BACKGROUND: Large language models (LLMs) have the potential to support promising new applications in health informatics. However, practical data on sample size considerations for fine-tuning LLMs to perform specific tasks in biomedical and health policy contexts are lacking.
OBJECTIVE: This study aims to evaluate sample size and sample selection techniques for fine-tuning LLMs to support improved named entity recognition (NER) for a custom data set of conflicts of interest disclosure statements.
METHODS: A random sample of 200 disclosure statements was prepared for annotation. All \"PERSON\" and \"ORG\" entities were identified by each of the 2 raters, and once appropriate agreement was established, the annotators independently annotated an additional 290 disclosure statements. From the 490 annotated documents, 2500 stratified random samples in different size ranges were drawn. The 2500 training set subsamples were used to fine-tune a selection of language models across 2 model architectures (Bidirectional Encoder Representations from Transformers [BERT] and Generative Pre-trained Transformer [GPT]) for improved NER, and multiple regression was used to assess the relationship between sample size (sentences), entity density (entities per sentence [EPS]), and trained model performance (F1-score). Additionally, single-predictor threshold regression models were used to evaluate the possibility of diminishing marginal returns from increased sample size or entity density.
RESULTS: Fine-tuned models ranged in topline NER performance from F1-score=0.79 to F1-score=0.96 across architectures. Two-predictor multiple linear regression models were statistically significant with multiple R2 ranging from 0.6057 to 0.7896 (all P<.001). EPS and the number of sentences were significant predictors of F1-scores in all cases ( P<.001), except for the GPT-2_large model, where EPS was not a significant predictor (P=.184). Model thresholds indicate points of diminishing marginal return from increased training data set sample size measured by the number of sentences, with point estimates ranging from 439 sentences for RoBERTa_large to 527 sentences for GPT-2_large. Likewise, the threshold regression models indicate a diminishing marginal return for EPS with point estimates between 1.36 and 1.38.
CONCLUSIONS: Relatively modest sample sizes can be used to fine-tune LLMs for NER tasks applied to biomedical text, and training data entity density should representatively approximate entity density in production data. Training data quality and a model architecture\'s intended use (text generation vs text processing or classification) may be as, or more, important as training data volume and model parameter size.

暂无摘要。

Equine strongylid parasites are ubiquitous around the world and are main targets of parasite control programmes. In recent years, automated fecal egg counting systems based on image analysis have become available allowing for collection and analysis of large-scale egg count data. This study aimed to evaluate equine strongylid fecal egg count (FEC) data generated with an automated system over three years in the US with specific attention to seasonal and regional trends in egg count magnitude and sampling activity. Five US regions were defined; North East, South East, North Central, South Central and West. The data set included state, region and zip code for each FEC. The number of FECs falling in each of the following categories were recorded: (1) 0 eggs per gram (EPG), (2) 1 ⩽ 200 EPG, (3) 201 ⩽ 500 EPG and (4) >500 EPG. The data included 58 329 FECs. A fixed effects model was constructed fitting the number of samples analysed per month, year and region, and a mixed effects model was constructed to fit the number of FECs falling in each of the 4 egg count categories defined above. The overall proportion of horses responsible for 80% of the total FEC output was 18.1%, and this was consistent across years, months and all regions except West, where the proportion was closer to 12%. Statistical analyses showed significant seasonal trends and regional differences of sampling frequency and FEC category. The data demonstrated that veterinarians tended to follow a biphasic pattern when monitoring strongylid FECs in horses, regardless of location.

The United States (US) Medicare claims files are valuable sources of national healthcare utilization data with over 45 million beneficiaries each year. Due to their massive sizes and costs involved in obtaining the data, a method of randomly drawing a representative sample for retrospective cohort studies with multi-year follow-up is not well-documented.
To present a method to construct longitudinal patient samples from Medicare claims files that are representative of Medicare populations each year.
Retrospective cohort and cross-sectional designs.
US Medicare beneficiaries with diabetes over a 10-year period.
Medicare Master Beneficiary Summary Files were used to identify eligible patients for each year in over a 10-year period. We targeted a sample of ~900,000 patients per year. The first year\'s sample is stratified by county and race/ethnicity (white vs. minority), and targeted at least 250 patients in each stratum with the remaining sample allocated proportional to county population size with oversampling of minorities. Patients who were alive, did not move between counties, and stayed enrolled in Medicare fee-for-service (FFS) were retained in the sample for subsequent years. Non-retained patients (those who died or were dropped from Medicare) were replaced with a sample of patients in their first year of Medicare FFS eligibility or patients who moved into a sampled county during the previous year.
The resulting sample contains an average of 899,266 ± 408 patients each year over the 10-year study period and closely matches population demographics and chronic conditions. For all years in the sample, the weighted average sample age and the population average age differ by <0.01 years; the proportion white is within 0.01%; and the proportion female is within 0.08%. Rates of 21 comorbidities estimated from the samples for all 10 years were within 0.12% of the population rates. Longitudinal cohorts based on samples also closely resembled the cohorts based on populations remaining after 5- and 10-year follow-up.
This sampling strategy can be easily adapted to other projects that require random samples of Medicare beneficiaries or other national claims files for longitudinal follow-up with possible oversampling of sub-populations.

