Artificial Intelligence in Respiratory Care: An Educational Review

1. Introduction

Respiratory diseases are among the most prevalent health conditions worldwide. COPD alone affects over 300 million people and is projected to become the third leading cause of death globally.[1] Asthma, pneumonia, tuberculosis, and ARDS further contribute to the burden, particularly in low- and middle-income countries. Traditional respiratory care relies on clinical expertise, imaging, and pulmonary function testing, but these methods often face limitations in accuracy, timeliness, and scalability.

Mechanical ventilation, while lifesaving, presents additional challenges including ventilator-induced lung injury (VILI), patient-ventilator asynchrony, and prolonged dependence. VILI contributes to mortality rates of up to 30% in ARDS patients when tidal volumes exceed 10 mL/kg.[2] Patient-ventilator asynchrony occurs in 25–50% of ventilated patients, prolonging ventilation by 2–3 days.[3] Manual ventilator adjustment is limited by clinician workload, cognitive overload, and inter-clinician variability.

AI technologies—including machine learning (ML), deep learning (DL), natural language processing (NLP), and reinforcement learning (RL)—offer new opportunities to enhance respiratory care. By analyzing large datasets and providing continuous real-time optimization, AI can detect subtle patterns, predict disease progression, support personalized treatment, and automate complex ventilator management.[4,5,6,7,8] The integration of AI into respiratory medicine represents a paradigm shift toward precision healthcare, though adoption remains limited across most domains despite substantial potential impact on quality improvement, patient safety, patient experience, and access to care.[9]

2. Applications of AI in Respiratory Diagnosis and Monitoring

2.1. Diagnosis and Monitoring

AI algorithms analyze chest imaging, spirometry, and wearable sensor data to detect early signs of respiratory disease. Convolutional neural networks (CNNs) can identify pneumonia or tuberculosis from chest X-rays with accuracy comparable to radiologists.[4,7,8] As of early 2023, there were 42 FDA-approved AI-based software devices specifically for thoracic radiology, primarily for pulmonary nodule detection, endotracheal tube placement monitoring, detection of emergent findings, and assessment of pulmonary parenchyma.[10] AI also supports continuous monitoring of patients with chronic conditions, alerting clinicians to early signs of deterioration.[4,5,6]

2.2. Imaging and Radiology

Radiology is a major area of AI application. Deep learning models enhance detection of lung nodules, interstitial lung disease, and COVID-19-related abnormalities.[4,7,10] AI reduces diagnostic variability and speeds up interpretation, particularly in resource-limited settings where radiologists are scarce. A comparative analysis identified 222 AI/ML-based medical devices approved in the USA and 240 in Europe between 2015 and 2020, with radiology dominating and most devices classified as low- or moderate-risk.[11] However, it is critical to note that 98% of 790 new devices in pulmonary, sleep, and critical care medicine cleared between 2014 and 2024 used the 510(k) pathway, indicating approval based on equivalence to existing devices rather than direct evidence of improved clinical outcomes.[12]

2.3. Pulmonary Function Testing

Spirometry is a cornerstone of respiratory diagnosis, but interpretation errors are common. AI models improve accuracy by analyzing flow-volume curves and integrating patient demographics.[4,7,8] Algorithms detect technical errors and classify spirometric patterns, with some claiming to outperform pulmonologists, though rigorous multicenter validation is essential for trust and adoption.[4] This reduces misdiagnosis and supports personalized treatment plans.

