Exploring the possibilities of machine learning in optimizing respiratory failure treatment

While machine learning (ML) has significant potential to predict and improve outcomes of acute respiratory failure by leveraging vast amounts of patient data, successful integration into clinical practice is conditioned on overcoming challenges related to data quality, system heterogeneity, clinician acceptance, and health equity. Critical Care.¹

With the global rise in acute respiratory failure (ARF), mechanical ventilation has been increasingly important but dangerous treatment.² In intensive care units (ICUs), where between 35% and 50% of patients need this support, ARF is associated with a high mortality rate of 67.2%. Treatment itself can be dangerous, with lung-related complications from mechanical ventilation contributing to 40% of hospital deaths.

Not only is invasive mechanical ventilation expensive, but ICU costs average $2,300 per day, increasing by over $3,900 after four days, but it also requires a major blow to patients and their families.³ To address this, researchers investigated whether ML could improve predictions of respiratory failure.¹ A panel of experts reviewed the existing literature and discussed ways to apply ML, and better predicted the onset and progression of this condition.

ML and data integration in clinical settings

Hospitals need to combine data from many sources to effectively manage patient information.¹ Integrating large-scale language models (LLMs) can be useful by incorporating unstructured data such as clinical notes, which improves predictive capabilities and clinical decision-making.

Both ML and deep learning are used in modern clinical decision support systems. Although the ML model has been applied to predict ARF during invasive ventilation, existing deep learning models still have challenges. Important barriers to implementation include integrating these analytical platforms into electronic health records (EHRS), gaining clinician acceptance, and maintaining the accuracy of the model across a wide range of patient populations and clinical practices.

Clinical utility of ML for respiratory outcomes

During the discussion of critical respiratory outcomes, panelists reached a consensus that predicted the most practical purpose of predicting the appearance and progression of respiratory failure. The requirements for invasive mechanical ventilation were also identified as a key consequence of interest. The main priority was to highlight the important impact of patients having early knowledge of possible progression of respiratory failure. Such sophisticated warnings provide clinicians with sufficient lead times to collect and interpret targeted diagnoses, thereby enabling potential interventions to avoid clinical degradation.

“There was debate about the length of the horizon for optimal prediction, but most people agreed that this would be a useful window for implementing a preventive strategy 12-24 hours before the onset of respiratory failure,” the study authors found.

Effective clinical collaboration and setting appropriate risk thresholds are essential for good decision-making. Notifying clinicians to significant thresholds can avoid alarm fatigue and excessive interpretation. This approach allows predicting patient response to interventions and informs both clinical decisions and trial design. The most valuable strategy the panelists agreed to is to predict and track patient progression through different levels of respiratory support.

Barriers to ML model development and implementation

The deep learning model faces challenges at both the patient level and the system level. Patient-level difficulties arise from a variety of underlying pathologies, types of respiratory failure, and treatment approaches that complicate model development. For ML models, source inconsistencies, format, frequency, and data quality can compromise accuracy. Furthermore, emergency procedures such as intubation are often not accurately recorded or time-engraved in the EHR, making it difficult to properly classify events and ensure model accuracy.

Substantial heterogeneity between healthcare systems, local practice patterns, and resource availability complicates the development of generalizable ML models. Successful integration and deployment conditions include a secure, efficient and appropriate interface between EHR and the analytics platform.

Physicians' hesitations regarding the deployment of AI models in ICU environments are an important barrier. This can be alleviated by promoting transparency and avoiding the “black box” model. As a result, a robust system for continuous monitoring and iterative improvement of model performance is a necessary component for successful implementations.

Strategies for successful ML integration and verification

A successful ML model should not only work well at the beginning, but also maintain effectiveness across diverse systems and patient groups. To prepare for a full clinical deployment, strategies need to be implemented that will increase predictive capabilities, detect bugs, and assess false positives and negatives. The panel emphasized that an important step in assessing the effectiveness of the model is to compare performance in a clinical setting with existing standard of care. To gain wider acceptance, panelists suggested that future trial designs should prioritize future multicenter studies.

Overcoming AI bias and resource gaps for healthcare

An important concern about technological advances in healthcare is the potential to disproportionately affect socioeconomically disadvantaged patients with limited access to support resources and infrastructure. Nevertheless, if ML algorithms are well designed, they could advance health equity rather than exacerbate existing disparities.

Additional considerations from a health equity perspective include the risk that low-resource health systems will delay adoption of these new technologies, thereby increasing health disparities. Furthermore, the available data may be inherently biased. Natural language processing and LLMS help analyse progress, but implicit biases also manifest through subtle language patterns. As a result, it is essential to address these health disparities during the development of any ML model to ensure fair performance and mitigate bias.

“Enhanced predictive capabilities through ML can promote a more proactive approach to patient care and potentially improve outcomes. However, many challenges need to be addressed to achieve meaningful integration into clinical care,” the study authors concluded.

reference

1. Can you predict the future of Pearce AK, Nemati S, Goligher EC, and other respiratory failure predictions? Critical Care. 2025; 29 (1). doi:10.1186/s13054-025-05484-7
2. Demem K, Tesfahun E, Nigussie F, Shibabaw AT, Ayenew T, Messelu MA. Time and predictors of adult patients undergoing mechanical ventilation in the intensive care unit of West Amhara Comprehensive Specialist Hospital, Ethiopia: Ethiopia: A retrospective follow-up study. BMC anesthesia. 2024; 24 (1). doi:10.1186/s12871-024-02495-9
3. Remove ventilators from patients faster: Facilitator guide. Healthcare research and quality agency. February 2017. Accessed on August 8, 2025. https://www.ahrq.gov/hai/tools/mvp/modules/vae/overview-off-ventilator-fac-guide.html