PLOS digital health最新文献

Timely diagnosis of non-ST-elevation myocardial infarction (NSTEMI) remains challenging, as current protocols rely on serial high-sensitivity cardiac troponin (hs-cTn) tests that may delay decisions and overcrowd emergency departments. We retrospectively analyzed 54,636 patients receiving hs-cTn testing at emergency departments across Taiwan (May 2016-Dec 2021). Excluding STEMI and incomplete cases, we developed a machine learning (ML) model using demographics and 23 routine lab tests from the initial blood draw to enable early NSTEMI risk stratification. An actionable clinical decision supporting algorithm was also created based on ML-derived risk scores. A total of 15,096 eligible patients (mean age 69.94 ± 15.66 years; 42.2% female) were included in model training and evaluation. The ML model outperformed hs-cTn alone in both internal and external validation sets in terms of area under the receiver-operating characteristic curve. Beyond model development, a clinically actionable decision algorithm using risk score was established. Thresholds (<1.8 and ≥38.5) to define low- and high-risk groups, the model achieved a negative predictive value (NPV) of 98.8% (98.5-99.1%) for rule-out and a positive predictive value (PPV) of 78.1% (73.2-82.4%) for rule-in, encompassing 48.3% and 2.6% of patients, respectively. When combined with the established 0 h/1 h algorithm, the ML model further enhanced early decision-making, safely ruling in/out 85.3% of patients within 1 hour, with PPV and NPV reaching 84.9% (79.5-87.7%) and 100% (99.6-100%), respectively. In conclusion, this ML-based approach offers not only accurate prediction but also an actionable guide to support rapid, safe NSTEMI triage in emergency care.

Understanding the time of the menstrual cycle would help women to avoid getting pregnant without the need for surgical, hormonal, or mechanical contraception. Women who do not use contraception and do not know when they are fertile are at a higher risk (17%) of unplanned pregnancy and abortion. Classifying knowledge of fertility periods using machine learning algorithms would help to automate decision-making, produce more precise and accurate classification, and scale up to manage big and complex datasets. Therefore, this study aimed to classify knowledge of the fertility period among adolescent girls in East Africa from 2012 to 2022 using a machine-learning algorithm. A community-based cross-sectional study design was used from 12 East African countries' DHS datasets spanning 2012-2022. The machine learning algorithms were applied to classify knowledge of the fertility period and identify its predictors using R software and Python, particularly Jupiter Notebook in Anaconda. Data cleaning, one-hot encoding, data splitting, data balancing, and ten-fold cross-validation were performed. Ten machine learning algorithms and SHAP were used to select and interpret the best model. From the 40,664 adolescent girls in East Africa, 13.22% (95% CI: 12.91, 13.54) of participants had knowledge of the fertility period. Logistic regression was found to be the best model for unbalanced training data with 74.38% of an AUC and 82.71% of an accuracy. While random forest outperformed on balanced training data, it achieved 91.12% of an AUC and 83.26% accuracy. The key determinant factors of the knowledge of the fertility period were education level, country, hearing about family planning, hearing about sexually transmitted infections, wealth index, knowledge of any method, and visiting health facilities. Governments, NGOs, policy makers, and researchers can utilize these findings to design targeted interventions for improving adolescents' reproductive health based on the identified gaps and disparities.

This multicenter study aims to enhance the preoperative prediction of pathological invasiveness in clinical stage I lung adenocarcinoma (LUAD) by developing and validating topologically distinct 2D and 3D intratumoral heterogeneity (ITH) scores derived from chest CT imaging. Patients with histopathologically confirmed LUAD were enrolled from three medical centers. We established a dual-scale computational framework to quantify ITH: the 2D ITH score was derived by integrating local radiomics features with global pixel distribution patterns on the largest cross-sectional slice, while the 3D ITH score captured volumetric heterogeneity using a voxel-based topology-aware approach. Subsequently, six machine learning models integrating clinicoradiologic (CR) features with these heterogeneity scores were developed. Model performance was optimized based on the area under the curve (AUC) across a training set and validated in both an internal test set and an independent external validation set. A total of 1,238 eligible patients were enrolled. Centers 1 and 2 provided 1,053 patients (Training: n=737; Internal Test: n=316), while Center 3 provided 185 patients for external validation. The CatBoost classifier integrating 2D/3D ITH scores with CR features (2DITH-3DITH-CR CatBoost) exhibited superior diagnostic performance, achieving AUCs of 0.867 in the internal test set and 0.881 in the external validation set. The integration of topologically distinct 3D ITH scores significantly improves the preoperative stratification of LUAD invasiveness. The 2DITH-3DITH-CR CatBoost model serves as a robust, non-invasive tool to guide individualized surgical decision-making in clinical practice.

