Daily Respiratory Research Analysis
Three studies advance respiratory care across decision support, imaging, and infectious diagnostics. An LLM-guided framework improved alignment of noninvasive respiratory support with guidelines and was associated with lower intubation and mortality/hospice rates. An AI classifier enhanced lung nodule risk stratification over BTS guidelines, potentially expediting cancer diagnosis, while mNGS outperformed conventional testing for lower respiratory tract infections and optimized antimicrobial use
Summary
Three studies advance respiratory care across decision support, imaging, and infectious diagnostics. An LLM-guided framework improved alignment of noninvasive respiratory support with guidelines and was associated with lower intubation and mortality/hospice rates. An AI classifier enhanced lung nodule risk stratification over BTS guidelines, potentially expediting cancer diagnosis, while mNGS outperformed conventional testing for lower respiratory tract infections and optimized antimicrobial use.
Research Themes
- AI-assisted respiratory support decision-making
- Imaging AI for early lung cancer risk stratification
- Metagenomic diagnostics for lower respiratory tract infections
Selected Articles
1. Enhancing predictive modeling for respiratory support with LLM-driven guideline adherence.
Integrating a guideline-driven LLM with a deep counterfactual model improved interpretability and alignment of HFNC vs NIV recommendations. Concordant care with LLM-enhanced recommendations was associated with lower IMV (e.g., 24.5% vs 52.9% for HFNC recommendations) and reduced mortality/hospice discharge (OR 0.670, p=0.046), though chart review flagged occasional potentially severe errors.
Impact: This hybrid decision-support approach addresses a key gap in applying evidence and guidelines to heterogeneous ICU patients and shows outcome signals that justify prospective evaluation.
Clinical Implications: Use LLM-enhanced, guideline-concordant recommendations to support HFNC vs NIV selection in high-risk ICU patients, while instituting guardrails (contraindication detection, clinician oversight) to mitigate rare harmful errors.
Key Findings
- Concordance with LLM-enhanced recommendations was associated with lower IMV rates (HFNC recommendation: 24.47% vs 52.94% when discordant).
- Mortality or hospice discharge was reduced with concordant care (OR 0.670, p=0.046).
- Chart review showed 95% guideline alignment; physicians agreed with 65% final recommendations, with 2/20 cases judged as potentially causing severe harm.
Methodological Strengths
- Real-world cohort with structured comparison of concordant vs discordant management and patient-centered outcomes (IMV, mortality/hospice).
- Integrated LLM configured in a HIPAA-compliant environment with explicit guideline prompting and independent chart review for safety.
Limitations
- Retrospective, observational design susceptible to residual confounding and selection bias.
- Limited chart review sample; two recommendations potentially causing severe harm underline the need for stronger safety guardrails and prospective validation.
Future Directions: Prospective, multi-center trials to test LLM-guided respiratory support pathways; enhance contraindication detection, bias auditing, and clinician-in-the-loop workflows.
BACKGROUND: Optimal respiratory support selection between high-flow nasal cannula (HFNC) and noninvasive ventilation (NIV) for intensive care units (ICU) patients at risk of invasive mechanical ventilation (IMV) remains unclear, particularly in cases not represented in prior clinical trials. We previously developed RepFlow-CFR, a deep counterfactual model estimating individualized treatment effects (ITE) of HFNC versus NIV. However, interpretability and guideline alignment remain challenges for clinical adoption. This study describes the development and integration of a clinical guideline-driven LLM to enhance deep counterfactual model recommendations for NIV versus HFNC in patients at high-risk for invasive mechanical ventilation. METHODS: We enhanced RepFlow-CFR by incorporating a large language model (LLM, Claude 3.5 Sonnet) to enforce clinical guideline adherence and generate explainable treatment recommendations. The LLM was configured in a HIPAA-compliant AWS environment and prompted using structured patient data, clinical notes, and formal guideline criteria. Recommendations from RepFlow-CFR and LLM were compared to actual treatment decisions to assess concordance. We evaluated IMV and mortality/hospice rates across concordant and discordant groups. Additionally, we conducted a structured chart review of 20 cases to assess the clinical validity and safety of LLM-driven recommendations.
2. A CADx tool improves lung nodule risk stratification when compared to British Thoracic Society guidelines on routine computed tomography (CT).
In a retrospective case-control cohort with definitive diagnoses, an AI-derived malignancy similarity index (mSI) adjusted BTS 2015 follow-up recommendations, upgrading 44% of cancers that would have received interval scans and reducing delayed malignant diagnoses from 45% to 25% (P<0.001). Among benign controls, 7% could be safely downgraded to no follow-up.
Impact: Demonstrates that a calibrated imaging AI can reduce delayed cancer diagnoses while avoiding some unnecessary follow-up, offering immediate translational value alongside established guidelines.
Clinical Implications: Incorporating mSI alongside BTS criteria may accelerate work-up for high-risk nodules and reduce surveillance for low-risk cases; multidisciplinary review and external validation are advised before broad adoption.
Key Findings
- Among 100 cancer cases, 45 would have had interval scans; mSI would upgrade 44% (20/45), cutting delayed malignant diagnoses from 45% to 25% (P<0.001).
- Among 100 benign cases, 7% would be downgraded to no follow-up and 7% upgraded to immediate PET-CT.
- mSI used thresholding (mSI >0.9 upgrade; <0.1 downgrade) to adjust BTS recommendations, improving classification performance.
