Mechanical ventilation is a vital tool used in life-threatening scenarios, making it essential for clinicians to understand its clinical and technical details for safer use. Since the pulmonary mechanics of patients on mechanical ventilation can vary widely, AI can be a useful tool to provide personalised care for each patient and alert clinicians to plan liberation from ventilation.

For example, Peine A et al. (34) developed and evaluated a RL algorithm, VentAI, to suggest a personalised mechanical ventilation regime, focusing on finding the optimal settings for positive end-expiratory pressure, fraction of inspired oxygen, and ideal body weight-adjusted tidal volume. It was noted that the performance of VentAI was significantly superior compared to physicians' standard clinical care, with an observed outcome of in-hospital or 90-day mortality. Rehm GB et al. (35) developed an ML-based classifier to detect abnormal waveform events, specifically injurious subtypes of patient-ventilator asynchronies (double trigger asynchrony and breath stacking asynchrony). The tool accurately detected double trigger asynchrony with a sensitivity of 96% and specificity of 97.5%, and breath stacking asynchrony with a sensitivity of 94.4% and specificity of 98.7%, and other non-patient-ventilator asynchrony events with a sensitivity of 96.7% and specificity of 98%. Adams JY et al. (36) developed an open-source data acquisition system for capturing ventilator waveform data and a modular, multi-algorithm analytic tool (ventMAP) for automated identification of off-target ventilation (OTV) in critically ill patients. The research demonstrated that ventMAP effectively identified harmful effects of OTV, such as excessive tidal volumes and common patient-ventilator asynchronies. Correcting artefacts notably enhanced the specificity of event detection without compromising sensitivity. Sottile PD et al. (37) assessed several ML techniques, including random forest, naïve Bayes, and AdaBoost, on data collected from 62 mechanically ventilated patients who had or were at risk of acute respiratory distress syndrome (ARDS). They chose 116 features based on clinical insight and signal description and were able to determine the presence of synchronous breathing and three types of patient-ventilator asynchrony, including double triggering, flow limited, and ineffective triggering, with an AUC >0.890. Additionally, Loo NL et al. (38) trained a CNN with 5,500 abnormal and normal breathing cycles each, aiming to develop an algorithm to distinguish normal from abnormal breathing cycles, and reported a sensitivity of 96.9% and a specificity of 63.7% for their model.

Another study investigated the scope of AI in predicting prolonged mechanical ventilation and the need for tracheostomy based on severity-of-illness scores calculated from readily available patient variables. The authors concluded that ML predictive models were effective for predicting prolonged mechanical ventilation and tracheostomy placement, with AUCs of 0.82 and 0.83, respectively (39).

Giri J et al. (40) introduced a proposal for a low-cost mechanical ventilator that uses AI technology. The authors suggested a novel design to solve the shortage of ventilators, particularly in resource-limited settings. They used MatLab/Simulink simulation software to develop system-design platforms, employing Mathematical Models to perform simulations under various conditions and parameters. The authors ran computational fluid dynamics simulations to examine the dynamics of airflow, pressure distribution, and gas exchange within the device. These simulations aided in ventilator design and performance optimisation, ensuring the ventilator's efficiency in providing suitable breathing assistance. The authors concluded that the simulation approach using ML was comparable to conventional ventilators, capable of adapting to the patient's breathing patterns, optimising ventilation parameters, and enhancing patient outcomes. Such technological innovations with AI support can be crucial in underserved areas or during emergencies when ventilator demand spikes.

AI has recently demonstrated considerable potential in enhancing the weaning process by predicting the ideal time for weaning and identifying patients who would benefit from weaning. Predicting timely and uneventful liberation from mechanical ventilation has always been of prime interest to respiratory therapists and other intensive care unit (ICU) clinicians. A study that assessed two-stage (time-based) AI prediction models (the try-weaning stage and weaning stage) concluded its effectiveness and precision in predicting the optimal timing to wean intubated patients from the ventilator, with AUCs ranging from 0.843 to 0.953 for stage-1, and 0.889 to 0.944 for stage-2. This approach could reduce patient discomfort, improve medical quality, and lower costs, proving beneficial during pandemics like coronavirus disease 2019 (COVID-19) (41). Smart weaning, an AI-based weaning technique, uses algorithms to modify the ventilator's support according to the patient's respiratory needs. Smart weaning has been shown to enhance patient outcomes, shorten time on mechanical ventilation, and decrease the risk of complications, such as ventilator-associated pneumonia (VAP). Several studies have compared the effectiveness of smart weaning to traditional methods. One study showed that smart weaning was associated with a substantially shorter ventilator duration and lower VAP incidence in a randomised controlled trial of 127 critically ill patients (42). Another study on 240 patients with respiratory failure found that smart weaning had a higher success rate and required less time to perform than traditional methods (43). In the future, it is anticipated that.

