Communications medicine最新文献

Background: The global population of People Living with Dementia (PLWD) is expected to grow rapidly in the coming decades, increasing the need for personalised, generalisable, and scalable prognosis and care planning support. However, current prognostic guidance does not adequately capture the heterogeneity in dementia trajectories, and existing predictive models of dementia progression rely on costly and inaccessible data, limiting their scalability in resource-constrained settings.

Methods: Using clinical assessments, demographic, and medical history data from 153 12-month clinical trajectories collected over three years, two machine learning algorithms were developed to predict 12-month cognitive and functional decline in Alzheimer's Disease (AD) and Mild Cognitive Impairment (MCI). Models were externally validated on 741 trajectories from the ADNI cohort. Cognitive and functional decline were estimated using the Mini-Mental State Exam (MMSE) and Bristol Activities of Daily Living (BADL).

Results: The MMSE model achieves a mean absolute error (MAE) of 1.84 (95% CI: 1.64-2.04) internally and 2.19 in external validation. The BADL model achieves an MAE of 3.88 (95% CI: 3.46-4.30). Baseline scores on ideational praxis, orientation, and word recall are among the strongest predictors of cognitive decline, while independence in food preparation, finances, and dressing are among the top predictors of functional decline.

Conclusions: Our models use only routinely collected and easily accessible data, offering high translational potential. If implemented, our scalable, data-driven prognostic support tool could streamline clinical workflows, support personalised care planning, and provide PLWD and their families with greater clarity and reassurance.

Background: Noise exposure at work can damage hearing at speech-frequency essential for speech perception, leading to communication difficulties, life quality decline, and adverse mental and cognitive outcomes. Early identification of individuals at high risk is crucial for occupational health management. This study aims to develop prediction models to estimate the risk of speech-frequency hearing loss among noise-exposed workers.

Methods: We developed and validated multimodal prediction models using epidemiological, hearing assessment, and genetic information from shipyard workers. The training cohort included 5053 workers and the testing cohort included 2086 workers recruited between 2012 and 2024. Noise exposure was estimated using detailed work durations and workplace measurements. Participants completed questionnaires, underwent standardized hearing examinations, and provided blood samples for genetic analysis. Sex-specific models were constructed based on two commonly used definitions of speech-frequency hearing loss. Longitudinal risk was evaluated using repeated-measures statistical approaches.

Results: Here we show that binaural hearing thresholds at 3 and 6 kHz are the strongest predictors of subsequent speech-frequency hearing loss, together with age and noise exposure (P < 0.001). Longitudinal prediction models demonstrate good discrimination and calibration, with AUCs exceeding 0.80 and C-indices above 0.78 in both training and testing cohorts. Incorporation of genetic variants further improves predictive performance, increasing discrimination by approximately 2% in males and 3% in females.

Conclusions: These findings provide evidence-based prediction tools that enable individualized risk assessment. Practically, identifying workers at high risk would benefit the hearing preservation in the frequencies more relevant to speech sounds and maintain good communication.

Background: Co-testing with human papillomavirus (HPV) DNA testing plus liquid-based cytology is still used in parts of China, although many screening programmes are moving toward HPV-based strategies. We aimed to compare co-testing with HPV-based and cytology-only approaches in routine county services in resource-limited areas.

Methods: We analysed a screening cohort of 33,387 women aged 35-64 years from four primary care sites. Because all women received both HPV testing and cytology, we reconstructed four strategies within the same population: co-testing, HPV primary screening with cytology triage, HPV-only, and cytology-only. For each strategy we estimated detection of cervical intraepithelial neoplasia grade 2 or worse (CIN2 + ), referrals for specialist examination of the cervix, and cytology workload per 1000 women screened.

Results: Here we show that co-testing detects 6.7 CIN2+ cases per 1000 women screened, compared with 6.5 for HPV primary screening with cytology triage, 4.3 for HPV-only, and 4.9 for cytology-only. However, co-testing requires more resources than HPV primary screening with cytology triage, including 33.1 additional colposcopy referrals and 888.8 extra cytology slides per 1,000 women screened, with little gain in detection. Cytology-only increases referrals while detecting fewer CIN2+ cases, whereas HPV-only reduces referrals but detects fewer CIN2 + .

Conclusions: In resource-limited county programmes, HPV primary screening with cytology triage provides the most favourable balance between detecting cervical pre-cancer and limiting unnecessary procedures. These findings support transitioning from routine co-testing to HPV-based screening tailored to local capacity.

Background: Artificial Intelligence tools such as ChatGPT are increasingly used by laypeople to support their care-seeking decisions, although the accuracy of newer models remains unclear. We aimed to evaluate the accuracy of care-seeking advice that is generated by all currently available ChatGPT models.