Decentralised clinical trials (DCTs) are a potentially efficient and cost-effective way of delivering research trials. My Kidneys & Me, a self-management digital health intervention for chronic kidney disease, was evaluated in a multi-centre randomised DCT (SMILE-K) (ISRCTN18314195). This study aims to evaluate recruitment outcomes and research staff experiences of delivering the SMIKE-K DCT, to inform the design of future DCTs. SMILE-K used fully remote trial processes, including online outcome measure collection. Recruitment and retention data were collected, including numbers invited, recruited, and completing outcome measures, and methods of invitation and consent. Quantitative data were analysed descriptively. Following trial recruitment, semi-structured interviews were conducted with research staff at external recruiting sites to explore their perspectives and experiences of remote trial processes. Qualitative data were analysed using thematic analysis. 420 participants were recruited to SMILE-K. The median time from expression of interest to consent was 1 day (range:0-100), and from consent to randomisation was 6 days (range:0-197). Thirteen research staff were interviewed. Six themes were identified: 'discordance between perceptions and experiences of recruiting participants', 'reallocation of available resources across research studies', 'more environmentally friendly', 'onus on participants', 'engaging disadvantaged groups of participants', and 'future considerations to improve recruitment'. Results suggest that a DCT design can reach a high number of eligible participants. An invitation flyer via post after a remote clinical appointment was the most successful method of recruitment. Research staff felt DCTs provided opportunities for a diverse and representative population to participate and study procedures were environmentally friendly; however, consideration must be given to the factors that may affect recruitment and participation. Our research highlights a clear disparity between the expected recruitment rate and the reality of recruiting for DCTs, with research staff indicating they faced unanticipated challenges during the process. We outline factors for consideration when designing and delivering DCTs.

This study aimed to develop and evaluate a machine learning pipeline using multiphase contrast-enhanced CT images and clinical data to classify renal tumors as benign, malignant-indolent, or malignant-aggressive, while assessing the contribution of each data source to the classification. In this retrospective study, 448 patients (mean age: 60.7 ± 12.6 years, 306 male, 142 female) who underwent nephrectomy and preoperative CECT between June 2008 and July 2018 were included. Tumors were histologically categorized as benign-indolent, malignant-indolent, or malignant-aggressive. Self-supervised feature extraction converted 4-phase CECT images into 512 real-valued features, combined with clinical data and tumor size for classification. Two machine learning classifiers, random forest (RF) and multi-layer perceptron (MLP), were used to predict tumor type. Nested five-fold cross-validation was employed for hyperparameter tuning and model evaluation, and performance was assessed using area under the curve (AUC) analysis. The best-performing models achieved an AUC of 0.90 (95% CI: 0.88-0.93) for classifying indolent versus aggressive tumors and 0.76 (95% CI: 0.71-0.81) for malignant versus benign tumors. Models incorporating tumor size significantly improved classification accuracy. RF classifiers excelled in distinguishing indolent from aggressive tumors, while MLP classifiers performed better for malignant versus benign classification. The machine learning pipeline demonstrated high accuracy in differentiating aggressive from indolent renal tumors, offering valuable prognostic insights for personalized treatment. Tumor size was a critical factor, complementing CECT images and clinical data. These findings highlight the potential of ML techniques in enhancing renal tumor risk stratification.

