Generalizable Model Research Articles

Multi-Context Point Cloud Dataset and Machine Learning for Railway Semantic Segmentation

Railway scene understanding is crucial for various applications, including autonomous trains, digital twining, and infrastructure change monitoring. However, the development of the latter is constrained by the lack of annotated datasets and limitations of existing algorithms. To address this challenge, we present Rail3D, the first comprehensive dataset for semantic segmentation in railway environments with a comparative analysis. Rail3D encompasses three distinct railway contexts from Hungary, France, and Belgium, capturing a wide range of railway assets and conditions. With over 288 million annotated points, Rail3D surpasses existing datasets in size and diversity, enabling the training of generalizable machine learning models. We conducted a generic classification with nine universal classes (Ground, Vegetation, Rail, Poles, Wires, Signals, Fence, Installation, and Building) and evaluated the performance of three state-of-the-art models: KPConv (Kernel Point Convolution), LightGBM, and Random Forest. The best performing model, a fine-tuned KPConv, achieved a mean Intersection over Union (mIoU) of 86%. While the LightGBM-based method achieved a mIoU of 71%, outperforming Random Forest. This study will benefit infrastructure experts and railway researchers by providing a comprehensive dataset and benchmarks for 3D semantic segmentation. The data and code are publicly available for France and Hungary, with continuous updates based on user feedback.

Meta-learners for few-shot weakly-supervised medical image segmentation

Most uses of Meta-Learning in visual recognition are very often applied to image classification, with a relative lack of work in other tasks such as segmentation and detection. We propose a new generic Meta-Learning framework for few-shot weakly supervised segmentation in medical imaging domains. The proposed approach includes a meta-training phase that uses a meta-dataset. It is deployed on an out-of-distribution few-shot target task, where a single highly generalizable model, trained via a selective supervised loss function, is used as a predictor. The model can be trained in several distinct ways, such as second-order optimization, metric learning, and late fusion. Some relevant improvements of existing methods that are part of the proposed approach are presented. We conduct a comparative analysis of meta-learners from distinct paradigms adapted to few-shot image segmentation in different sparsely annotated radiological tasks. The imaging modalities include 2D chest, mammographic, and dental X-rays, as well as 2D slices of volumetric tomography and resonance images. Our experiments consider in total 9 meta-learners, 4 backbones, and multiple target organ segmentation tasks. We explore small-data scenarios in radiology with varying weak annotation styles and densities. Our analysis shows that metric-based meta-learning approaches achieve better segmentation results in tasks with smaller domain shifts compared to the meta-training datasets, while some gradient- and fusion-based meta-learners are more generalizable to larger domain shifts. Guidelines learned from the comparative performance assessment of the analyzed methods are summarized to support those readers interested in the field.

Trade‐Off Between Light Deprivation and Desiccation in Intertidal Seagrasses Due To Periodic Tidal Inundation and Exposure: Insights From a Data‐Calibrated Model

AbstractSome seagrass species thrive in shallow intertidal zones globally, adapting to periodic tidal inundation and exposure with distinctive physiological traits and offering crucial ecosystem services. However, predicting the responses of intertidal seagrasses to external stressors is hampered by the complexity of the dynamic and harsh environments they occupy. Intertidal seagrass growth models, especially those incorporating dynamic physiological responses, are scarce in the literature. Our study comprehensively collated relevant data from the literature to parameterize the relationship between air exposure, seagrass leaf water content and photosynthetic efficiency to inform new growth rate functions for generalizable intertidal seagrass growth models. We tested the applicability of these model formulations for scenarios with varying physiological process assumptions, seagrass species, tidal conditions, meadow elevations and water turbidity. We found that neglecting air‐exposed physiological responses (i.e., leaf water content loss and reduced photosynthetic efficiency) can substantially overestimate seagrass growth rates. We also observed a trade‐off between light deprivation and desiccation on intertidal seagrass growth under specific tidal ranges and turbidity conditions. This can yield an “optimal” elevation where overall stress of desiccation (increasing with meadow elevation) and light deprivation (decreasing with meadow elevation) are minimized. The predicted optimal elevation, that is, the most suitable habitat for intertidal seagrass, moves upward as water turbidity increases. Our study provides conceptual and quantitative guidance for ecological modelers to include air exposure responses of intertidal seagrasses in coastal ecosystem models. The model also helps to evaluate the viability of intertidal seagrass habitats and inform decisions on coastal ecosystem management under changing environmental conditions.