A honeybee colony, as a super-organism, is regulated through age-polyethism. A honeybee worker\'s age is considered by means of a chronological and biological approach. The biological age is estimated with physiologically related biological markers, e.g., total hemolymph protein content (THP) and hypopharyngeal gland size (HGs), which also vary seasonally. Contemporary insights into the age-related spatial workers\' distribution within the hive nest space regarding biological age are insufficiently clarified. This study aimed to monitor changes in selected physiological markers during the entire season in relation to worker age and their spatial position in the hive nest. THP content and HG size analysis was performed in nine colonies for the entire season to compare the physiological markers within and among the groups of the workers whose ages were known and sampled in different hive parts. Seasonal impact on the biomarkers\' development was confirmed in known-age workers. In the case of HGs, this impact was the most apparent in 4- and 5-week-old workers. For THP, the seasonal impact was the most obvious in 2-week-old workers. The highest THP was found in 1- and 2-week-old workers during the entire season. Biologically younger workers of the same age were located predominantly in upper hive parts consistently throughout the year and vice versa. These workers showed significantly higher THP in comparison with those sampled below. Regarding the chronological age, the downwards, spatially shifting mechanism of workers within the hive nest while they aged was characterized. We recommend storage of diluted hemolymph samples up to one month before performing an assay if necessary. The physiological context, relation to division of labor and benefits for beekeeping practices are discussed.

BACKGROUND: Medical image segmentation is an important processing step in most of medical image analysis. Thus, high accuracy and robustness are required for them. The current deep neural network based medical segmentation methods have good effect on image with balanced foreground and background, but it will loss the characteristics of small targets on image with imbalanced foreground and background after multiple convolutions.
METHODS: In order to retain the features of small targets in the deep network, we proposed a new medical image segmentation model based on the U-Net with squeeze-and-excitation and attention modules which form a spiral closed path,callled as Spiral Squeeze-and-Excitation and Attention NET (SEA-NET) in this paper. The segmentation model used squeeze-and-extraction modules to adjust the channel information to enhance the useful information and used attention modules to adjust the spatial information of the feature map to highlight the target area for small target segmentation when up-sampling. The deep semantic information is integrated into the shallow feature map by the attention model. Therefore, the deep semantic information cannot be scattered by continuous up-sampling. We used cross entropy loss + Tversky loss function for fast convergence and well processing the imbalanced data sets. Our proposed SEA-NET was tested on the brain MRI dataset LPBA40 and peripheral blood smear images.
CONCLUSIONS: On brain MRI data, the average value of the Dice coefficient we obtained reached 98.1[Formula: see text]. On the peripheral blood smear dataset, our proposed model has a good segmentation effect on adhesion cells.
RESULTS: The experimental results proved that the proposed SEA-Net performed better than U-Net, U-Net++, etc. in medical image segmentation.

With regard to the use and transfer of research participants\' personal information, samples and other data nationally and internationally, it is necessary to construct a data management plan. One of the key objectives of a data management plan is to explain the governance of clinical, biochemical, laboratory, molecular and other sources of data according to the regulations and policies of all relevant stakeholders. It also seeks to describe the processes involved in protecting the personal information of research participants, especially those from vulnerable populations. In most data management plans, the framework therefore consists of describing the collection, organization, use, storage, contextualization, preservation, sharing and access of/to research data and/or samples. It may also include a description of data management resources, including those associated with analyzed samples, and identifies responsible parties for the establishment, implementation and overall management of the data management strategy. Importantly, the data management plan serves to highlight potential problems with the collection, sharing, and preservation of research data. However, there are different forms of data management plans and requirements may vary due to funder guidelines and the nature of the study under consideration. This paper leverages the detailed data management plans constructed for the \'NESHIE study\' and is a first attempt at providing a comprehensive template applicable to research focused on vulnerable populations, particularly those within LMICs, that includes a multi-omics approach to achieve the study aims. More particularly, this template, available for download as a supplementary document, provides a modifiable outline for future projects that involve similar sensitivities, whether in clinical research or clinical trials. It includes a description of the management not only of the data generated through standard clinical practice, but also that which is generated through the analysis of a variety of samples being collected from research participants and analyzed using multi-omics approaches.

The role of community conservation areas for large mammals is rarely evaluated. We investigated the species richness and frequency of sightings of large mammals in the Dodola Community Conservation Area. The study area was stratified into three habitat types, and 49 lines transect was laid (27 Dry evergreen Afromontane forests, 20 Sub-afro-alpine habitats, and 2 plantation forests) based on the topography, land use, and vegetation cover of the study area. A total of 24 species of large mammals were identified and recorded in the study area. Though the community conservation area is home to diverse species of mammals, including some endemic and endangered ones such as mountain nyala and Bale Monkey, human encroachment, agriculture, and overgrazing are prominent in the area, putting huge pressure on flora and fauna. Therefore, we recommend the participatory approach be strengthened to ensure sustainable coexistence between people and wildlife.

Surface-enhanced Raman spectroscopy (SERS) is a highly sensitive technique that can assist in trace analysis for biomedical, diagnostic, and environmental applications. However, a major limitation of SERS is surface contamination of the substrates used, which can complicate the spectral reproducibility, limits of detection, and detection of unknown analytes. This is especially prevalent with commercially available substrates as shipping under a controlled and clean environment is difficult. Here we report a method using dilute bleach solutions to remove surface contamination from commercially available substrates consisting of gold-coated nanopillar arrays that maintains functionality. The results show that this method can be used to remove background signals associated with typical surface contamination in commercially available substrates as well as remove thiolated self-assembled monolayers (SAMs). Results indicate the bleach oxidizes the surface contaminants, which can then be easily washed away. Although the metallic surface also becomes oxidized in this process, the surface can be reduced without loss of SERS activity. The SERS intensity of SAMs improved following bleach treatment across all concentrations studied.

暂无摘要。