Research Appraisals
Evidence-based critical appraisals of the latest medical research, systematically evaluated using Oxford CEBM methodology.
Showing 3 appraisals
Sleep
Multimodal cardiovascular risk profiling using self-supervised learning of polysomnography
STUDY OBJECTIVES: Polysomnography (PSG) provides a comprehensive assessment of brain, cardiac, and respiratory activity during sleep. While it is widely used for diagnosing sleep disorders, its potential to assess future health risks has not been fully explored. This study aimed to develop and evaluate an interpretable framework to identify physiological patterns in PSG data linked to cardiovascular disease (CVD) outcomes, without relying on manual annotations (e.g. sleep stages). METHODS: We developed a self-supervised deep learning model that extracts meaningful patterns from multi-modal signals (electroencephalography [EEG], electrocardiography [ECG], and respiratory signals [Resp]). The model was trained on data from 4398 participants. Projection scores were derived by contrasting embeddings from individuals with and without CVD outcomes. External validation was conducted in an independent cohort with 1093 participants. The source code is available on https://github.com/miraclehetech/sleep-ssl. RESULTS: The projection scores revealed distinct and clinically meaningful patterns across modalities. ECG-derived features were predictive of both prevalent and incident cardiac conditions, particularly CVD mortality. EEG-derived features were predictive of incident hypertension and CVD mortality. Resp added complementary predictive value. Combining projection scores with the Framingham Risk Score consistently improved prediction, with area under the curve values ranging from 0.607 to 0.965 in the Sleep Heart Health Study test cohort and 0.710 to 0.807 across four of five outcomes in the external validation cohort, demonstrating robust cross-cohort generalizability. CONCLUSIONS: Our findings demonstrate that the proposed framework can generate individualized CVD risk scores directly from PSG data. The resulting projection scores have the potential to be integrated into clinical practice, enhancing risk assessment and supporting personalized care. Statement of Significance This study is among the first to apply a self-supervised framework for cardiovascular risk profiling using polysomnography (PSG) data. By transforming electroencephalography, electrocardiography, and Resp into interpretable projection scores, we identified physiological markers predictive of multiple cardiovascular outcomes. These risk profiles, combined with traditional risk scores, significantly improved prediction across both internal and external cohorts. Our findings highlight the untapped potential of PSG signals beyond sleep staging, providing an interpretable, scalable, and clinically actionable approach for personalized cardiovascular risk stratification.
17 Apr 2026
Read appraisal →Heart (British Cardiac Society)
Parsimonious machine learning model to predict 1-year mortality and procedural futility after transcatheter aortic valve replacement
BACKGROUND: Current risk scores inadequately predict long-term mortality after transcatheter aortic valve replacement (TAVR), limiting their ability to guide decisions around procedural futility. We aimed to develop and externally validate a machine learning (ML) model using only preprocedural variables to predict 1-year all-cause mortality. METHODS: An ML model was trained on a retrospective cohort of 1025 TAVR patients using 52 clinical and echocardiographic variables. Feature selection and model tuning were performed via a multiobjective evolutionary algorithm to optimise predictive performance and model simplicity. The final model used 13 preprocedural variables and was externally validated in an independent cohort of 270 patients. Performance was compared with European System for Cardiac Operative Risk Evaluation II (EuroSCORE II), FRANCE-2 and TAVI2-SCORE using the area under the curve (AUC), calibration and net reclassification improvement (NRI). RESULTS: The ML model demonstrated excellent discrimination, with AUCs of 0.81 in the discovery cohort and 0.84 in the external validation cohort. This exceeded the performance of EuroSCORE II (AUC: 0.61 and 0.70), FRANCE-2 (0.52 and 0.58) and TAVI2-SCORE (0.56 and 0.64). Calibration plots showed strong agreement between predicted and observed risks. NRI in the test set compared with FRANCE-2 was 0.62 (95% CI: 0.49 to 0.75); compared with TAVI2-SCORE, it was 0.36 (95% CI: 0.14 to 0.61). The final model incorporated age, atrial fibrillation, creatinine, haemoglobin, pulmonary function, frailty markers (Katz Index, poor mobility) and tricuspid regurgitation. Misclassification analysis revealed that most errors were clustered near the decision threshold, with no evidence of systematic bias. Performance was consistent across subgroups and robust to temporal and institutional variation. CONCLUSION: This externally validated ML model, using 13 routinely available variables, significantly outperforms existing risk scores in predicting 1-year mortality post-TAVR. Its simplicity and generalisability support its potential use in real-world clinical decision-making to identify patients at high risk of procedural futility.
12 Apr 2026
Read appraisal →Journal of hazardous materials
Structuring large language models for chemical health risk reasoning in environmental and occupational exposure
Large language models (LLMs) have revolutionized numerous scientific fields. However, their application in environmental health remains limited, as they are prone to hallucinations and struggle with reasoning over complex, real-world exposure data. Chemical risks, especially from rare or poorly studied compounds, span diverse occupational and everyday scenarios, where structured knowledge is often incomplete and poorly integrated. To address these challenges, we present a health risk assistant that integrates deep learning-based compound property prediction, retrieval-augmented generation (RAG), and prompt engineering into a unified reasoning framework. Central to our design are two synergistic strategies: semantic structuring, which reorganizes heterogeneous knowledge into context-consistent and semantically rich profiles, and domain-aware prompting, which guides large language models through structured and role-specific reasoning aligned with toxicological goals. The framework enables the system to map physicochemical attributes to exposure pathways and disease outcomes with improved relevance and fidelity. Evaluated on a 100-question benchmark spanning chronic toxicity, occupational hazards, and exposure-specific risks, our LLM-RAG system consistently outperforms baseline models in correctness, faithfulness, and domain-specific helpfulness. Beyond answering queries, our approach demonstrates how integrated knowledge curation and tailored prompting can enable trustworthy and scalable AI reasoning for public and environmental health protection.
4 Apr 2026
Read appraisal →