Lexical-Semantic Content, Not Syntactic Structure, Is the Main Contributor to ANN-Brain Similarity of fMRI Responses in the Language Network.

Representations from artificial neural network (ANN) language models have been shown to predict human brain activity in the language network. To understand what aspects of linguistic stimuli contribute to ANN-to-brain similarity, we used an fMRI data set of responses to n = 627 naturalistic English sentences (Pereira et al., 2018) and systematically manipulated the stimuli for which ANN representations were extracted. In particular, we (i) perturbed sentences' word order, (ii) removed different subsets of words, or (iii) replaced sentences with other sentences of varying semantic similarity. We found that the lexical-semantic content of the sentence (largely carried by content words) rather than the sentence's syntactic form (conveyed via word order or function words) is primarily responsible for the ANN-to-brain similarity. In follow-up analyses, we found that perturbation manipulations that adversely affect brain predictivity also lead to more divergent representations in the ANN's embedding space and decrease the ANN's ability to predict upcoming tokens in those stimuli. Further, results are robust as to whether the mapping model is trained on intact or perturbed stimuli and whether the ANN sentence representations are conditioned on the same linguistic context that humans saw. The critical result-that lexical-semantic content is the main contributor to the similarity between ANN representations and neural ones-aligns with the idea that the goal of the human language system is to extract meaning from linguistic strings. Finally, this work highlights the strength of systematic experimental manipulations for evaluating how close we are to accurate and generalizable models of the human language network.

Effects of primary health care and socioeconomic aspects on the dispersion of COVID-19 in the Brazilian Northeast: Ecological study of the first pandemic wave.

The COVID-19 pandemic has had a negative impact on socioeconomic and public health conditions of the population. To measure the temporal evolution of COVID-19 cases in cities near the countryside outside metropolitan areas of northeastern Brazil and the impact of the primary care organization in its containment. This is a time-series study, based on the first three months of COVID-19 incidence in northeastern Brazil. Secondary data were used, the outcome was number of COVID-19 cases. Independent variables were time, coverage and quality score of basic health services, and demographic, socioeconomic and social isolation variables. Generalizable Linear Models with first order autoregression were applied. COVID-19 spreads heterogeneously in cities near the countryside of Northeastern Brazilian cities, showing associations with the city size, socioeconomic and organizational indicators of services. The Family Health Strategy seems to mitigate the speed of progression and burden of the disease, in addition to measures such as social isolation and closure of commercial activities. The spread of COVID-19 reveals multiple related factors, which require coordinated intersectoral actions in order to mitigate its problems, especially in biologically and socially vulnerable populations.

Enhancing quality control in emulsion-type sausage production: Predicting chemical composition of intact samples with near infrared spectroscopy

This research actively explores the potential of near infrared spectroscopy (NIR) for analyzing the chemical composition of emulsion-type sausages, focusing on critical factors like residual nitrite, moisture, protein, and fat content. To establish robust and generalizable models, we utilized a dataset of 100 experimentally prepared sausages encompassing a wide range of pork back fat replacement levels (5%, 15%, 30%, 45%, and 60%) and added sodium nitrite amounts (0, 80, 125, 250, and 375 ppm). An external validation set of 20 commercially sourced sausages further assessed the model’s real-world applicability. Partial least squares (PLS) regression calibration models with multiplicative scatter correction (MSC) pre-treatment demonstrated impressive accuracy for moisture (RMSECV = 0.57%, RPD = 9.8), fat (RMSECV = 1.17%, RPD = 9.5), and protein (RMSECV = 0.30%, RPD = 7.6). While residual nitrite prediction presented challenges due to its inherent complexity, the external validation yielded a competitive root mean square error of prediction (RMSEP) of 12.02 ppm, surpassing the average performance reported in similar studies (RMSEP ∼15 ppm) by 3 ppm. Importantly, sample homogenization did not significantly affect parameter prediction, highlighting the robustness of the NIR-based approach. These findings suggest that NIR spectroscopy, with its non-destructive, rapid, and cost-effective nature, could provide valuable tools for quality control and monitoring in the emulsion-type sausage industry. More importantly, improved nitrite prediction could pave the way for enhanced precision and control in sausage production, ultimately contributing to improved food safety and sustainability.