2.4. Wearable Devices and Remote Monitoring

AI-powered wearables track respiratory rate, oxygen saturation, and activity levels. These devices enable continuous monitoring of patients with COPD or asthma, allowing early intervention when exacerbations are detected.[4,5,6] Connected devices integrated with environmental data help forecast asthma and COPD exacerbations, while telehealth and predictive models enable earlier, more personalized interventions.[4,5] Remote monitoring reduces hospital visits and empowers patients to manage their conditions. A systematic review and meta-analysis of 35 RCTs found that most digital health interventions involving real-time monitoring or teleconsultations for asthma and COPD were cost-effective or less expensive than standard care, with incremental cost-effectiveness ratios ranging from €3,530.93/QALY to €286,369.28/QALY.[13]

3. AI in Mechanical Ventilation: Maintenance and Weaning

3.1. The Challenge of Conventional Ventilator Management

Conventional ventilator management relies on intermittent human assessment, whereas AI systems provide continuous optimization using high-frequency physiologic input. Modern RL-based controllers can propose tidal volume, PEEP, FiO₂, and pressure-support changes to minimize VILI while preserving respiratory muscle activity.[14,15] Unlike manual adjustments, which are intermittent and often subject to cognitive overload, AI systems evaluate thousands of data points per minute.[14]

3.2. Closed-Loop Ventilation Systems

AI-assisted mechanical ventilation provides continuous, automated interpretation of lung compliance, resistance, gas-exchange metrics, airway pressures, and respiratory drive signals.[16] Digital lung-twin models simulate recruitment, overdistension, and derecruitment patterns, allowing adaptive PEEP strategies tailored to the patient's lung morphology.[14] Studies demonstrate improved synchrony and shortened weaning times when AI is used adjunctively with clinician oversight.[17,18]

SmartCare/PS and Adaptive Support Ventilation (ASV) are early examples of autonomous systems.[16] IntelliVent-ASV represents a more advanced closed-loop system. A prospective crossover study of IntelliVent-ASV in 85 ICU patients showed improved oxygenation and reduced variability compared to manual settings.[19] Automated systems reduced hypoxemia episodes by 40% and hypercapnia events by 35%.[19] A systematic review demonstrated that IntelliVent-ASV maintained consistent adherence to protective ventilation, keeping tidal volumes 8 mL/kg in >95% of cases, thereby reducing the incidence of VILI compared to conventional ventilation.[20]

More advanced systems integrate ultrasound, bioacoustic signals, esophageal pressure monitoring, and proportional assist indicators.[14] Machine learning models achieved >90% sensitivity in detecting patient-ventilator asynchrony compared to manual waveform inspection.[21]

3.3. AI in Ventilator Weaning

Weaning from mechanical ventilation remains a critical challenge. Each failed extubation adds approximately 5 ICU days, costing roughly $15,000 per patient.[22] AI-based prediction models have improved extubation success rates from approximately 70% (traditional indices) to >85%.[23] AI models achieved AUC scores of 0.87–0.92 in predicting weaning readiness, outperforming the Rapid Shallow Breathing Index (RSBI) which typically achieves AUC ~0.65.[23]

A retrospective study evaluating AI integration into weaning protocols in medical and COVID-19 ICUs found that in the COVID-19 ICU, the AI-assisted intervention group (n=88) had a mean ventilation time of 244.2 hours versus 426.0 hours in the control group (P=0.011), ICU length of stay of 11.0 days versus 18.7 days (P=0.001), and hospital length of stay of 23.5 days versus 40.4 days (P0.001).[24] These reductions represent a 42.6% reduction in ventilation time, 41.2% reduction in ICU stay, and 41.8% reduction in hospital stay.[24]

A two-stage AI prediction system tailored weaning timing, reducing reintubation rates from 18% to 10%.[25] AI-guided reduction in PEEP levels improved weaning success rates by 15% compared to clinician-led protocols.[26] Across 12 trials, AI-driven ventilation reduced ICU stay by 1.5–2.5 days on average.[27]

3.4. Safety, Explainability, and Ethical Oversight

Ethical oversight requires transparency, explainability, and safety limits to prevent adverse ventilatory decisions.[14] AI does not replace clinician judgment but augments it through tireless pattern recognition and rapid multivariable optimization.[14] AI ventilation systems require integration of >20 physiological variables; missing data reduced accuracy by 15%.[27] Liability remains a concern; clinicians must retain oversight despite automation.[27]

4. Evidence of Clinical Impact Across Respiratory Care

4.1. Quantitative Outcomes in COPD Management

Wang et al. developed and validated an artificial neural network (ANN) model to predict 30-day readmission risk in COPD patients across eight hospitals within a large urban health system.[28] The model, implemented via a smartphone application (Re-Admit), used four clinical variables and achieved an AUC of 0.77 with sensitivity of 0.75 and specificity of 0.67. Following implementation of a targeted clinical care plan for high-risk patients, the study reported a statistically significant 48% reduction in 30-day readmission rates in the high-risk subgroup.[28] This represents one of the few studies demonstrating a direct, quantifiable improvement in a patient-centered outcome attributable to an AI intervention in respiratory care.

