Daily Respiratory Research Analysis
Analyzed 145 papers and selected 3 impactful papers.
Summary
An updated Bayesian meta-analysis of 32 RCTs finds that selective decontamination of the digestive tract (SDD) likely reduces in-hospital mortality among mechanically ventilated adults. Advances in respiratory informatics include Influ-BERT, a domain-adaptive genomic language model that achieves >97% F1 in influenza A subtype identification, and a multicenter prospective model that predicts early failure of high-flow nasal oxygen (HFNO) at initiation in COVID-19 AHRF.
Research Themes
- Critical care infection prevention and mortality outcomes (SDD in ventilated adults)
- AI/genomics for respiratory virus surveillance and subtype identification
- Early triage and escalation decisions in acute hypoxemic respiratory failure (HFNO failure prediction)
Selected Articles
1. Selective Decontamination of the Digestive Tract in Adult Mechanically Ventilated Patients - An Updated Systematic Review with Bayesian Meta-Analysis.
This Bayesian update of a comprehensive RCT meta-analysis (32 trials; 27,687 patients) shows SDD is associated with lower in-hospital mortality (RR 0.91; 95% CrI 0.82–0.99) in mechanically ventilated adults. The probability of benefit supports reconsideration of SDD within antimicrobial stewardship frameworks.
Impact: Clarifies mortality impact of SDD with modern evidence synthesis and large sample size, informing ICU infection prevention policy after years of equipoise.
Clinical Implications: ICUs may consider SDD to reduce mortality in ventilated adults alongside robust antimicrobial stewardship, surveillance for resistance, and protocolized implementation.
Key Findings
- Pooled hospital mortality risk with SDD vs usual care/placebo: RR 0.91 (95% CrI 0.82–0.99).
- 32 RCTs included (27,687 participants); 30 trials (27,332) contributed to the primary outcome.
- Bayesian framework indicates a high probability that SDD reduces in-hospital death among ventilated adults.
Methodological Strengths
- Comprehensive RCT-only synthesis with large aggregate sample.
- Bayesian meta-analytic approach providing credible intervals and probability of benefit.
Limitations
- Heterogeneity in SDD protocols and co-interventions across trials.
- Potential gaps in antimicrobial resistance and adverse effect reporting across included studies.
Future Directions: Prospective, stewardship-embedded pragmatic trials and surveillance platforms to quantify resistance/ecologic effects, and define optimal SDD protocols.
BACKGROUND: There is uncertainty whether the use of selective decontamination of the digestive tract (SDD) as a preventive antimicrobial strategy reduces mortality in adult patients receiving mechanical ventilation in the intensive care unit (ICU). Following the publication of new data from a contemporary randomized clinical trial, an updated systematic review and meta-analysis was conducted to determine whether the use of SDD reduced hospital mortality compared to standard care. METHODS: An updated systematic review of a previously published meta-analysis was conducted including a search from September 12, 2022, to August 18, 2025, for randomized clinical trials (RCTs) of adults receiving mechanical ventilation in an ICU that compared SDD to standard care. Data were pooled using a Bayesian framework. The primary outcome was hospital mortality or closest approximation. RESULTS: One additional trial was identified, giving a total of 32 RCTs (27,687 participants), with 30 of the 32 RCTs (27,332 participants) contributing data to the primary outcome. The pooled estimated relative risk of hospital mortality for SDD compared to usual care or placebo was 0.91; 95% credible interval, 0.82 to 0.99, CONCLUSIONS: There is a high probability that in mechanically ventilated adults in the ICU, SDD, compared to standard care, is associated with a reduction in the risk of in-hospital death.
2. Influ-BERT: a domain-adaptive genomic language model for advancing influenza A virus research.
Influ-BERT, trained on ~900,000 influenza genomes with a custom tokenizer and two-stage regimen, surpassed conventional ML and general genomic LLMs, achieving F1 >97% across five subtype tasks and improving performance for underrepresented subtypes. It generalized to respiratory virus identification, pathogenicity prediction, and gene/fragment localization with biologically meaningful attention.
Impact: Establishes a state-of-the-art domain-adaptive genomic LLM for influenza, with direct applications in surveillance, risk assessment, and diagnostics for respiratory viruses.
Clinical Implications: Supports faster, more accurate subtype attribution and pathogenicity inference from sequence data, enhancing outbreak detection, vaccine strain assessment, and diagnostic pipelines.
Key Findings
- Two-stage, domain-adaptive training on ~900,000 influenza genomes with a custom BPE tokenizer.
- Achieved F1 >97% across five influenza A subtype classification tasks; robust for rare subtypes (e.g., H5N8, H1N2, H13N6).
- Generalized to respiratory virus identification, IAV pathogenicity prediction, and gene/fragment localization; interpretability highlighted biologically relevant regions.
Methodological Strengths
- Large, domain-specific pretraining corpus with tailored tokenizer and two-stage optimization.