Methodological Strengths
- Ground-truth diagnoses enabled direct assessment of reclassification impact versus BTS guidelines.
- Predefined, transparent reclassification thresholds (upgrade/downgrade/maintain) support reproducibility.
Limitations
- Single-site retrospective study with limited sample size; potential selection bias.
- External generalizability of the AI classifier was not extensively tested; clinical utility requires prospective validation.
Future Directions: Prospective, multi-center impact studies to measure time-to-diagnosis, diagnostic yield, and downstream outcomes when mSI is integrated into lung nodule pathways.
AIM: The aim of this study was to evaluate whether a computed tomography (CT) image-based lung nodule artificial intelligence (AI) classifier can improve lung nodule risk classification and clinical management decisions relative to British Thoracic Society (BTS) 2015 guidelines. MATERIALS AND METHODS: This is a retrospective single-site case-control study of incidental lung nodules identified from routine clinical CT scans. An AI-based classification tool (RevealAI-Lung) was used to compute a malignancy similarity index (mSI) of lung nodules from CT scans of patients with known definitive diagnoses. For comparison to current clinical best practices, nodules were retrospectively classified for follow-up using the BTS 2015 guidelines by a re-review of all cases. The mSI was used to adjust BTS recommendations (msi > 0.9: upgrade, mSI < 0.1: downgrade, and 0.1 < mSI < 0.9: maintain), and performance of the mSI reclassification against BTS guidelines was compared. RESULTS: A total of 100 control patients and 100 cancer patients were analysed from an UK National Health Service (NHS) referral population who had proven diagnosis (benign or malignant). Forty-five of 100 cancer patients would have received interval scans before eventual malignant diagnosis. mSI would have upgraded 44% (n=20/45) of these cases and downgraded none, reducing delayed cancer diagnoses from 45% to 25% (P<0.001). A total 100 control patients would have received followup scans before eventual benign diagnosis. mSI would have downgraded 7% (n=7/100) of these cases to no follow-up and upgraded 7% (n=7/100) from an interval scan to an immediate positron emission tomography (PET)-CT. CONCLUSION: mSI significantly improved nodule classification in incidental lung nodules when combined with current best practice BTS guidelines.
3. Diagnostic value of metagenomic next-generation sequencing in the etiological diagnosis of lower respiratory tract infection.
Across 165 suspected LRTI cases, mNGS achieved a much higher detection rate than conventional methods (86.7% vs 41.8%), identified 29 pathogens missed by standard testing, and prompted treatment changes in 72.1% with antibiotic de-escalation in 32.7%. Performance was consistent across specimen types.
Impact: mNGS offers comprehensive, rapid pathogen profiling that can reshape LRTI management by improving pathogen detection, guiding de-escalation, and uncovering atypical etiologies.
Clinical Implications: When available, mNGS can complement routine microbiology to increase diagnostic yield, tailor antimicrobial therapy (including de-escalation), and detect uncommon or polymicrobial etiologies, particularly in immunocompromised hosts.
Key Findings
- mNGS positivity significantly exceeded conventional testing (86.7% vs 41.8%, P<0.05) and was robust across BALF, blood, tissue, and pleural samples.
- Detected 29 pathogens exclusively via mNGS, including NTM, anaerobes, Prevotella, Legionella gresilensis, Orientia tsutsugamushi, and viruses.
- Led to treatment changes in 72.13% of patients, with antibiotic reduction in 32.73%.
Methodological Strengths
- Head-to-head comparison of mNGS versus conventional methods across multiple specimen types in 165 suspected LRTI cases.
- Assessment of clinical impact on antimicrobial management, including de-escalation.
Limitations
- Single-center observational design without a uniform clinical adjudication framework for causality.
- Potential for over-detection/contamination inherent to mNGS; turnaround time and cost-effectiveness not evaluated.
Future Directions: Prospective diagnostic stewardship trials integrating mNGS into LRTI pathways to assess time-to-targeted therapy, outcomes, and cost-effectiveness.
Metagenomic next-generation sequencing (mNGS) has been widely used in infectious diseases. However, reports on mNGS for lower respiratory tract infection (LRTI) diagnosis remain limited, potentially offering significant value for improving pathogen identification. This study evaluates the diagnostic performance and clinical value of mNGS compared to traditional methods in LRTI. We analyzed traditional and mNGS detection results from 165 patients with suspected LRTI using different specimens including bronchoalveolar lavage fluid (BALF), blood, tissue samples, and pleural effusion. We compared diagnostic differences and characteristics between mNGS and traditional methods, and evaluated the effect of mNGS results on antibiotic treatment.Among 165 cases, 146 (88.48%) patients with LRTI had microbial etiology finally identified. Compared with traditional diagnostic methods, mNGS showed significantly higher positive rate (143/165, 86.7% vs 69/165, 41.8%, P < 0.05). The diagnostic performance of mNGS was not affected by sample types. mNGS demonstrated significant advantage in detecting poly-microbial infections and rare pathogens. Twenty-nine kinds of pathogens were detected only by mNGS, including non-tuberculous mycobacteria (NTM), Prevotella, anaerobic bacteria, Legionella gresilensis, Orientia tsugamushi, and viruses. The pathogen spectrum differed between immunocompetent and immunocompromised individuals. mNGS resulted in treatment changes in 119 patients (72.13%), with 54 patients (32.73%) having reduced antibiotics. mNGS has obvious advantages over traditional detection methods with results unaffected by sample types. mNGS demonstrates significant value for pathogen detection and may provide guidance in clinical practice.