AI will be crucial in assisting clinicians in managing ventilated patients. Dual, hybrid, and intelligent ventilator modes are already in place to enhance patient outcomes and lower complications by providing more individualised ventilation using advanced algorithms and patient monitoring. To offer personalized ventilation, dual-mode ventilation integrates inbuilt AI algorithms to combine distinct ventilation techniques, such as pressure support ventilation and volume-controlled ventilation into a single, adaptive system (44). Hybrid mode ventilation enables patients to breathe independently while still receiving ventilator support when necessary, by combining spontaneous breathing and required ventilation (45). Intelligent mode ventilation may reduce the risk of over- or underventilation by adjusting ventilator settings in real-time based on patient requirements and respiratory mechanics using advanced algorithms and monitoring (46). Some literature has reported the relevance of predictive AI models in various aspects of mechanical ventilation. For instance, one group created two models using gradient boosting regression with ICU data, outperforming existing standards with lower errors. Their findings suggested that high average heart rate, diagnosis of respiratory infection, and admissions from locations other than the operating theatre were associated with longer ventilation durations. For non-invasive ventilation, higher respiratory rates and lower Glasgow Coma Scale values were associated with longer durations (47). To err is human, and such errors should not occur with patients on life support devices like mechanical ventilator. Therefore, it is advisable to facilitate and accept technically approved algorithm-based decisions supported by AI programmes, rather than relying solely on human judgments, which are prone to subjective errors.

Extracorporeal membrane oxygenation (ECMO) is revolutionising critical care as a life-saving therapy for patients with severe, potentially reversible respiratory or cardiac failure (48). AI is now emerging as a transformative tool in ECMO management, aiding in decision-making and prognostication, and optimising clinical outcomes. One key challenge in ECMO management is determining the optimal timing for decannulation. Traditional methods rely heavily on clinical intuition and weaning trials. The Continuous Evaluation of Veno-venous ECMO Outcomes (CEVVO) is a DL-based model designed to predict decannulation success in patients supported on VV-ECMO (49). This model integrates discrete clinical information with continuous data from ECMO devices, employing a long short-term memory-based network to identify temporal patterns. The model categorises patients into risk groups, guiding clinicians on when to initiate weaning trials. The CEVVO model has shown superior performance compared to contemporary models (logistic regression, dense models, decision tree etc) in terms of risk stratification, decannulation outcome and clinical utility, with significantly higher successful decannulation rates in low-risk groups. In venoarterial ECMO (VA-ECMO), predicting patient outcomes is complex due to the high-risk nature of the therapy. The ECMO Predictive Algorithm (ECMO PAL) is the first AI-driven survival score developed using a large international patient cohort (50). Trained on data from the Extracorporeal Life Support Organisation registry, ECMO PAL demonstrated high sensitivity and precision in predicting in-hospital mortality. The model outperformed existing prognostication scores such as ’Survival After Veno-arterial ECMO' (SAVE) and Modified SAVE, highlighting its potential for broader clinical application. Similarly, another study developed an ML model to predict survival to discharge for patients on VA-ECMO, using data from the initial 48 h of support (51). This model, which outperformed the SAVE score, identified key variables, such as lactate, age, total bilirubin, and creatinine as significant predictors of survival, demonstrating the potential of AI to enhance clinical decision-making. Furthermore, AI's role extends beyond ECMO-specific models to broader applications in managing respiratory conditions that may require ECMO. A multimodal AI system was developed to predict the prognosis and required interventions for COVID-19 patients, incorporating clinical findings, laboratory data, and chest x-ray features (52). This system achieved higher predictive accuracy than single-modal models, demonstrating the value of combining different data types for comprehensive prognostication. Such approaches can inform decisions about the necessity of ECMO and other interventions, optimising resource use in critical care settings. Higher-volume ECMO centers have been associated with lower mortality rates, emphasising the importance of experience and expertise in ECMO management (53). AI models can further enhance outcomes in these settings by providing data-driven insights to support clinical decisions. Encouraging the use of AI in high-volume centers could lead to better standardisation and optimisation of ECMO practices. AI and ML are increasingly being explored for their potential to enhance extracorporeal cardiopulmonary resuscitation (ECPR) by optimising patient selection, predicting outcomes, and improving clinical decision-making. AI and ML techniques can analyse large volumes of structured and unstructured data to develop predictive models that identify which patients are most likely to benefit from ECPR. These technologies can also explore treatment heterogeneity, allowing for personalised treatment strategies based on individual patient characteristics (54). Studies have demonstrated that precise patient selection is crucial for achieving better neurological outcomes (55, 56). Additionally, AI can optimise the logistics of ECPR initiation. AI-driven systems can streamline preparation and setup processes, thereby reducing the time to cannulation and initiation of ECMO. Shortened setup times have been associated with improved rates of return of spontaneous circulation and overall survival (57). Moreover, AI can provide real-time monitoring and decision support during ECPR. By continuously analysing physiological parameters, AI systems can assist in adjusting ECMO settings and other supportive measures to optimise patient outcomes. This capability is particularly valuable in managing complex cases where rapid decision-making is critical (58). Finally, AI-driven simulation tools can enhance the training of medical personnel in ECPR procedures. These tools offer realistic scenarios and feedback, thereby improving the proficiency and readiness of the ECPR team (57). Despite the promising applications of AI and ML in ECPR, several challenges need to be addressed to ensure their effective integration into clinical practice.