A versatile, semi-automated image analysis workflow for time-lapse camera trap image classification

Camera traps are a powerful, practical, and non-invasive method used widely to monitor animal communities and evaluate management actions. However, camera trap arrays can generate thousands to millions of images that require significant time and effort to review. Computer vision has emerged as a tool to accelerate this image review process. We propose a multi-step, semi-automated workflow which takes advantage of site-specific and generalizable models to improve detections and consists of (1) automatically identifying and removing low-quality images in parallel with classification into animals, humans, vehicles, and empty, (2) automatically cropping objects from images and classifying them (rock, bait, empty, and species), and (3) manually inspecting a subset of images. We trained and evaluated this approach using 548,627 images from 46 cameras in two regions of the Arctic: “Finnmark” (Finnmark County, Norway) and “Yamal” (Yamalo-Nenets Autonomous District, Russia). The automated steps yield image classification accuracies of 92% and 90% for the Finnmark and Yamal sets, respectively, reducing the number of images that required manual inspection to 9.2% of the Finnmark set and 3.9% of the Yamal set. The amount of time invested in developing models would be offset by the time saved from automation after 960 thousand images have been processed. Researchers can modify this multi-step process to develop their own site-specific models and meet other needs for monitoring and surveying wildlife, balancing the acceptable levels of false negatives and positives.

Likelihood-based generalization of Markov parameter estimation and multiple shooting objectives in system identification

This paper considers system identification (ID) of linear and nonlinear non-autonomous systems from noisy and sparse data. We analyze an optimization objective derived from Bayesian inference for the dynamics of hidden Markov models. We then relate this objective to that used in several state-of-the-art approaches for both linear and nonlinear system ID. In the former, we analyze least squares approaches for Markov parameter estimation, and in the latter, we analyze the multiple shooting approach. We demonstrate the limitations of the optimization problems posed by these existing methods by showing that they can be seen as special cases of the proposed optimization objective under certain simplifying assumptions: conditional independence of data and zero model error. Furthermore, we observe that the proposed approach has improved smoothness and inherent regularization that make it well-suited for system ID and provide mathematical explanations for these characteristics’ origins. Finally, numerical simulations demonstrate a mean squared error over 8.7 times lower compared to multiple shooting when data are noisy and/or sparse. Moreover, the proposed approach identifies accurate and generalizable models even when there are more parameters than data or when the system exhibits chaotic behavior.

Multi-sample ζ\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\zeta $$\\end{document}-mixup: richer, more realistic synthetic samples from a p-series interpolant

Modern deep learning training procedures rely on model regularization techniques such as data augmentation methods, which generate training samples that increase the diversity of data and richness of label information. A popular recent method, mixup, uses convex combinations of pairs of original samples to generate new samples. However, as we show in our experiments, mixup can produce undesirable synthetic samples, where the data is sampled off the manifold and can contain incorrect labels. We propose ζ\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\zeta $$\\end{document}-mixup, a generalization of mixup with provably and demonstrably desirable properties that allows convex combinations of T≥2\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$${T} \\ge 2$$\\end{document} samples, leading to more realistic and diverse outputs that incorporate information from T\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$${T}$$\\end{document} original samples by using a p-series interpolant. We show that, compared to mixup, ζ\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\zeta $$\\end{document}-mixup better preserves the intrinsic dimensionality of the original datasets, which is a desirable property for training generalizable models. Furthermore, we show that our implementation of ζ\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\zeta $$\\end{document}-mixup is faster than mixup, and extensive evaluation on controlled synthetic and 26 diverse real-world natural and medical image classification datasets shows that ζ\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\zeta $$\\end{document}-mixup outperforms mixup, CutMix, and traditional data augmentation techniques. The code will be released at https://github.com/kakumarabhishek/zeta-mixup.