Methods: We evaluated 22 ChatGPT models using 45 validated vignettes, each prompted ten times (9,900 total assessments). Each model classified the vignettes as requiring emergency care, non-emergency care, or self-care. We evaluated accuracy against each case's gold standard solution (determined by two physicians), examined the variability across trials, and tested algorithms to aggregate multiple recommendations to improve accuracy.

Results: We show that o1-mini achieves the highest accuracy (74%), but we cannot observe an overall improvement with newer models - although reasoning models (e.g., o4-mini) improved their accuracy in identifying self-care cases. Selecting the lowest urgency level across multiple trials improves accuracy by 4 percentage points.

Conclusions: Although newer increasingly provide self-care advice, their accuracy remains insufficient for standalone use. However, making use of output variability with aggregation algorithms can improve the performance of existing models.

Background: Although case reports and observational studies suggest COVID-19 increases the risk of kidney diseases, real-world evidence comparing it with influenza is limited. Our study aims to assess the association between COVID-19 infections and subsequent kidney diseases, using influenza as a positive control and incorporating a negative control to establish clearer associations.

Methods: A large retrospective cohort study with strata matching was conducted using the MarketScan database with records from Jan. 2020 to Dec. 2021. We used the ICD-10 codes to identify individuals and build three cohorts: (1) COVID-19 group, (2) Positive control group: Influenza but no COVID-19, and (3) Negative control group: no COVID-19 / Influenza. The outcomes were acute kidney injury (AKI), chronic kidney disease (CKD), end-stage renal disease (ESRD), and glomerular diseases. Multivariable stratified Cox proportional hazards regression analyses were performed.

Results: The study includes 939,241 individuals with COVID-19, 1,878,482 individuals in the negative control group, and 199,071 individuals with influenza. COVID-19 is significantly associated with increased risks of AKI (adjusted hazards ratio, aHR: 2.74; 95% CI, 2.61-2.87), CKD (aHR: 1.38, 1.32-1.45), ESRD (aHR: 3.22, 2.67-3.88), and glomerular diseases (aHR:1.28, 1.09-1.50), while influenza has no impact on CKD, ESRD, and glomerular diseases. Time-specific analyses indicate that COVID-19 has stronger effects on AKI in the short term but has stable long-term effects on CKD.

Conclusions: In this large real-world study of working-age, commercially insured adults in the United States, COVID-19 infection is associated with a 2.3-fold risk of developing AKI, a 1.4-fold risk of CKD, and a 4.7-fold risk of ESRD compared to influenza. Greater attention to kidney diseases post-COVID-19 is essential to prevent future adverse health outcomes.

Background: The molecular features determining the risk of metachronous metastases in clear cell renal cell carcinoma (ccRCC) are poorly defined. This study aimed to identify molecular factors associated with the risk of metachronous metastasis.

Methods: Using a systematic tumor transcriptome deconvolution approach, we investigated the genomic and transcriptomic profiles of 192 ccRCC primary tumors with extended clinical follow-up to identify cancer- and stromal cell-specific molecular features associated with metastatic risk. Based on these features, we applied multivariate Cox regression to develop a compact 5-gene predictive model for metachronous metastasis.

Results: At the genomic level, we identify a significantly higher frequency of copy number loss at 1p31-36 in primary tumors that later progress with metastases. Tumor transcriptome deconvolution identifies significant down-regulation of epithelial cell polarity, including PATJ (1p31), and fatty acid metabolism, including CYP4A11 (1p33), in cancer cells of tumors that develop metastatic progression. We develop and benchmark a compact 5-gene predictive model (5G) that demonstrates improved accuracy over existing ccRCC gene signatures in the prediction of metachronous metastasis risk.

Conclusions: Overall, our study highlights convergent genomic and transcriptomic alterations in chromosome 1p, driving dysregulation of epithelial cell polarity and fatty acid metabolism, as putative risk factors of metachronous metastasis in ccRCC.

Background: Reliable biomarkers for predicting disease progression in multiple sclerosis (MS) are crucial for advancing precision medicine and optimising treatment strategies. This study evaluates the predictive potential of serum nuclear magnetic resonance (NMR)-based metabolomics, individually and in combination with well-established biomarkers of neuroinflammation (serum glial fibrillary acidic protein, sGFAP) and axonal damage (neurofilament light chain, sNfL), in an extreme-phenotype subset of the Swiss Multiple Sclerosis Cohort (SMSC).

Methods: Serum samples were analysed using NMR-based metabolomics, along with quantification of sNfL and sGFAP. Supervised multivariate analysis was performed to differentiate MS phenotypes and identify future progressors. Multivariable receiver operating characteristic (ROC) analysis evaluated predictive performance, with key metabolite findings validated in an independent Oxford MS cohort.