High-quality, reliable data are essential for developing accurate ML models while using these tools in ECMO patients. Rigorous validation processes are necessary to ensure the reliability of these models. Furthermore, the potential risks of self-fulfilling forecast and feedback loops must be managed carefully.

AI clinical decision-support tools, though advanced, face implementation challenges due to limited datasets and imbalanced data, particularly in neonatal medicine. Despite these challenges, AI applications in neonatal respiratory care include tools for monitoring vital signs, predicting diseases like respiratory distress syndrome (RDS) and bronchopulmonary dysplasia (BPD), assessing risks such as retinopathy of prematurity, and providing diagnostic and prognostic support for neurological conditions like sleep stage classification.

RDS, common in premature infants due to pulmonary surfactant deficiency, affects over 80% of infants born before 29 weeks gestation (59). Infants with severe RDS are at higher risk of developing BPD (60). Traditional biomarkers like the lecithin/sphingomyelin ratio have limited bedside utility due to technical complexity. Ahmed W et al. (61) explored an ML algorithm to analyse other biomarkers in biological samples, offering a potential bedside tool for clinicians to guide interventions in preterm infants with RDS. Research in paediatric respiratory care includes risk stratification for BPD to identify infants who may benefit from preventive approaches like corticosteroids or patent ductus arteriosus management. The BPD Outcome Estimator, endorsed by the US National Institute of Child Health and Human Development, is widely used for guiding treatment decisions and counselling families (62). However, this tool's scope was limited to neonates of White, Black, or Hispanic descent. Patel M et al. (63) extended this ML algorithm to predict respiratory outcomes in extremely premature infants of Asian heritage, confirming its feasibility. Another study developed an ML algorithm to predict survival without BPD by retrospectively analysing perinatal factors and respiratory support in preterm infants. The combined model, using all perinatal features and 14 days of respiratory data, provided the best prediction with AUCs of 0.921 and 0.899 in the training and testing datasets, respectively. The model also suggested that extubating the patient to CPAP is superior to non-invasive positive pressure ventilation (NIPPV) in BPD-free survival (64). Apnoea of prematurity (AOP) affects over 50% of preterm babies and is nearly universal in infants with extremely low birth weight, linking it to brain damage and poor neurodevelopmental outcomes (65). A ML-based AOP detection model was developed to identify true apnoea in infants by analysing electrocardiographic data. With a false positive rate threshold of 0.3 in the mean receiver operating characteristic curve, the model achieved a high detection accuracy for all cardiorespiratory events (78.2%), and even higher for events followed by bradycardia (93.4%) and desaturation (95.2%). The study concluded that AI-based detection improved apnoea recognition with fewer false alarms than traditional methods (66). Correa M et al. studied 60 lung ultrasound (LUS) frames from 21 children under 5 in Peru to develop a DL algorithm to classify pneumonia. Using neural networks to analyse 1,450 pneumonia vectors and 1,605 normal lung vectors, the authors concluded that the DL model had 100% specificity and 91% sensitivity in diagnosing pneumonia in paediatric inpatients (67). Additionally, AI is used to analyse respiratory sounds in the paediatric population. Techniques like ANN and k-nearest neighbour are successful in recognising and categorising aberrant respiratory sounds (68). AI-based methods for wheeze identification in children and cough sound analysis also enhance diagnostic accuracy (69, 70).