Photonic data analysis in 2050

Photonic data analysis is a field at the intersection of imaging, spectroscopy, machine learning, and computer science. The diversity of both data types and application scenarios requires flexibility in the methods applied, combining a full range of computational methods, from classical chemometric techniques to state-of-the-art deep learning solutions. Interdisciplinary and international collaborations are needed to accelerate the progress of photonic data science. An underlying data infrastructure and standardization will be needed to provide collaborative platforms for research on data comparability, enabling the integration of novel photonic techniques into routine applications. The increasing complexity of the questions being investigated requires the application of more sophisticated data-driven models, which may only be optimized for large data sets. Unfortunately, novel techniques in the early stages of development can rarely provide a variability of measured samples sufficient to build a generalizable complex model. To overcome this problem, state-of-the-art methods will emerge for working with extremely limited or unbalanced data, as well as for dealing with device-to-device variations. Further developments are also foreseen in computable artificial intelligence methods, which will allow the validation of models of any architecture by comparing them with the knowledge of the researchers.

The common drivers of children and young people’s health and wellbeing across 13 local government areas: a systems view

BackgroundSystem dynamics approaches, including group model building (GMB) and causal loop diagrams (CLDs), can be used to document complex public health problems from a community perspective. This paper aims to apply Social Network Analysis (SNA) methods to combine multiple CLDs created by local communities into a summary CLD, to identify common drivers of the health and wellbeing of children and young people.MethodsThirteen community CLDs regarding children and young people health and wellbeing were merged into one diagram involving three steps: (1) combining variable names; (2) CLD merging, where multiple CLDs were combined into one CLD with a set of unique variables and connections; (3) paring, where the Decision-Making Trial and Evaluation Laboratory (DEMATEL) method was used to generate a cut-point to reduce the number of variables and connections and to rank the overall importance of each variable in the merged CLD.ResultsCombining variable names resulted in 290 variables across the 13 CLDS. A total of 1,042 causal links were identified in the merged CLD. The DEMATEL analysis of the merged CLD identified 23 common variables with a net importance between 1.0 and 4.5 R + C values and 57 causal links. The variables with the highest net importance were ‘mental health’ and ‘social connection & support’ classified as high net receivers of influence within the system.ConclusionsCombining large CLDs into a simple diagram represents a generalisable model of the drivers of complex health problems.

Aligning open educational resources to new taxonomies: How AI technologies can help and in which scenarios

Aligning open educational resources (OER) to skill taxonomies is a common task in the education field and helps teachers better locate material that aligns with the standards of their curriculum. When taxonomies change, as they periodically do, re-tagging the increasing mass of open educational resources is needed. The process of manual tagging is, however, exceedingly labor intensive. We propose and evaluate a novel combination of machine learning methods to help automate tagging open educational resources with skills from an existing taxonomy as well as skills from any newly introduced taxonomy. We collected text, image figures, and videos from tens of thousands of educational resources from two major digital learning platforms to answer the research questions of: how effective are machine learning models in automatically updating OER classification to reflect a new taxonomy (RQ1), and which models may be of practical use in different scenarios (RQ2)? Using several taxonomies, including the US Common Core, we find that while full automation is not practically viable, our most generalizable model can reach non-expert human labeling performance requiring only 100 labeled examples and near expert level with 5000. We believe these novel findings may have immediate utility for practitioners and policymakers and better ready the growing landscape of open educational resources for the advent of new taxonomies ahead. We publicly release our pre-trained US Common Core and new taxonomy tagging models, providing guidance on their viability in various real-world scenarios.

