Dr. Jennifer Obi, MD
Founder, The Clinical AI Institute · Triple Board-Certified Physician
Pulmonary and critical care medicine sits at one of the most demanding intersections in all of clinical practice. The patients are acutely ill, the decisions are time-sensitive, the data streams are dense, and the margin for error is narrow. It is precisely this environment — high stakes, high volume, high complexity — that has made pulmonary and critical care one of the most active frontiers for artificial intelligence in medicine.
The applications being developed and deployed in this specialty are not incremental improvements to existing workflows. They represent a fundamental shift in how clinical information is gathered, interpreted, and acted upon. AI systems are now reading chest imaging with radiologist-level accuracy, predicting sepsis hours before clinical deterioration, stratifying ventilator management in real time, and flagging early signs of acute respiratory distress syndrome before they meet formal diagnostic criteria. The question for pulmonary and critical care physicians is no longer whether AI will change their practice. It already is. The question is how to engage with these tools as a clinician — critically, deliberately, and with the patient's safety as the governing standard.
The most mature AI applications in pulmonary medicine are in chest imaging. Deep learning models trained on hundreds of thousands of chest radiographs and CT scans have demonstrated performance that, in controlled studies, meets or exceeds that of experienced radiologists on specific detection tasks. The detection of pulmonary nodules on CT is perhaps the most well-documented example. AI systems can identify nodules as small as 3 millimeters with high sensitivity, flag incidental findings that might be missed on a busy read, and stratify nodules by malignancy risk using volumetric and morphological features that are difficult to assess visually.
For pulmonary embolism, AI-assisted CT pulmonary angiography interpretation has shown particular promise. Studies have demonstrated that AI systems can detect filling defects with sensitivity comparable to attending radiologists, and in some settings have reduced the time from image acquisition to preliminary read — a meaningful advantage when anticoagulation decisions are time-sensitive. Similarly, AI tools for interstitial lung disease pattern recognition are beginning to assist in the classification of fibrotic patterns on high-resolution CT, a task that has historically required subspecialty expertise and carries significant inter-reader variability.
What these imaging applications share is a common limitation that every pulmonary physician must understand: they are trained on datasets that may not reflect the patient population in a given institution. A model trained predominantly on CT scans from academic medical centers in one region may perform differently when deployed in a community hospital serving a different demographic. Performance metrics from validation studies are not guarantees of performance in local practice. Before any AI imaging tool is integrated into a clinical workflow, it must be validated on local data — and that validation must be stratified by patient characteristics to detect differential performance across subgroups.
Sepsis remains one of the leading causes of death in the intensive care unit, and its early recognition is one of the most consequential clinical challenges in critical care. AI-based early warning systems — trained on electronic health record data including vital signs, laboratory values, nursing assessments, and medication administration records — have been developed to identify patients at risk of sepsis hours before clinical deterioration.
The most widely studied of these systems, including the Epic Sepsis Model and various institutional variants, have demonstrated the ability to generate risk scores that precede clinical recognition of sepsis by two to six hours. In theory, this lead time creates an opportunity for earlier intervention — fluid resuscitation, blood cultures, antibiotic administration — that could reduce mortality. In practice, the clinical impact has been more complicated. A landmark study published in JAMA Internal Medicine found that the Epic Sepsis Model, when evaluated in a real-world health system, had a positive predictive value of approximately 12 percent — meaning that for every 100 alerts generated, roughly 88 were false positives. The alert burden this creates has measurable consequences: alert fatigue, workflow disruption, and the risk that clinicians begin to ignore warnings that carry genuine signal.
This is not an argument against AI-based sepsis prediction. It is an argument for the kind of rigorous, locally validated, physician-governed implementation that responsible AI adoption requires. The performance of any sepsis prediction model is highly dependent on the characteristics of the patient population, the completeness of the EHR data, and the clinical context in which the alert fires. Institutions that have achieved meaningful reductions in sepsis mortality with AI-assisted detection have typically done so through careful workflow integration — ensuring that alerts reach the right clinician at the right time, with enough contextual information to support rapid clinical judgment rather than replace it.
The management of mechanical ventilation in the ICU is one of the most technically demanding and consequential aspects of critical care practice. Ventilator settings must be continuously adjusted in response to changes in lung mechanics, oxygenation, and patient effort — a task that requires sustained attention and clinical expertise that is not always available at 3 AM on a busy unit. AI systems designed to assist with ventilator management represent one of the most clinically significant applications in critical care, and also one of the most ethically complex.
Closed-loop ventilation systems — in which an AI algorithm continuously adjusts tidal volume, respiratory rate, PEEP, and FiO₂ in response to real-time patient data — have been studied in clinical trials with promising results. These systems have demonstrated the ability to maintain patients within lung-protective ventilation targets more consistently than manual management, reduce the time to weaning readiness, and decrease the incidence of ventilator-induced lung injury in select populations. The ARDSnet protocols that transformed critical care practice in the early 2000s established that standardized, evidence-based ventilation targets reduce mortality — closed-loop AI systems represent the logical extension of that principle into real-time adaptive management.
