Publications

Safety reporting is largely neglected across trials of mental-health-related interventions, thus hindering the assessment of risk-benefit ratios for rational decision making in mental health care.

Scientific journals may counter the misuse, misreporting, and misinterpretation of statistics by providing guidance to authors. We assessed the nature and prevalence of statistical guidance at 15 journals (top-ranked by Impact Factor) in each of 22 scientific disciplines across five high-level domains (N = 330 journals). The frequency of statistical guidance varied across domains (Health & Life Sciences: 122/165 journals, 74%; Multidisciplinary: 9/15 journals, 60%; Social Sciences: 8/30 journals, 27%; Physical Sciences: 21/90 journals, 23%; Formal Sciences: 0/30 journals, 0%). In one discipline (Clinical Medicine), statistical guidance was provided by all journals and in two disciplines (Mathematics and Computer Science) no journals provided statistical guidance. Of the 160 journals providing statistical guidance, 93 had a dedicated statistics section in their author instructions. The most frequently mentioned topics were confidence intervals (90 journals) and p-values (88 journals). For six ‘hotly debated’ topics (statistical significance, p-values, Bayesian statistics, effect sizes, confidence intervals, and sample size planning/justification) journals typically offered implicit or explicit endorsement and rarely provided opposition. The heterogeneity of statistical guidance provided by top-ranked journals within and between disciplines highlights a need for further research and debate about the role journals can play in improving statistical practice.

To achieve high bit rates link designers are using more sophisticated communication techniques, often turning to 4PAM transmission or decision-feedback equalization (DFE). Interestingly, with only minor modification the same hardware needed to implement a 4PAM system can be used to implement a loop-unrolled single-tap DFE receiver. To get the maximum performance from either technique, the link has to be tuned to match the specific channel it is driving. Adaptive equalization using data based update filtering allows continuous updates while minimizing the required sampler front-end hardware and significantly reduces the cost of implementation in multi-level signaling schemes. A transceiver chip was designed and fabricated in 0.13 /spl mu/m CMOS process to investigate dual-mode operation and the modifications of the standard adaptive algorithms necessary to operate in high-speed link environments.

ABSTRACT Background COVID-19 exhibits a steep age gradient of infection fatality rate. There has been debate about whether extra protection of elderly and other vulnerable individuals (precision shielding) is feasible, and, if so, to what extent. Methods We used systematically retrieved data from national seroprevalence studies conducted in the pre-vaccination era. Studies were identified through SeroTracker and PubMed searches (last update May 17, 2022). Studies were eligible if they targeted representative general populations without high risk of bias. Seroprevalence estimates were noted for children, non-elderly adults, and elderly adults, using cut-offs of 20, and 60 years (or as close to these ages, if they were not available). Results Thirty-eight national seroprevalence studies from 36 different countries were included in the analysis. 26/38 also included pediatric populations. 25/38 studies were from high-income countries. The median ratio of seroprevalence in the elderly versus non-elderly adults (or non-elderly in general, if pediatric and adult population data were not offered separately) was 0.90-0.95 in different analyses with large variability across studies. In 5 studies (all of them in high-income countries), there was significant protection of the elderly with ratio <0.40. The median was 0.83 in high-income countries and 1.02 in other countries. The median ratio of seroprevalence in children versus adults was 0.89 and only one study showed a significant ratio of <0.40. Conclusion Precision shielding of elderly community-dwelling populations before the availability of vaccines was feasible in some high-income countries, but most countries failed to achieve any substantial focused protection of this age group. summary 38 COVID-19 nationally representative seroprevalence studies conducted before vaccination campaigns were systematically identified. Median seroprevalence ratio in elderly versus non-elderly adults was 0.90-0.95, indicating no generally achieved precision shielding of elderly. In 5 studies, substantial protection (ratio <0.40) was observed.

Despite a few hundred systematic reviews and meta-analyses, highly convincing evidence of a clear role of vitamin D does not exist for any outcome, but associations with a selection of outcomes are probable.

State-of-the-art deep learning models for computer vision tasks are based on the transformer architecture and often deployed in real-time applications. In this scenario, the resources available for every inference can vary, so it is useful to be able to dynamically adapt execution to trade accuracy for efficiency. To create dynamic models, we leverage the resilience of vision transformers to pruning and switch between different scaled versions of a model. Surprisingly, we find that most FLOPs are generated by convolutions, not attention. These relative FLOP counts are not a good predictor of GPU performance since GPUs have special optimizations for convolutions. Some models are fairly resilient and their model execution can be adapted without retraining, while all models achieve better accuracy with retraining alternative execution paths. These insights mean that we can leverage CNN accelerators and these alternative execution paths to enable efficient and dynamic vision transformer inference. Our analysis shows that leveraging this type of dynamic execution can lead to saving 28\% of energy with a 1.4\% accuracy drop for SegFormer (63 GFLOPs), with no additional training, and 53\% of energy for ResNet-50 (4 GFLOPs) with a 3.3\% accuracy drop by switching between pretrained Once-For-All models.