A vast amount of clinical data are still stored in unstructured text. Automatic extraction of medical information from these data poses several challenges: high costs of clinical expertise, restricted computational resources, strict privacy regulations, and limited interpretability of model predictions. Recent domain adaptation and prompting methods using lightweight masked language models showed promising results with minimal training data and allow for application of well-established interpretability methods. We are first to present a systematic evaluation of advanced domain-adaptation and prompting methods in a lower-resource medical domain task, performing multi-class section classification on German doctor’s letters. We evaluate a variety of models, model sizes (further-pre)training and task settings, and conduct extensive class-wise evaluations supported by Shapley values to validate the quality of small-scale training data and to ensure interpretability of model predictions. We show that in few-shot learning scenarios, a lightweight, domain-adapted pretrained language model, prompted with just 20 shots per section class, outperforms a traditional classification model, by increasing accuracy from $48.6\%$ to $79.1\%$. By using Shapley values for model selection and training data optimization, we could further increase accuracy up to $84.3\%$. Our analyses reveal that pretraining of masked language models on general-language data is important to support successful domain-transfer to medical language, so that further-pretraining of general-language models on domain-specific documents can outperform models pretrained on domain-specific data only. Our evaluations show that applying prompting based on general-language pretrained masked language models combined with further-pretraining on medical-domain data achieves significant improvements in accuracy beyond traditional models with minimal training data. Further performance improvements and interpretability of results can be achieved, using interpretability methods such as Shapley values. Our findings highlight the feasibility of deploying powerful machine learning methods in clinical settings and can serve as a process-oriented guideline for lower-resource languages and domains such as clinical information extraction projects.

Due to adverse snow and cloud conditions, only a few inventories are available for the maritime glaciers in New Zealand. These are difficult to compare as different approaches and baseline data have been used to create them. In consequence, glacier fluctuations in New Zealand over the past two decades are only known for a few glaciers based on field observations. Here we present the results of a new inventory for the ‘year 2000’ (some scenes are from 2001 and 2002) that is based on glacier outlines from a recently published inventory for the year 2016 and allowed consistent change assessment for nearly 3000 glaciers over this period. The year 2000 inventory was created by manual on-screen digitizing using Landsat ETM+ satellite imagery (15 m panchromatic band) in the background and the year 2016 outlines as a starting point. Major challenges faced were late and early seasonal snow, clouds and shadow, the geo-location mismatch between Landsat and Sentinel-2 as well as the correct interpretation of ice patches and ice under debris cover. In total, we re-mapped 2967 glaciers covering an area of 885.5 km2 in 2000, which is 91.7 km2 (or 10.4%) more than the 793.8 km2 mapped in 2016. Area change rates (mean rate −0.65% a−1) increase towards smaller glaciers. Strongest area loss from 2000 to 2016 occurred at elevations ~1900 m but the highest relative loss was found below 800 m a.s.l. In total, 109 glaciers split into two or more entities and 264 had wasted away by 2016.

When people judge risk or the probability of a risky prospect, single case narratives can bias judgments when a statistical base-rate is also provided. In this work we investigate various methodological and procedural factors that may influence this narrative bias. We found that narratives had the strongest effect on a non-numerical risk measure, which was also the best predictor of behavioral intentions. In contrast, two scales for subjective probability reflected primarily statistical variations. We observed a negativity bias on the risk measure, such that the narratives increased rather than decreased risk perceptions, whereas the effect on probability judgments was symmetric. Additionally, we found no evidence that the narrative bias is solely produced by adherence to conversational norms. Finally, changing the absolute number of narratives reporting the focal event, while keeping their relative frequency constant, had no effect. Thus, individuals extract a representation of likelihood from a sample of single-case narratives, which drives the bias. These results show that the narrative bias is in part dependent on the measure used to assess it and underline the conceptual distinction between subjective probability and perceived risk.

Glacier monitoring has been internationally coordinated for more than 125 years. Despite this long history, there is no authoritative answer to the popular question: ‘Which glaciers are the largest in the world?’ Here, we present the first systematic assessment of this question and identify the largest glaciers in the world – distinct from the two ice sheets in Greenland and Antarctica but including the glaciers on the Antarctic Peninsula. We identify the largest glaciers in two domains: on each of the seven geographical continents and in the 19 first-order glacier regions defined by the Global Terrestrial Network for Glaciers. Ranking glaciers by area is non-trivial. It depends on how a glacier is defined and mapped and also requires differentiating between a glacier and a glacier complex, i.e. glaciers that meet at ice divides such as ice caps and icefields. It also depends on the availability of a homogenized global glacier inventory. Using separate rankings for glaciers and glacier complexes, we find that the largest glacier complexes have areas on the order of tens of thousands of square kilometers whereas the largest glaciers are several thousands of square kilometers. The world's largest glaciers and glacier complexes are located in the Antarctic, Arctic and Patagonia.