While DL and newer AI algorithms hold significant potential in paediatric respiratory care, further research is needed to fully understand their capabilities. Medical professionals must be trained in AI tool usage, and regulatory procedures must ensure data security and confidentiality in AI applications.

AI-driven systems in pulmonary rehabilitation utilise wearable devices and sensors to monitor vital parameters like respiratory rate, oxygen saturation, and physical activity levels. These data are crucial for managing conditions such as COPD and interstitial lung disease. By applying ML algorithms, these systems can analyse real-time data to detect anomalies and predict exacerbations, enabling timely interventions and reducing hospital readmissions (71, 72). One study used ML to predict the effectiveness of pulmonary rehabilitation in convalescent COVID-19 patients, based on improved performance in the 6-minute walking test (6MWT). They classified 6MWT performance into three categories (low, medium, and high) and employed various algorithms, including random forest, adaptive boosting, and gradient boosting. Among 189 patients, random forest achieved the highest accuracy (83.7%), sensitivity (84.0%), and AUC (94.5%), while adaptive boosting had the highest specificity (92.7%). The study concluded that ML models are effective in predicting rehabilitation outcomes for convalescent COVID-19 patients (73). AI and ML are pivotal in personalising pulmonary rehabilitation by analysing patient-specific data, such as genetic information, environmental exposures, and health records. These algorithms can identify unique disease phenotypes and predict individual responses to rehabilitation. Personalised rehabilitation protocols can then be optimised to enhance effectiveness and minimise adverse effects. For example, ML models can tailor exercise regimens to match patients' physical capabilities and progression, ensuring safe and effective rehabilitation (74, 75). AI-powered virtual assistants and chatbots are revolutionising patient engagement in pulmonary rehabilitation. These tools utilise NLP technology to provide personalised education, medication and exercise reminders, and real-time symptom management. They also collect patient feedback to refine care plans and improve compliance (76, 77). Predictive analytics driven by AI and ML can forecast disease progression and identify high-risk patients. For instance, RNNs can analyse home spirometry data to predict lung function decline, enabling early intervention. This capability is essential for managing progressive diseases and improving patient outcomes (78).

In conclusion, AI, ML, and DL play a multifaceted role in pulmonary rehabilitation, including diagnostics, personalised treatment, patient engagement, and predictive analytics. These technologies enhance the accuracy and efficiency of pulmonary rehabilitation, empowering patients to actively manage their health and improving clinical outcomes and quality of life for those with chronic respiratory diseases (79).

