerably high AUC of 79.5%. Namely, most of the value contributed by the testing episode can be gained by earlier information. Our ability to predict early the outcomes of COVID-19 tests is pivotal for breaking transmission chains, and can be utilized for a more efficient testing policy.

We propose a novel approach for textual input which is based on air-writing recognition using smart-bands. The proposed approach enables the user to hand-write in the air in an intuitive and natural way, where text is recognized by analyzing the motion signals captured by an off-the-shelf smart-band worn by the user. Unlike existing studies that proposed the use of motion signals to recognize written letters, our approach does not require an extra dedicated device, nor it imposes unnecessary limitations on the writing process of the user. To test the feasibility of the new approach, we developed two air-writing recognition methods: a user-dependent method, based on K-Nearest-Neighbors with Dynamic-Time-Warping as the distance measure, and a user-independent method, based on a Convolutional-Neural-Network. The first creates a tailored model for each user, using a set of reference samples collected from the user in an enrollment phase, and therefore has the potential to be more accurate. The latter involves a preliminary training phase which generates a single model to fit all users, and therefore does not require an enrollment phase for new users. In order to evaluate our methods, we collected 15 sets of the English alpha-bet letters (written on the air and collected using a smart-band) from 55 different subjects. The results of our evaluation demonstrate the ability of the proposed methods to successfully recognize air-written letters with a high degree of accuracy, obtaining 89.2% average accuracy for the user-dependent method, and 83.2% average accuracy (95.6% when applying an auto-correction phase) for the user-independent method.

Pertussis incidence in developed countries, including Israel, has increased over the past two decades despite the addition of two booster doses in children. However, as pertussis is characterized by a multi-annual periodicity, and since clinical diagnosis can miss cases, determining disease trends at the population level is challenging. To bridge this gap, we developed a simple statistical model to capture the temporal patterns of pertussis incidence in Israel. Our model was calibrated and tested using laboratory-confirmed cases of pertussis for the Israeli population between 1998 and 2019. The model identifies a clear four-year periodicity of pertussis incidence over the past two decades that is identical to the one observed in the pre-vaccine era. Accounting for this periodicity, the model shows a 325% increase in pertussis incidence from 2002 to 2014. These multi-year trends were interrupted shortly after the introduction of routine immunization of Tdap vaccine in pregnancy in 2015, after which we found a 59.7% (95% CI: 57.7-61.6%)decline in pertussis incidence and a 49.5% (36.0-61.6%) decline in hospitalizations compared to the model's projection. While this sharp decline cannot be fully attributed to the newly introduced vaccination policy, sharper reductions of 71.2% (65.6-76.1%) in incidence and 58.4% (39.6-72.7%) in hospitalizations, have been observed in infants of age two months and below - young infants that have yet to become vaccinated and are more likely to be protected by maternal vaccination. Our work suggests that Tdap vaccination during pregnancy is a promising policy for controlling pertussis. Furthermore, due to the stable periodicity of pertussis, public health decision-makers should invest continuous efforts in the implementation of this strategy with additional reinforcement in expected peak years.

Label ranking is a prediction task which deals with learning a mapping between an instance and a ranking (i.e., order) of labels from a finite set, representing their relevance to the instance. Boosting is a well-known and reliable ensemble technique that was shown to often outperform other learning algorithms. While boosting algorithms were developed for a multitude of machine learning tasks, label ranking tasks were overlooked. In this paper, we propose a boosting algorithm which was specifically designed for label ranking tasks. Extensive evaluation of the proposed algorithm on 24 semi-synthetic and real-world label ranking datasets shows that it significantly outperforms existing state-of-the-art label ranking algorithms.

Label ranking tasks are concerned with the problem of ranking a finite set of labels for each instance according to their relevance. Boosting is a well-known and reliable ensemble technique that was shown to often outperform other learning algorithms. While boosting algorithms were developed for a multitude of machine learning tasks, label ranking tasks were overlooked. Herein, we present a novel boosting algorithm, BoostLR, that was specifically designed for label ranking tasks. Similarly to other boosting algorithms, BoostLR, proceeds in rounds, where in each round, a single weak model is trained over a sampled set of instances. Instances that were identified as harder to predict in the current round, receive a higher (boosted) weight, and therefore also a higher probability to be included in the sample of the forthcoming round. Extensive evaluation of our proposed algorithm on 24 semi-synthetic and real-world label ranking datasets concludes that our algorithm significantly outperforms the current state-of-the-art label ranking methods.

Recommender systems have become extremely common in recent years, and are applied in a variety of domains. Existing recommender systems exhibit two major limitations: (1) Privacy - each service provider holds a database that contains information about all of its users; and (2) Partial view - when recommending to users, each such service can rely only on data that were collected by the service itself. The Open Personal Data Store (openPDS) architecture was recently suggested for storing personal data in a privacy preserving way. Inspired by openPDS, we suggest a novel architecture for recommender systems that overcomes the two limitations mentioned above. The suggested architecture allows the recommender system to utilize rich data collected about the user (possibly through other services) to produce more accurate recommendations, while allowing its users to manage and gain control over their own data. We evaluate the suggested architecture on two different use cases: movies and web browsing, and compare its performance with that of a popular non-privacy-aware collaborative-filtering algorithm. We find that in comparison to the alternative approach, our approach is able to enhance privacy significantly without sacrificing the accuracy level of the recommendations (and in some cases providing even higher level of accuracy).