The future contributions of the Antarctic Ice Sheet to sea level rise will depend on the evolution of its surface mass balance (SMB), which could amplify/dampen mass losses increasingly observed at the ice sheet's edge. In situ constraints of SMB over annual-to-decadal timescales consist mostly of firn/ice cores that have a surface footprint $\sim$cm$^{2}$. SMB constraints also come from climate models, which have a higher temporal resolution but a larger surface footprint of several km$^{2}$. We use ice-penetrating radar data to obtain an intermediate spatial and temporal resolution SMB record over three ice rises along the Princess Ragnhild Coast. The co-located ice cores allow us to obtain absolute radar-derived SMB rates at a multi-annual-to-decadal temporal resolution. By comparing the ice core SMB measurements and the radar-derived SMB records, we determine that pointwise measurements of SMB are representative of a small surface area, $\sim 200-500$ m radius extending from the ice core drill site for the ice rises studied here, and that the pointwise measurements are systematically 7–15 cm w.e. a$^{-1}$ lower than the mean SMB value calculated for the whole ice rises. However, ice core records are representative of an entire ice rise's temporal variability at the temporal resolution examined.

Glaciers in the Alps and several other regions in the world have experienced strong negative mass balances over the past few decades. Some of them are disappearing, undergoing exceptionally negative mass balances that impact the mean regional value, and require replacement. In this study, we analyse the geomorphometric characteristics of 46 mass-balance glaciers in the Alps and the long-term mass-balance time series for a subset of nine reference glaciers. We identify regime shifts in the mass-balance time series (when non-climatic controls started impacting) and develop a glacier vulnerability index (GVI) as a proxy for their possible future development, based on criteria such as hypsometric index, breaks in slope, thickness distribution and elevation change pattern. We found that the subset of 46 mass-balance glaciers reflects the characteristics of the total glacier sample very well and identified a region-specific variability of the mass balance. As the GVI is strongly related to cumulative glacier mass balances, it can be used as a pre-selector of future mass-balance glaciers. We conclude that measurements on rapidly shrinking glaciers should be continued as long as possible to identify regime shifts in hind-cast and better understand the impacts of climatic variability on such glaciers.

Increased fruit and vegetable (FV) intake is associated with reduced blood pressure (BP). However, it is not clear whether the effect of FV on BP depends on the type of FV consumed. Furthermore, there is limited research regarding the comparative effect of juices or whole FV on BP. Baseline data from a prospective cohort study of 10 660 men aged 50–59 years examined not only the cross-sectional association between total FV intake but also specific types of FV and BP in France and Northern Ireland. BP was measured, and dietary intake assessed using FFQ. After adjusting for confounders, both systolic BP (SBP) and diastolic BP (DBP) were significantly inversely associated with total fruit, vegetable and fruit juice intake; however, when examined according to fruit or vegetable sub-type (citrus fruit, other fruit, fruit juices, cooked vegetables and raw vegetables), only the other fruit and raw vegetable categories were consistently associated with reduced SBP and DBP. In relation to the risk of hypertension based on SBP >140 mmHg, the OR for total fruit, vegetable and fruit juice intake (per fourth) was 0·95 (95 % CI 0·91, 1·00), with the same estimates being 0·98 (95 % CI 0·94, 1·02) for citrus fruit (per fourth), 1·02 (95 % CI 0·98, 1·06) for fruit juice (per fourth), 0·93 (95 % CI 0·89, 0·98) for other fruit (per fourth), 1·05 (95 % CI 0·99, 1·10) for cooked vegetable (per fourth) and 0·86 (95 % CI 0·80, 0·91) for raw vegetable intakes (per fourth). Similar results were obtained for DBP. In conclusion, a high overall intake of fruit, vegetables and fruit juice was inversely associated with SBP, DBP and risk of hypertension, but this differed by FV sub-type, suggesting that the strength of the association between FV sub-types and BP might be related to the type consumed, or to processing or cooking-related factors.

Recent years has seen an increasing interest in the hallucinatory experience, including investigations of its phenomenological prevalence and character both in pathological and normal (predisposed) populations. We investigated the multi-dimensionality of hallucinatory experiences in 265 subjects from the normal population, who completed a modified version of the Launay-Slade Hallucinations Scale. Principal components analysis was performed on the data. Four factors were obtained loading on items reflecting (1) sleep-related hallucinatory experiences (2) vivid daydreams (3) intrusive thoughts or realness of thought and (4) auditory hallucinations. The results offer further evidence of the multi-dimensionality of hallucinatory disposition in the normal population. Directions for future research in hallucinatory predisposition are discussed.