AI is increasingly used to interpret and analyse CT and x-ray images, employing supervised learning to identify and track respiratory illnesses (80). Over the past two decades, LUS has become a crucial bedside imaging tool in acute care settings (81). AI-enhanced ultrasound machines can now perform complete lung scans in minutes (82). However, less experienced staff may find it challenging to gather footage for analysis, which can limit the utility of automated or AI-driven techniques. Nevertheless, with proper use, even novices can diagnose conditions such as pneumonia, pneumothorax, or effusions using this technology. Recent advances in ML have enabled automated photo recognition to match the accuracy of trained physicians (83). COVID-19 pneumonia often presents with B-line artifacts and pleural line irregularities on LUS, reflecting interstitial thickening and inflammation, which worsens as the disease progresses (84). LUS findings in the anterior lung fields, such as B-lines, consolidation, and pleural thickening, have been associated with the need for intubation in COVID-19 patients (85). Arntfield R et al. (86) developed a DL algorithm trained on a diverse dataset of LUS videos from patients with COVID-19 pneumonia, non-COVID ARDS, and hydrostatic pulmonary oedema, achieving high accuracy in distinguishing these pathologies. Born J et al. (87) introduced a CNN model (POCOVID-Net) to differentiate COVID-19 patients from those with bacterial pneumonia or healthy individuals, achieving an accuracy of 89% and a video accuracy of 92%. Kuroda Y et al. (88) evaluated an AI-based pneumonia detection method using a pocket-sized point-of-care ultrasound (AI-POCUS) device for COVID-19 pneumonia, achieving high sensitivity (92.3%) and specificity (100%). Another group developed a DL algorithm to quantify B-lines in LUS assessments, comparing it with expert human interpretations for both binary and severity classifications. The DL model achieved a sensitivity of 93% and a specificity of 96% for detecting B-lines compared to human experts (89). In a novel DL method called automated M-mode classification, the absence of lung sliding motion was detected with a balanced accuracy of 89%, sensitivity of 82%, and specificity of 92% (90). Another study developed and validated a binary classifier to distinguish between the presence and absence of lung sliding on LUS. Using a labelled dataset of 2,535 clips from 614 patients, the DL binary classifier achieved a sensitivity of 93.5%, specificity of 87.3%, and an AUC of 0.973 (91). Researchers also evaluated a DL algorithm using a spatial transformer network to detect and quantify pleural effusion based on images, finding improved accuracy compared to clinical standards (92).

AI integration in LUS holds great potential; however, robust applications require substantial high-quality data for training. The limited availability of LUS imaging data compared to other widely available imaging modalities poses a challenge to developing these applications.

AI is increasingly used to evaluate sleep quality, analysing heart rate, breathing rate, and body movements to provide clinicians with a comprehensive understanding of a patient's sleep, thereby aiding informed treatment decisions (93). AI has the potential to revolutionise sleep medicine by enabling early detection, personalised treatment, and better management of sleep disorders. In conditions like obstructive sleep apnoea (OSA), AI is expanding beyond the traditional apnoea-hypopnoea index for diagnosis (94). While OSA is usually diagnosed using polysomnography—which requires an overnight stay in a sleep lab or home testing—researchers are developing more accessible and affordable methods using AI. For example, smartwatches can monitor blood oxygen levels, providing a tool for early OSA detection (95, 96). AI-based predictive models have also shown promise in identifying individuals at risk of developing OSA using predictors like age, sex, and body mass index (97, 98). Furthermore, AI aids in diagnosing and subtyping narcolepsy by analysing clinical and polysomnographic data (99, 100). Li C et al. (101) proposed a DL model (EEGSNet) using multi-layer CNNs to extract features from EEG and bi-directional long short-term memory networks to classify sleep stages, outperforming other technologies, particularly for the challenging N1 stage. DL algorithms improve the precision and reliability of sleep studies, leading to more accurate diagnosis and treatment of sleep disorders. AI can also track sleep-related movement disorders such as REM sleep behaviour disorder, restless legs syndrome, and periodic limb movement disorder. Diagnosing narcolepsy is difficult due to the variability of multiple sleep latency tests, but an AI algorithm has shown promise in diagnosing type 1 narcolepsy in a single night with reasonable accuracy (102). Furthermore, Brennan HL and Kirby SD (103). in their review concluded with the utility of AI to individualise treatment regimens for sleep disorders. They underline the potential of AI to improve the treatment of OSA through predicting outcomes, and evaluating the treatment based on the disease process and physiology.

AI algorithms can identify patterns in patient data, such as sleep study results, medical history, and lifestyle factors, that may not be evident to physicians. However, clinicians should be aware of potential biases in these applications and the need for human supervision when designing sleep-related diagnostic and therapeutic plans.