The clinical and ethical questions these systems raise are not trivial. When an AI algorithm adjusts a ventilator setting on a critically ill patient, who is responsible for that decision? How does the bedside nurse or respiratory therapist know when to override the system? What happens when the algorithm encounters a clinical scenario outside its training distribution — a patient with an unusual chest wall compliance, a concurrent pneumothorax, or a rapidly evolving clinical picture? These are not hypothetical concerns. They are the questions that physician-led governance must answer before any closed-loop system is deployed in a clinical environment.
ARDS remains one of the most challenging diagnoses in critical care, in part because its clinical presentation overlaps substantially with other causes of acute hypoxemic respiratory failure, and in part because its severity and trajectory are difficult to predict at the time of diagnosis. AI applications in ARDS span the full clinical arc — from early detection to severity stratification to outcome prediction.
Early detection models, trained on chest imaging and clinical data, have demonstrated the ability to identify radiographic patterns consistent with ARDS before the syndrome meets Berlin Definition criteria — a window that could allow earlier initiation of lung-protective ventilation and prone positioning. Severity stratification models using a combination of imaging features, oxygenation parameters, and inflammatory biomarkers have shown promise in identifying patients most likely to benefit from specific interventions, including neuromuscular blockade and venovenous ECMO. Outcome prediction models have been developed to estimate mortality risk and duration of mechanical ventilation, information that is increasingly relevant to goals-of-care conversations with families.
Each of these applications carries the same fundamental requirement: the model must be validated in the population where it will be used, its outputs must be interpretable to the clinician receiving them, and there must be a clear governance structure that defines how the model's outputs are incorporated into clinical decision-making. An ARDS severity score generated by an AI system is not a diagnosis. It is a data point — one that must be integrated with the full clinical picture by a physician who understands both the patient and the limitations of the tool.
Beyond the ICU, AI applications are expanding across the broader scope of pulmonary medicine. In pulmonary function testing, machine learning models are being developed to improve the accuracy of spirometry interpretation, reduce the variability introduced by suboptimal patient effort, and detect early patterns of obstructive or restrictive disease that may not meet conventional diagnostic thresholds. For patients with COPD, AI-based risk stratification tools are being used to identify those at highest risk of exacerbation — enabling proactive outreach and medication optimization before a hospitalization occurs.
In sleep medicine, AI-powered polysomnography analysis has moved from research to clinical practice in several health systems. Automated sleep staging algorithms have demonstrated agreement with expert human scoring that meets or exceeds inter-rater reliability between human scorers, and they have the potential to significantly reduce the time and cost of sleep study interpretation. For a specialty facing a chronic shortage of board-certified sleep medicine physicians, this is not a marginal efficiency gain — it is a structural change in how care can be delivered.
The breadth and pace of AI development in pulmonary and critical care medicine creates a genuine obligation for physicians in this specialty. We cannot afford to be passive recipients of technology that others design and deploy in our clinical environments. The decisions made about which AI tools to adopt, how to integrate them into workflow, how to monitor their performance, and when to override or discontinue them are clinical decisions — and they require clinical leadership.
This means engaging with the evidence on AI tools with the same critical rigor we apply to any other intervention. A vendor's validation study is not sufficient evidence for clinical deployment. It is the beginning of an evaluation process that must include local validation, demographic performance analysis, workflow integration assessment, and ongoing monitoring. It means insisting on transparency — understanding what data a model uses, what outcome it predicts, and what its known failure modes are. And it means building the governance structures that ensure AI tools in our ICUs and pulmonary clinics are subject to the same accountability as any other clinical intervention.
The patients in our ICUs are among the most vulnerable in medicine. They cannot advocate for themselves when an algorithm generates a spurious alert, when a closed-loop system makes a suboptimal ventilator adjustment, or when an imaging AI misses a finding that a careful human reader would have caught. That advocacy is our responsibility — and it is one that no AI system can fulfill on our behalf.
The promise of artificial intelligence in pulmonary and critical care medicine is real. The tools being developed have the potential to reduce mortality, improve the consistency of evidence-based care, and extend the reach of subspecialty expertise to settings where it is not currently available. Realizing that promise requires physicians who are willing to engage with these tools not as passive users, but as clinical leaders — asking hard questions, demanding rigorous evidence, and holding AI to the same standard of accountability we hold ourselves.
Dr. Jennifer Obi, MD is a triple board-certified physician in Pulmonary Medicine, Critical Care Medicine, and Internal Medicine, and the Founder of The Clinical AI Institute. She advises health systems and physician organizations on the responsible implementation of artificial intelligence in clinical practice.
The Clinical AI Institute works with health systems, physician groups, and conference organizers to build the governance structures and clinical competencies that responsible AI adoption requires.
Be the first to join this conversation.
Physicians, attorneys, healthcare leaders — this is a conversation that matters. What is your line in the sand?
Continue Reading