4.2. Pulmonary Rehabilitation Outcomes

A systematic review and meta-analysis pooled three RCTs (456 participants) comparing AI-assisted pulmonary rehabilitation to standard care in chronic respiratory diseases, demonstrating a statistically significant improvement in 6-minute walk distance (mean difference: 22.08 meters; 95% CI: 4.96–39.20; p=0.01).[29] While this is a validated functional outcome, the number of included studies is small, and data on long-term quality of life or survival were not reported.

4.3. Diagnostic Accuracy and Clinician Performance

A randomized clinical vignette survey study directly measured the impact of AI on diagnostic accuracy for acute respiratory failure.[30] The study demonstrated that while standard AI models improved diagnostic accuracy, systematically biased models decreased physician performance, and providing AI explanations did not mitigate this harm.[30] Notably, a majority of clinicians were unaware that AI models could be systematically biased.[30] This highlights critical patient safety concerns regarding automation bias and over-reliance on AI predictions.

4.4. Comparative Effectiveness Evidence

A systematic review by Lam et al. found that in 77% of 39 RCTs across specialties, AI-assisted interventions outperformed usual care, with clinically relevant outcomes improving in 70% of those studies.[31] However, most trials were small and single-center, and only a minority directly addressed respiratory conditions. A randomized clinical trial at the Mayo Clinic evaluated an AI-assisted clinical decision support tool for childhood asthma management and found no significant difference in asthma exacerbation frequency between AI and standard care groups (12% vs. 15%; OR 0.82; 95% CI 0.374–1.96; P=0.626), though the AI intervention significantly reduced clinician EHR review time (median 3.5 minutes vs. 11.3 minutes; P0.001).[32]

4.5. Limitations of Current Evidence

Systematic reviews of AI in ARDS report high model performance (AUC 0.8–1.0) for diagnostic and prognostic tasks, but most studies are retrospective and focused on model development rather than prospective clinical impact.[33,34,35] Similarly, AI-based risk scores for community-acquired pneumonia enhance prediction accuracy but lack direct evidence of improved survival or reduced readmissions from prospective trials.[36,37] The predominance of retrospective, single-center studies, the paucity of prospective multicenter RCTs, and the lack of data on long-term patient outcomes such as survival and sustained quality of life represent significant limitations of the current evidence base.[4,5,35,38,39]

5. Regulatory Status and Real-World Adoption

The FDA has articulated a tailored regulatory framework for AI-enabled medical devices, including a 5-point action plan and guidance for clinical decision-support software, emphasizing a risk-based approach.[40] As of 2025, approximately 1,000 AI-enabled devices have been approved, with a focus on fostering collaboration, promoting harmonized standards, advancing regulatory approaches supporting innovation, and supporting research related to AI performance evaluation.[40] The FDA's Total Product Life Cycle approach emphasizes continuous monitoring and iterative improvement of AI-enabled devices throughout their deployment.[40]

Real-world adoption remains heterogeneous. While AI tools for chest imaging and spirometry are increasingly used, widespread adoption is contingent upon rigorous multicenter validation and real-world evidence.[4] The Global Initiative for Chronic Obstructive Lung Disease (GOLD) 2026 report acknowledges that AI and telehealth may offer improved healthcare access and outcomes for COPD patients, but notes that evidence is still emerging and that AI comes with risks and limitations requiring careful consideration before deployment.[1]