AI has the potential to revolutionise pulmonary function evaluation, enhancing diagnostic precision by improving the interpretation of pulmonary function tests (PFTs) and eliminating inconsistencies in subjective test interpretations among pulmonologists (104). AI-based software has been successfully implemented for the computerised interpretation of PFTs, aiding in diagnosing lung diseases (105). Spirometry remains the most reliable method for detecting obstructive airway diseases, yet its underuse and misinterpretation have led to underdiagnosis (106). The rapid advancement of AI in medicine has improved PFT interpretations. A decision tree model that incorporates lung function and clinical variables has shown improved accuracy in detecting common lung diseases, such as COPD, asthma, interstitial lung disease, and neuromuscular disorders, compared to using PFT data alone (107). The study, which included 968 new patients in a pulmonary practice, demonstrated that the American Thoracic Society (ATS) and European Respiratory Society (ERS) algorithm correctly diagnosed 38% of patients, with COPD having the highest positive predictive value (74%). The decision tree algorithm doubled overall accuracy to 68% and improved positive predictive values for COPD (83%), asthma (66%), interstitial lung disease (52%), and neuromuscular disorders (100%). Another study using a neuro-fuzzy system with spirometric parameters and clinical symptoms achieved over 99% accuracy in classifying asthma and COPD (108). The neuro-fuzzy system integrates neurology-based learning with fuzzy logic's interpretability (e.g., If-Then). Topalovic M et al. (104) compared pulmonologists with AI software in identifying PFT patterns, finding that pulmonologists matched guidelines in about 75% of cases, with a correct diagnosis in 44.6 ± 8.7% of cases. In contrast, the AI software matched PFT pattern interpretations in 100% of cases, with an accurate diagnosis in 82%. Andrade D et al. (109) studied a diagnostic model using ML and the forced oscillation technique to detect respiratory changes in systemic sclerosis patients, concluding that ML algorithms combined with oscillation principles are effective in diagnosing respiratory changes. Another study by Amaral JL et al. (110) classified the severity of airway involvement in COPD using ML algorithms and forced oscillation measures, yielding promising results for precise diagnosis and monitoring. These studies underscore the potential of AI in improving the accuracy and efficiency of PFT interpretation and diagnosis.

AI can help reduce the workload for students and teachers while offering more engaging learning opportunities (111). It is expected to tailor learning experiences to individual students and incorporate active learning into education (112). AI can facilitate student learning through direct teaching (acting like a teacher), support teaching (collaborating with students), and empowering the learner (enabling group problem-solving with teacher feedback) (113–115). AI algorithms can analyse individual learning styles, progress, and strengths or weaknesses, creating personalised learning paths to optimise the educational experience. Liventsev V et al. (116) highlighted the importance of AI in healthcare education, particularly in patient simulators. Tools like “Auto-ALS” and GraphSim align with respiratory care curriculum objectives, using reinforcement learning to enhance training (116). AI virtual environments, such as SimX and Body Interact, are crucial in healthcare education, offering immersive, interactive scenarios for students to practise and refine their skills. These platforms provide dynamic, realistic patient interactions, essential for developing critical thinking and decision-making skills, which are vital in respiratory care. Through simulations, students encounter various clinical scenarios without risking real patients, enhancing clinical competence and confidence in handling real-life challenges (117). Adaptive e-learning systems like Firecracker, Smart Sparrow, and Coursera have significantly contributed to medical education by tailoring learning experiences to each student's needs. These platforms adjust content and pace based on individual performance, presenting more challenging cases to those excelling and providing extra resources to those struggling. This personalised approach is crucial in mastering the vast subject matter in medical studies. By continuously analysing performance, these systems optimise the learning journey in real-time, preventing students from feeling overwhelmed or under-challenged (118). In health professions education, AI-driven assessments have transformed student performance evaluations. Platforms such as ExamSoft and Gradescope offer precise, efficient evaluation methods, providing thorough and unbiased assessments. AI swiftly analyses outcomes objectively, delivering timely, precise feedback essential for students to identify strengths and address knowledge gaps, enhancing their learning journey (119). These technologies adapt simulation settings based on individual skill levels (120), ensuring a targeted, efficient educational journey and better translation of theoretical concepts into clinical proficiency (119). Health professions education must adapt to evolving technologies and their ethical implications. As AI brings significant changes, respiratory care professionals in academic and clinical contexts should receive proper training to employ these tools effectively (119, 121).