Female genital schistosomiasis (FGS) is a sequela of infection with a waterborne parasite prevalent in sub-Saharan Africa and is associated with increased HIV risk. Diagnosis of FGS involves visual colposcopic identification of lesions on the cervix or vaginal walls. Previous studies have utilized digital image processing methods with statistical validation, and more recently, an artificial intelligence (AI)-based approach has also been explored. In this work, we sought to evaluate the performance of an AI model for identifying the presence of FGS from cervical photographs. Colposcopy images were obtained from 340 subjects in Zambia. Ground truth for presence or absence of FGS was determined by trained expert human examiners using visual assessment of images. Examiners also provided a FGS severity score between 0-8 for each image based on the number of lesions and the cervical quadrants affected, where 8 denotes highest severity and 0 denotes no FGS. The images were pre-processed with specular reflection artifact removal and image cropping to focus on the regions corresponding to the cervix and the transformation zone. The preprocessed dataset was randomly divided into training (FGS = 71, no FGS = 71) and testing (FGS = 21, no FGS = 177) cohorts. Image representations in the latent space were obtained using an ensemble of pre-trained machine learning models to further classify the image into FGS and no FGS. The best performance in the testing dataset was obtained at subject-level with area under the curve (AUC) =0.70 (95% Confidence interval: 0.58 - 0.82), Specificity = 0.68, and Sensitivity = 0.71, against the ground truth. Subjects with higher FGS severity scores (between 5-8) had high prediction rate by the machine classifier compared to those with lower severity scores (between 1-4). Machine learning shows promise in detecting FGS from limited colposcopy images. Early, accurate diagnosis may enhance reproductive health, and reduce HIV transmission risks, safeguarding maternal and child health.

Wearable activity trackers have been recognised as effective tools for physical activity promotion, leading to their integration in healthcare services. Although, some qualitative literature indicated that device users may experience internal conflict. The current study is the first of our knowledge to directly examine the conflict faced by wearable activity tracker users. A qualitative, exploratory design was followed, with inductive thematic analysis conducted on semi-structured interview transcripts. The current study consisted of 11 regular wearable activity tracker users (8 female), aged between 18-59 years (M = 30.73). Four themes and nine sub-themes captured participants' internal conflict. Themes were; Who knows best? Who's in charge? Who am I without it? And What is happening to me?. Themes highlighted that device users faced conflict around navigating a data mismatch, how a wearable activity tracker impacted their behaviour, the amount of control a tracker had over them, whether their device use was positive, and how they would act and feel if they no longer used their wearable activity tracker. Participants experienced substantial internal conflict from wearable activity tracker use. The intensity of device-user relationship was clear, suggesting device dependency and perceived device importance. Findings hold crucial implications around the integration of activity trackers in healthcare services, recommendations around healthy use, and the potential long-term negative impact of using these devices on bodily intuition. Theoretical underpinnings remain unclear around wearable activity tracker use; results suggested blurred boundaries between intrinsic and extrinsic motivation - likely due to device embodiment - and highlighted the role of pressure in driving increased physical activity.

Veterinary medical records represent a large data resource for application to veterinary and One Health clinical research efforts. Use of the data is limited by interoperability challenges including inconsistent data formats and data siloing. Clinical coding using standardized medical terminologies enhances the quality of medical records and facilitates their interoperability with veterinary and human health records from other sites. Previous studies, such as DeepTag and VetTag, evaluated the application of Natural Language Processing (NLP) to automate veterinary diagnosis coding, employing long short-term memory (LSTM) and transformer models to infer a subset of Systemized Nomenclature of Medicine - Clinical Terms (SNOMED-CT) diagnosis codes from free-text clinical notes. This study expands on these efforts by incorporating all 7,739 distinct SNOMED-CT diagnosis codes recognized by the Colorado State University (CSU) Veterinary Teaching Hospital (VTH) and by leveraging the increasing availability of pre-trained language models (LMs). 13 freely available pre-trained LMs (GatorTron, MedicalAI ClinicalBERT, medAlpaca, VetBERT, PetBERT, BERT, BERT Large, RoBERTa, GPT-2, GPT-2 XL, DeBERTa V3, ModernBERT, and Clinical ModernBERT) were fine-tuned on the free-text notes from 246,473 manually-coded veterinary patient visits included in the CSU VTH's electronic health records (EHRs), which resulted in superior performance relative to previous efforts. The most accurate results were obtained when expansive labeled data were used to fine-tune relatively large clinical LMs, but the study also showed that comparable results can be obtained using more limited resources and non-clinical LMs. The results of this study contribute to the improvement of the quality of veterinary EHRs by investigating accessible methods for automated coding and support both animal and human health research by paving the way for more integrated and comprehensive health databases that span species and institutions.