Vehicular Fuel Consumption and CO2 Emission Estimation Model Integrating Novel Driving Behavior Data Using Machine Learning

Fossil fuel vehicles significantly contribute to CO2 emissions due to their high consumption of fossil fuels. Accurate estimation of vehicular fuel consumption and the associated CO2 emissions is crucial for mitigating these emissions. Although driving behavior is a vital factor influencing fuel consumption and CO2 emissions, it remains largely unaddressed in current CO2 emission estimation models. This study incorporates novel driving behavior data, specifically counts of occurrences of dangerous driving behaviors, including speeding, sudden accelerating, and sudden braking, as well as driving time and driving distances on expressways, national highways, and local roads, respectively, into monthly fuel consumption estimation models for individual gasoline and hybrid vehicles. The CO2 emissions are then calculated as a secondary parameter based on the estimated fuel consumption, assuming a linear relationship between the two. Using regression algorithms, it has been demonstrated that all the proposed driving behavior data are relevant for monthly CO2 emission estimation. By integrating the driving behavior data of various vehicle categories, a generalizable CO2 estimation model is proposed. When utilizing all the proposed driving behavior data collectively, our random forest regression model achieves the highest prediction accuracy, with R2, RMSE, and MAE values of 0.975, 13.293 kg, and 8.329 kg, respectively, for monthly CO2 emission estimation of individual vehicles. This research offers insights into CO2 emission reduction and energy conservation in the road transportation sector.

Gendered sustainability: Are public spaces designed for girls good for everyone?: Examining female participation as a strategy for inclusive public space

This article examines female participation as a strategy for sustainable development. The increasing interest in female participation as a sustainability strategy on an international scale, underscores the need for critical examinations of this approach. Centring on contemporary Sweden as a prominent site of this development, this article investigates the municipal use of gendered design processes and gender-sensitive policies. Drawing on the study of five public spaces, this article discusses categorization and processes of inclusion/exclusion and ordination/subordination in relation to the design and use of public spaces. Central to the analysis is a recurring ambivalence identified in the material, where spaces are referred to both as spaces for girls and spaces for everyone. The authors argue that spaces can only fulfil the double demand of being for girls and for everyone if we accept that girls stay subordinated to boys in their use of public space. Drawing on an analysis of the interplay of inclusion/exclusion and ordination/subordination they propose a generalizable model of participation and public space. The article concludes by advocating for the creation of spaces for girls and the creation of spaces for everyone to be treated as two related yet separate tasks.

Integrating convolutional neural network and constitutive model for rapid prediction of stress-strain curves in fibre reinforced polymers: A generalisable approach

Despite recent advancements in using machine learning (ML) techniques to establish the microstructure-property linkage for composites’ representative volume elements (RVEs), challenges persist in effectively characterising the effect of microstructural randomness on material properties. This complexity arises from the difficulty of expressing randomness as definitive variables and its intertwined relations with other factors, such as material constituents. Such complexities result in limitations in generalising ML models across different material constituents. Conventional solutions to these challenges usually necessitate large datasets, which require considerable computational resources, for an accurate and generalisable ML models to be trained. This paper presents an innovative approach to tackling these challenges by integrating a high-accuracy convolutional neural network (CNN) with a novel microstructure-factored constitutive model (MCM). The MCM, rooted from classic empirical constitutive modelling, effectively segregates the microstructural and constituting material effects, extending the generalisability and thus significantly enhancing the efficacy of the CNN. This new approach enabled a CNN trained on the transverse stress-strain curves of one set of material constituents (CF/PEEK at 270 °C) to be generalised for the rapid prediction of various sets of material constituents at different temperatures, unseen by the CNN during training, with an average mean absolute percentage error around 3 %.