6. Algorithmic Bias and Health Equity

6.1. Evidence of Systematic Bias

State-of-the-art AI classifiers for chest X-ray pathology consistently and selectively underdiagnose disease in female, Black, Hispanic, and low socioeconomic status patients, with the underdiagnosis rate being particularly pronounced in intersectional subgroups such as Hispanic women.[41,42] Deep learning models for chest X-ray disease detection encode protected characteristics such as race and biological sex within their feature representations, resulting in shifted true and false positive rates across demographic groups.[43]

AI models for predicting asthma exacerbation in children performed worse for those with lower socioeconomic status, linked to incomplete EHR data in lower SES groups.[44] Bias can be introduced through the physical design of medical technologies, unrepresentative training data, and conflation of biological with social determinants of health.[45] A scoping review found that 74.7% of clinical ML models evaluated for bias demonstrated bias on sociodemographic factors, with bias disproportionately affecting groups with socioeconomic disadvantage.[46]

6.2. Regulatory and Equity Frameworks

Current regulatory guidance does not consistently include equity as a core criterion and lacks mechanisms to systematically detect or prevent bias.[47] Proprietary datasets and algorithms hamper transparency and can obscure algorithmic bias, and underrepresentation of racial and ethnic minority groups in training data is rampant.[48] Realizing AI's benefits will require proactive bias detection and mitigation with inclusive sampling, equity audits, and stratified performance reporting.[4,47,48]

7. Patient Safety and Implementation Barriers

7.1. Automation Bias and Over-Reliance

Long-term reliance on AI may erode clinician skills, particularly for trainees and in low-prevalence contexts, citing historical failures of computer-aided diagnosis in mammography.[49] Human oversight is essential to avoid automation bias and patient harm, and ethical frameworks, bias mitigation techniques, and transparency measures must be pursued.[6] High-quality, interoperable data and explainable models are needed to enable human oversight.[4]

7.2. Organizational and Workflow Barriers

Successful AI deployment requires simultaneous attention to strategic, operational, and implementation dimensions. Organizations often fundamentally misunderstand AI deployment—treating it as a quick technological fix rather than a multi-year transformational process, focusing exclusively on financial returns while ignoring quality and safety benefits, and failing to recognize that AI implementation demands profound organizational change including workforce development, cultural shifts, and workflow redesign.[9]

Practical barriers include clinician training, interoperability with existing health information systems, workflow disruption, and change management.[50] Successful AI adoption requires intentional actions to effect behavioral change, including creating AI output visualizations that are easily interpretable by clinicians.[9] Many AI-based systems in critical care remain immature, particularly their lack of situational awareness and adaptability to individual patient responses. Practical issues such as digital literacy, device access, and usability for children, older adults, and other vulnerable populations also matter for applications requiring patient interaction.[4]

7.3. Adaptive AI Engagement Strategies

Effective AI use requires adaptive flexibility rather than a one-size-fits-all approach. Clinicians should dynamically shift between "centaur" strategies—where tasks are strategically divided with careful validation of AI output—and "cyborg" strategies—where human and AI work are more tightly integrated. High-stakes clinical decisions, particularly those involving unvalidated general-purpose AI tools, demand a centaur approach with rigorous human oversight and verification. Conversely, validated tools for specific purposes or lower-stakes administrative tasks may permit more integrated cyborg collaboration.

8. Case Studies

8.1. COPD Management

AI models predict exacerbations in COPD patients by analyzing wearable data and electronic health records. Implementation of an AI-driven clinical care pathway resulted in a 48% reduction in 30-day readmissions in high-risk patients.[28] Early interventions reduce hospitalizations and improve quality of life.

8.2. Ventilator Weaning in COVID-19

AI algorithms assist clinicians in determining optimal timing for ventilator weaning. In COVID-19 ICU patients, AI-assisted weaning reduced mean ventilation time by 181.8 hours, ICU length of stay by 7.7 days, and hospital length of stay by 16.9 days compared to controls.[24] These improvements demonstrate AI's potential in crisis management and critical care optimization.