Results: NMR-based metabolomics reliably distinguishes relapsing-remitting MS (RRMS) from secondary-progressive MS (SPMS) and predicts individual transitions. The identified predictive metabolites (lipoproteins, glutamine, alanine, valine, glucose) are also associated with progression independent of relapse activity (PIRA), a clinically relevant marker of sustained disability worsening. This demonstrates that the approach can both stage disease and forecast progression irrespective of stage. ROC analysis shows strong predictive performance (AUC = 0.81, p = 0.001), with external validation confirming robustness. Integration of NMR-metabolomics with sGFAP and sNfL further improves accuracy, yielding AUCs of 0.91 (p < 0.0001) and 0.87 (p = 0.0002), respectively, supported by independent validation.

Conclusions: The integration of metabolic and protein biomarkers enables both accurate staging of RRMS versus SPMS and, critically, early prediction of progression irrespective of stage. This dual capability provides a clinically actionable, serum-based tool that can refine monitoring, improve therapeutic decision-making, and support a shift towards stage-agnostic, progression-focused care in MS.

Background: Diabetic retinopathy (DR) is a complication of diabetes that affects the eyes. This study aims to identify the genetic variants associated with DR in type 2 diabetes (T2D) patients from the UK Biobank cohort (n = 16,988).

Methods: We conducted a genome-wide association study (GWAS) of DR and integrated genomic results with multi-omics data to identify and prioritize susceptibility variants and genes. The findings are set to undergo validation in four replication cohorts.

Results: Here we show that the lead SNP rs6066146 in EYA2 reaches genome-wide significance (p = 4.21×10^-8) and is replicated in three independent cohorts. The SNP-based heritability for DR is estimated at 14.6% (standard deviation: 0.11). Colocalization analysis at the EYA2 locus suggests moderate colocalization (PP.H4 = 0.553) alongside distinct association signals for DR and T2D, and cis-Mendelian randomization (MR) within the EYA2 region provides gene-centric evidence that T2D exerts a significant causal effect on DR. Exploratory multivariable MR identifies proinsulin as a significant mediator of T2D on DR, which may partly account for the moderate evidence for colocalization. Tissue expression, chromatin interaction, and transcriptome-wide association analyses point to the spleen, while gene set analysis identifies B-cell pathways. Together, these convergent signals suggest that splenic B-cell abundance could serve as a predictive marker for DR risk.

Conclusions: Our study demonstrates a genomic risk locus in gene EYA2 associated with DR in type 2 diabetes, which offers deeper insights into broader trait architecture on DR.

Background: Achieving Universal Health Coverage (UHC) in low- and middle-income countries (LMICs) requires a robust digital infrastructure capable of monitoring healthcare services and associated costs. A major barrier is the absence of a standardized and comprehensive nomenclature for billable healthcare services. Assessment across five hospitals in Burundi confirmed this problem by showing significant inconsistencies in service naming and coding. This study presents the development of a Universal Nomenclature of Health Services (UNHS) for Burundi, a meta-classification for billable health services, designed to align international classification systems with local operational needs.

Methods: The methodology comprised a need assessment, literature review, the selection of relevant international standards, national adaptation, integration of operational sub-codes, and validation through stakeholder engagement.

Results: The developed meta-classification based on six international standards (ICD-10-PCS, CPT, HCPCS, LOINC, RxNorm, and UB04) produces 82,433 codes covering 97.7% of health services relevant to UHC tracking.

Conclusions: This paper details the methodology, structure, coverage, and implementation of the UNHS, offering a scalable model for improving health information system interoperability and UHC monitoring in LMICs.

Background: Equitable deployment of clinical artificial intelligence systems requires consistent performance across diverse patient populations. However, race information in electronic health records is often missing/inconsistently documented, limiting the ability to construct representative cohorts or assess algorithmic bias. This study evaluates model performance and fairness in predicting race from clinical text.

Methods: We compared four transformer-based deep learning models with a hierarchical convolutional neural network designed to capture the multilevel structure of clinical narratives. A two-phase active learning framework guided annotation of a primary care database. A fairness-aware loss function was applied to mitigate disparities across racial groups. Each model was trained with and without fairness-aware optimization. Performance and equity were evaluated using 10-fold cross-validation and subgroup audits across race, sex, age, and their intersections.

Results: Here we show that the hierarchical convolutional neural network achieves higher accuracy and performance equity than transformer models (macro F1 = 98.4%). Fairness constraints enhance parity across most transformer architectures, but degrade hierarchical model performance and cause one clinical model to collapse toward majority predictions, demonstrating that fairness interventions are highly model dependent. Persistent disparities across race, sex, and age indicate that inequities reflect architectural limitations and systemic biases.

Conclusions: This study demonstrates that fairness can be integrated into clinical language models, though effects vary by model type. Architectures aligned with clinical text structure inherently promote fairness, yet mixed fairness constraint outcomes highlight the need for tailored interventions. Persistent demographic disparities show that algorithmic bias often reflects upstream documentation inequities. This framework offers a scalable path toward equitable NLP for clinical artificial intelligence.