MOB-CBAM: A dual-channel attention-based deep learning generalizable model for breast cancer molecular subtypes prediction using mammograms

Background and objectiveDeep Learning models have emerged as a significant tool in generating efficient solutions for complex problems including cancer detection, as they can analyze large amounts of data with high efficiency and performance. Recent medical studies highlight the significance of molecular subtype detection in breast cancer, aiding the development of personalized treatment plans as different subtypes of cancer respond better to different therapies. MethodsIn this work, we propose a novel lightweight dual-channel attention-based deep learning model MOB-CBAM that utilizes the backbone of MobileNet-V3 architecture with a Convolutional Block Attention Module to make highly accurate and precise predictions about breast cancer. We used the CMMD mammogram dataset to evaluate the proposed model in our study. Nine distinct data subsets were created from the original dataset to perform coarse and fine-grained predictions, enabling it to identify masses, calcifications, benign, malignant tumors and molecular subtypes of cancer, including Luminal A, Luminal B, HER-2 Positive, and Triple Negative. The pipeline incorporates several image pre-processing techniques, including filtering, enhancement, and normalization, for enhancing the model's generalization ability. ResultsWhile identifying benign versus malignant tumors, i.e., coarse-grained classification, the MOB-CBAM model produced exceptional results with 99 % accuracy, precision, recall, and F1-score values of 0.99 and MCC of 0.98. In terms of fine-grained classification, the MOB-CBAM model has proven to be highly efficient in accurately identifying mass with (benign/malignant) and calcification with (benign/malignant) classification tasks with an impressive accuracy rate of 98 %. We have also cross-validated the efficiency of the proposed MOB-CBAM deep learning architecture on two datasets: MIAS and CBIS-DDSM. On the MIAS dataset, an accuracy of 97 % was reported for the task of classifying benign, malignant, and normal images, while on the CBIS-DDSM dataset, an accuracy of 98 % was achieved for the classification of mass with either benign or malignant, and calcification with benign and malignant tumors. ConclusionThis study presents lightweight MOB-CBAM, a novel deep learning framework, to address breast cancer diagnosis and subtype prediction. The model's innovative incorporation of the CBAM enhances precise predictions. The extensive evaluation of the CMMD dataset and cross-validation on other datasets affirm the model's efficacy.

Weighing in on the average weights: Measuring corporate social performance (CSP) score using DEA

This study develops a new DEA approach to incorporate such market-level information (i.e., average weights over firms) to create a composite corporate social performance (CSP) measure. The new model addresses the challenge that stakeholder expectations—expressed as the CSP weights of different CSP dimensions—are usually unknown at the firm level, but some reliable estimates of the average weights (across firms) are available at the industry or market level. Distinct from the traditional DEA, the new approach can incorporate constraints on the average values of weight distributions and is built on a computational framework new to the literature. The new model is compared with the traditional CSP approach, using longitudinal investment data by socially responsible funds (SRF). The regression results show evidence that the traditional CSP score, which is widely used in the management literature, may not be a reliable CSP score: firms with higher CSP scores tend to have lower SRF investment (in both the number of SRF and total shares). In contrast, the DEA-based CSP scores show the correct and positive correlation with SRF investment. Overall, this paper contributes to the literature by presenting a more reliable CSP approach and a new, generalizable DEA model. This is also the first empirical work documenting the potential validity issue of the traditional CSP scores.

A Generalizable Decision-Making Framework for Selecting Onsite versus Send-out Clinical Laboratory Testing.

Laboratory networks provide services through onsite testing or through specimen transport to higher-tier laboratories. This decision is based on the interplay of testing characteristics, treatment characteristics, and epidemiological characteristics. Our objective was to develop a generalizable model using the threshold approach to medical decision making to inform test placement decisions. We developed a decision model to compare the incremental utility of onsite versus send-out testing for clinical purposes. We then performed Monte Carlo simulations to identify the settings under which each strategy would be preferred. Tuberculosis was modeled as an exemplar. The most important determinants of the decision to test onsite versus send-out were the clinical utility lost due to send-out testing delays and the accuracy decrement with onsite testing. When the sensitivity decrements of onsite testing were minimal, onsite testing tended to be preferred when send-out delays reduced clinical utility by >20%. By contrast, when onsite testing incurred large reductions in sensitivity, onsite testing tended to be preferred when utility lost due to delays was >50%. The relative cost of onsite versus send-out testing affected these thresholds, particularly when testing costs were >10% of treatment costs. Decision makers can select onsite versus send-out testing in an evidence-based fashion using estimates of the percentage of clinical utility lost due to send-out delays and the relative accuracy of onsite versus send-out testing. This model is designed to be generalizable to a wide variety of use cases. The design of laboratory networks, including the decision to place diagnostic instruments at the point-of-care or at higher tiers as accessed through specimen transport, can be informed using the threshold approach to medical decision making.The most important determinants of the decision to test onsite versus send-out were the clinical utility lost due to send-out testing delays and the accuracy decrement with onsite testing.The threshold approach to medical decision making can be used to compare point-of-care testing accuracy decrements with the lost utility of treatment due to send-out testing delays.The relative cost of onsite versus send-out testing affected these thresholds, particularly when testing costs were >10% of treatment costs.