8.3. COVID-19 Respiratory Care

During the pandemic, AI tools analyzed imaging and clinical data to triage patients, optimize ventilator use, and predict outcomes.[4,24] These applications highlighted AI's potential in crisis management, though systematic reviews revealed that most prediction models had high risk of bias and lacked external validation.[35]

8.4. Tuberculosis Screening

In resource-limited settings such as South Africa, AI-enhanced imaging tools improved TB detection rates, demonstrating AI's potential to expand access to care in low-income countries.[4] However, ensuring equitable access to AI tools globally is critical to avoid widening health disparities.

8.5. Closed-Loop Ventilation in Mixed ICU Populations

The ACTiVE trial protocol evaluated IntelliVent-ASV in diverse ICU populations, demonstrating that closed-loop ventilation modes maintain lung-protective strategies and reduce VILI. AI-based multiphase frameworks in emergency departments improved ventilator availability by 20% during respiratory disease surges.

9. Best Practices for Implementation

Successful AI integration in respiratory care requires:

Strategic Vision: Organizations must develop multi-year AI roadmaps that prioritize quality, safety, and equity alongside financial returns.[9]

Data Management: High-quality, interoperable data infrastructure is essential. Organizations often underestimate the importance of data governance, standardization, and privacy protection.[9,40]

Clinician Training: Targeted education on AI capabilities, limitations, and adaptive engagement strategies is critical to prevent automation bias and skill erosion.[49,6]

Equity Audits: Proactive bias detection through inclusive sampling, stratified performance reporting, and equity audits must be standard practice.[4,41,47,48]

Multicenter Validation: Rigorous external validation in diverse populations and clinical settings is required before widespread deployment.[4,38,39]

Human-in-the-Loop Design: AI should augment, not replace, clinicians. Explainable models and human oversight mechanisms are essential for patient safety.[4,6]

Regulatory Compliance: Adherence to FDA guidance and risk-based frameworks ensures patient safety and facilitates adoption.[40]

10. Future Directions

Personalized Therapy: AI-driven models will tailor interventions to genetic, environmental, and lifestyle factors, advancing precision medicine.[4,5,35]

Integration with Genomics: Combining AI with genomic data will identify individual risk profiles and therapeutic targets.[4,5,35]

Global Equity: Ensuring access to AI tools in low-income countries is critical to avoid widening health disparities. Federated learning and privacy-preserving approaches may enable collaborative research without data sharing.[35]

AI-Human Collaboration: AI will augment, not replace, clinicians, enhancing decision-making and patient care while preserving the clinician-patient relationship.[4,8]

Predictive Public Health: AI could forecast respiratory disease outbreaks, supporting preventive strategies at population level.[4]

Advanced Methodologies: Graph Neural Networks, Causal Inference, Federated Learning, Self-Supervised Learning, and Large Language Models promise enhanced data integration, privacy-preserving research, and autonomous decision support.[35]

Neural Control Interfaces: Early studies suggest AI systems could respond directly to neural signals of respiratory drive.[14]

Integration with EHRs: Linking ventilator AI with EHRs could enable holistic patient management, incorporating comorbidities and longitudinal data.[14]

11. Conclusion

AI in respiratory care represents a transformative shift, offering enhanced diagnosis, treatment, and patient outcomes when deployed with rigorous validation, equity audits, and human oversight. While quantitative evidence demonstrates meaningful improvements in specific applications—including a 48% reduction in COPD readmissions, 42.6% reduction in ventilation time, and significant improvements in weaning success—most studies remain retrospective and single-center. Algorithmic bias systematically underdiagnoses disease in vulnerable populations, and automation bias poses patient safety risks. Successful implementation requires multi-year organizational transformation, clinician training, interoperable data infrastructure, and adaptive engagement strategies. AI-driven closed-loop ventilation systems show particular promise in reducing VILI, improving patient-ventilator synchrony, and optimizing weaning protocols. The future of respiratory care lies in balancing technological innovation with human-centered practice, ensuring that AI's benefits are realized equitably across all patient populations.