- Benchmarking against multiple strong baselines and tasks plus interpretability via perturbation analysis.
Limitations
- Lacks prospective clinical validation within real-world surveillance/diagnostic workflows.
- Potential dataset and lineage biases; generalizability to emergent variants outside training distribution requires ongoing assessment.
Future Directions: Prospective integration into public health pipelines, continual learning for emerging variants, and calibration for clinical decision support.
Influenza A virus (IAV) poses a persistent threat to global public health due to its broad host adaptability, frequent anti-genic variation, and potential for cross-species transmission. Accurate identification of IAV subtypes is essential for effective epidemic surveillance and precise disease control. Here, we present Influ-BERT, a domain-adaptive pretrained model based on the Transformer architecture. Optimized from DNABERT-2, Influ-BERT was developed using a dedicated corpus of ~900 000 influenza genome sequences. We constructed a custom Byte Pair Encoding tokenizer, and employed a two-stage training strategy involving domain-adaptive pretraining followed by task-specific fine-tuning. This approach significantly enhanced identification performance for IAV subtypes. Experimental results demonstrate that Influ-BERT outperforms both traditional machine learning approaches and general genomic language models, such as DNABERT-2, Necleotide Transformer, and MegaDNA, in the task of IAV subtype identification. The model consistently achieved F1-scores above 97% across five subtype classification tasks and exhibited stable performance gains for subtypes that are underrepresented in sequencing data, including H5N8, H1N2, and H13N6. Beyond subtype identification, Influ-BERT was successfully applied to additional tasks including respiratory virus identification, IAV pathogenicity prediction, and identification of IAV genomic fragments and functional genes, demonstrating robust performance throughout. Further interpretability analysis using sliding window perturbation confirmed that the model focuses on biologically significant genomic regions, providing insight into its improved predictive capability.
3. Predicting Failure at Initiation of High-Flow Nasal Oxygen in Patients With COVID-19: Literature Review, Development and Internal Validation of a Prediction Model.
In a 10-center, prospective cohort of 608 COVID-19 AHRF patients starting HFNO, 46% required intubation. A parsimonious model using age, urea, platelet count, respiratory rate, oxygen saturation, and FiO2 predicted HFNO failure with good internal validation via bootstrapping.
Impact: Provides an actionable, early triage tool at the moment of HFNO initiation, potentially improving timing of escalation and resource allocation in AHRF.
Clinical Implications: Use of this model may guide earlier intubation or closer monitoring in high-risk patients and support standardized decision-making pathways for HFNO in COVID-19 AHRF.
Key Findings
- Failure (intubation) occurred in 277/608 (46%) HFNO starts across 10 Dutch centers.
- Independent predictors: higher age, higher urea, lower platelet count, higher respiratory rate, lower oxygen saturation, and higher FiO2 requirement.
- Model showed good performance with internal validation by bootstrapping.
Methodological Strengths
- Multicenter prospective design with pre-specified predictors.
- Internal validation using bootstrapping to mitigate optimism.
Limitations
- No external validation; model derived in COVID-19 may need recalibration for other viral AHRF or practice settings.
- Outcome limited to intubation; competing risks and clinician thresholds may vary.
Future Directions: External validation across pathogens and health systems, dynamic updating, and integration into bedside decision support to optimize HFNO pathways.
BACKGROUND AND OBJECTIVE: High-Flow Nasal Oxygen (HFNO) can reduce the need for invasive mechanical ventilation in patients with acute hypoxemic respiratory failure (AHRF) from viral pneumonias, like COVID-19. Early prediction of HFNO failure is useful for timely decision-making at HFNO initiation. This study aimed to develop a prediction model for HFNO failure using predictors available just prior to HFNO initiation in patients with COVID-19 AHRF and compare its performance to existing models. METHODS: This multicenter, prospective observational cohort study included hospitalized patients from 10 centers in the Netherlands between December 2020 and July 2021. Adults who tested positive for SARS-CoV-2, had no treatment limitations, and initiated HFNO for hypoxemia were included. The primary outcome was HFNO failure, defined as the event of endotracheal intubation. Pre-defined candidate predictors were selected by multivariable logistic regression for prediction model development. Internal validation was conducted using bootstrapping. RESULTS: Out of 608 patients, 277 (46%) experienced HFNO failure. Independent predictors of HFNO failure included (odds ratio [95% CI]): age (1.02 [1.00-1.03]), urea (1.04 [1.00-1.08]), platelet count (0.94 [0.92-0.97]), respiratory rate (1.05 [1.02-1.08]), oxygen saturation (0.89 [0.84-0.94]), and FiO CONCLUSIONS: This newly developed model, using variables available at HFNO initiation, effectively predicted HFNO failure in hospitalized hypoxemic patients due to COVID-19 pneumonia with good performance. REGISTRATION NUMBER CLINICAL TRIAL: Dutch Trial Registry: DTR, NL9067.