A pretrain-finetune approach for improving model generalizability in outcome prediction of acute respiratory distress syndrome patients

BackgroundEarly prediction of acute respiratory distress syndrome (ARDS) of critically ill patients in intensive care units (ICUs) has been intensively studied in the past years. Yet a prediction model trained on data from one hospital might not be well generalized to other hospitals. It is therefore essential to develop an accurate and generalizable ARDS prediction model adaptive to different hospital or medical centers. MethodsWe analyzed electronic medical records of 200,859 and 50,920 hospitalized patients within 24 h after being diagnosed with ARDS from the Philips eICU Institute (eICU-CRD) and the Medical Information Mart for Intensive Care (MIMIC-IV) dataset, respectively. Patients were sorted into three groups, including rapid death, long stay, and recovery, based on their condition or outcome between 24 and 72 h after ARDS diagnosis. To improve prediction performance and generalizability, a “pretrain-finetune” approach was applied, where we pretrained models on the eICU-CRD dataset and performed model finetuning using only a part (35%) of the MIMIC-IV dataset, and then tested the finetuned models on the remaining data from the MIMIC-IV dataset. Well-known machine-learning algorithms, including logistic regression, random forest, extreme gradient boosting, and multilayer perceptron neural networks, were employed to predict ARDS outcomes. Prediction performance was evaluated using the area under the receiver-operating characteristic curve (AUC). ResultsResults show that, in general, multilayer perceptron neural networks outperformed the other models. The use of pretrain-finetune yielded improved performance in predicting ARDS outcomes achieving a micro-AUC of 0.870 for the MIMIC-IV dataset, an improvement of 0.046 over the pretrain model.Conclusions.The proposed pretrain-finetune approach can effectively improve model generalizability from one to another dataset in ARDS prediction.

Optimising brain age estimation through transfer learning: A suite of pre-trained foundation models for improved performance and generalisability in a clinical setting.

Estimated age from brain MRI data has emerged as a promising biomarker of neurological health. However, the absence of large, diverse, and clinically representative training datasets, along with the complexity of managing heterogeneous MRI data, presents significant barriers to the development of accurate and generalisable models appropriate for clinical use. Here, we present a deep learning framework trained on routine clinical data (N up to 18,890, age range 18-96 years). We trained five separate models for accurate brain age prediction (all with mean absolute error ≤4.0 years, R2 ≥ .86) across five different MRI sequences (T2 -weighted, T2 -FLAIR, T1 -weighted, diffusion-weighted, and gradient-recalled echo T2 *-weighted). Our trained models offer dual functionality. First, they have the potential to be directly employed on clinical data. Second, they can be used as foundation models for further refinement to accommodate a range of other MRI sequences (and therefore a range of clinical scenarios which employ such sequences). This adaptation process, enabled by transfer learning, proved effective in our study across a range of MRI sequences and scan orientations, including those which differed considerably from the original training datasets. Crucially, our findings suggest that this approach remains viable even with limited data availability (as low as N = 25 for fine-tuning), thus broadening the application of brain age estimation to more diverse clinical contexts and patient populations. By making these models publicly available, we aim to provide the scientific community with a versatile toolkit, promoting further research in brain age prediction and related areas.