While Digital Therapeutics (DTx) are widely considered a key strategy to reach certain populations with unmet healthcare needs, a range of differences in the impact and adoption of DTx still exists. These differences are not just rooted in access, but also in gaps in knowledge about how to produce community-relevant DTx, primarily stemming from the implicit or explicit exclusion of those with both relevant trained expertise (gained through formal education or professional experience) and lived expertise (gained through personal and direct experience). This paper expands the traditional conceptualization of the digital divide beyond access to encompass four interconnected domains: the Digital Knowledge Divide, Digital Evidence Generation Divide, Digital Production Divide, and Digital Adoption Divide. Drawing on Ridgeway's cultural schema theory of status, we demonstrate how conventional team hierarchies in DTx development systematically allocate status and decision-making authority through automatic cultural defaults, credentials, professional roles, demographic characteristics, rather than through contextual assessment of who possesses the most relevant expertise for specific decisions. To address this challenge, we propose a theoretical framework for dynamic expertise integration that deliberately disrupts rapid-stabilizing hierarchies by creating explicit relational spaces where teams can recognize and value both lived and trained expertise contextually. We operationalize this framework through the DTx Team Building Worksheet, a practical tool that integrates team science approaches with Community-Led Transformation principles and Culturally and Community Responsive Design. The Worksheet provides structured processes for assessing diverse forms of expertise, defining roles dynamically, and identifying decision-making priorities that shift appropriately across the DTx lifecycle. This integrated approach including problem analysis, theoretical framework, and practical tool, offers a pathway toward more equitable DTx development by enabling teams to make status dynamics explicit, expand what counts as expertise, and establish new consensual norms about contextually-appropriate status allocation. We invite stakeholders across sectors to test and refine these tools in diverse contexts, recognizing that creating equitable DTx requires sustained commitment to partnerships that genuinely honor multiple forms of expertise and willingness to disrupt comfortable hierarchies in service of producing interventions truly designed for and with the communities they aim to serve.

This study aims to characterize the temporal discordance between CGM-derived glucose exposure and HbA1c over time in individuals with type 1 diabetes, and to explore the development of a statistical model to adjust the relationship between these measures based on previously observed individual discrepancies. We paired CGM-data in a 60-day window prior to each HbA1c measurement and included individuals with type 1 diabetes with multiple pairs to assess and model discordance over time. Discordance was defined as difference between HbA1c and Glucose Management Indicator at each pair. At baseline (first pair), participants were categorized into three groups based on the degree of discordance: positive (≥0.5%), negative (≤-0.5%), and neutral (within ±0.5%). A multiple linear regression model incorporating historical discordance values, HbA1c levels, and the current GMI was utilized for an adjustment. 477 individuals were included and 1,523 instances of paired HbA1c and CGM-data were analyzed. Absolute discordance of ≥0.5% was observed in 31% of cases. In 51% of instances, the direction of discordance in each pair was maintained. In the modeling analysis, GMI accounted for 69% of the variance in HbA1c levels (r = 0.83, p < 0.001, MAE = 0.42%). Adjusting improved variance explainability to 82% (r = 0.90, p < 0.001, MAE = 0.33%). HbA1c-CGM discordance is highly prevalent, and while inter-individual discordance shows some degree of persistence, it also appears to vary over time for a substantial proportion of individuals. Adjusting for individual discordance in the short term can improve the alignment between adjusted GMI and laboratory-measured HbA1c.