Your privacy, your choice

We use essential cookies to make sure the site can function. We also use optional cookies for advertising, personalisation of content, usage analysis, and social media.

By accepting optional cookies, you consent to the processing of your personal data - including transfers to third parties. Some third parties are outside of the European Economic Area, with varying standards of data protection.

See our privacy policy for more information on the use of your personal data.

for further information and to change your choices.
Skip to main content
Download PDF

Subphenotyping prone position responders with machine learning

Critical Care volume 29, Article number: 116 (2025) Cite this article

Abstract

Background

Acute respiratory distress syndrome (ARDS) is a heterogeneous condition with varying response to prone positioning. We aimed to identify subphenotypes of ARDS patients undergoing prone positioning using machine learning and assess their association with mortality and response to prone positioning.

Methods

In this retrospective observational study, we enrolled 353 mechanically ventilated ARDS patients who underwent at least one prone positioning cycle. Unsupervised machine learning was used to identify subphenotypes based on respiratory mechanics, oxygenation parameters, and demographic variables collected in supine position. The primary outcome was 28-day mortality. Secondary outcomes included response to prone positioning in terms of respiratory system compliance, driving pressure, PaO2/FiO2 ratio, ventilatory ratio, and mechanical power.

Results

Three distinct subphenotypes were identified. Cluster 1 (22.9% of whole cohort) had a higher PaO2/FiO2 ratio and lower Positive End-Expiratory Pressure (PEEP). Cluster 2 (51.3%) had a higher proportion of COVID-19 patients, lower driving pressure, higher PEEP, and higher respiratory system compliance. Cluster 3 (25.8%) had a lower pH, higher PaCO2, and higher ventilatory ratio. Mortality differed significantly across clusters (p = 0.03), with Cluster 3 having the highest mortality (56%). There were no significant differences in the proportions of responders to prone positioning for any of the studied parameters. Transpulmonary pressure measurements in a subcohort did not improve subphenotype characterization.

Conclusions

Distinct ARDS subphenotypes with varying mortality were identified in patients undergoing prone positioning; however, predicting which patients benefited from this intervention based on available data was not possible. These findings underscore the need for continued efforts in phenotyping ARDS through multimodal data to better understand the heterogeneity of this population.

Background

Acute respiratory distress syndrome (ARDS) is a common and severe cause of admission to the intensive care unit (ICU) with a mortality rate of up to 45% [1, 2]. ARDS can be defined as a heterogeneous phenotype of clinical, biological, and radiological signs secondary to pulmonary or extrapulmonary aggression [2, 3]. Recent guidelines have stressed the importance of identifying subphenotypes based on data-driven assessment of multimodal traits [4]. Subphenotyping is a challenging task that requires collection of clinically relevant variables related to the pathophysiology of ARDS to provide personalized care [5] or prognostic enrichment. To personalize therapeutic strategies, biological and clinical hyper-inflammatory and hypo-inflammatory subphenotypes of ARDS have been described, which may benefit from different levels of positive end-expiratory pressure (PEEP) [6, 7] or fluid strategies [8], and can be identified early [9].

Prone positioning [10] is recommended for moderate-to-severe ARDS [4] and remains one of the most useful techniques for decreasing mortality in this population [11]. Its utilization has increased greatly during the coronavirus disease 2019 (COVID-19) pandemic [12]. However, apart from findings on prone positioning being most beneficial in a subgroup with a PaO2/FiO2 ratio under 150 mmHg, little is known about potential ARDS subphenotypes that might benefit the most or could even be harmed from the technique [13, 14]. While the LIVE trial [15] showed no significant mortality difference in the intention-to-treat analysis, its per-protocol findings suggest a potential benefit of personalized proning strategies, emphasizing the importance of tailored approaches in ARDS management.

Nonetheless, the question of which subphenotype might benefit most from prone positioning subsists, as patients without improvement in gas exchange may still benefit [16]. The morphological subphenotype of ARDS, assessed through chest radiography, ultrasound, or computed tomography, fails to predict improvement in oxygenation during proning [17]. Recent machine learning methods used on electronic health record data failed to predict the success of prone positioning during COVID-19 related ARDS, defined as an improvement of at least 10% in the PaO2/FiO2 ratio, ventilatory ratio, respiratory system compliance, or mechanical power [18]. Additionally, Hannon et al. [19] highlighted the potential of machine learning in predicting mortality after prone positioning despite poor model performance in terms of specificity. Furthermore, it is unknown whether transpulmonary pressure measurements can improve identification of subphenotypes. As a key element in the personalization of mechanical ventilation, esophageal pressure monitoring, which allows for the estimation of transpulmonary pressures [20,21,22] and transpulmonary mechanical power [23], can improve the assessment and quantification of the risk of lung injury [24].

In this study, we aimed to identify subphenotypes of patients undergoing prone positioning based on respiratory system mechanics, transpulmonary pressure measurements, and gas exchange parameters and assess their association with mortality and response to prone position.

Methods

Study population and data collection

In this retrospective observational study, we enrolled all consecutive critically ill adult patients receiving invasive mechanical ventilation for at least 24 h and at least one cycle of prone positioning between 2020 and 2022 at the Beth Israel Deaconess Medical Center in Boston, Massachusetts. Prone positioning was performed in accordance with a standardized institutional protocol (E-Table 1), which detailed criteria for initiation, duration, and discontinuation based on ARDS severity. PEEP levels were adjusted based on individual patient characteristics and clinical judgment rather than a standardized protocol. Esophageal manometry was used at the discretion of the attending clinicians, without strict protocol, to guide PEEP adjustments. These adjustments were based on achieving target transpulmonary pressures, with a goal expiratory transpulmonary pressure of -2 to + 2 cmH₂O and an inspiratory transpulmonary pressure ≤ 15 cmH₂O, as per the clinical team's judgment. Deidentified data were extracted from the electronic health records of the eligible patients. Respiratory parameters during the 12 h preceding the first cycle of proning, the period of proning, and 12 h after the first cycle of prone positioning were derived and changes between these periods calculated. If several measurements for a variable in a period were available, medians were calculated after cleaning for clinically implausible values for the respective parameter [25]. Patients with missing data for critical variables necessary to calculate mechanical power specifically respiratory rate (RR), tidal volume (Vt), peak inspiratory pressure (PIP), and positive end-expiratory pressure (PEEP), were excluded from analyses. Other missing covariates were deemed to be missing at random and imputed through a single imputation algorithm using a Factorial Analysis for Mixed Data (FAMD) model. This allows for imputation, considering similarities between both individuals and relationships between variables and further dimension reduction analysis [26]. This approach was employed to preserve the sample size and minimize potential biases that might emerge from excluding patients with incomplete data. During each period, mechanical power, driving pressure, and respiratory system compliance were computed for each patient. Mechanical power was calculated from the median ventilator parameters, using the following equation: mechanical power (J/min) = 0.098 × RR × Vt × (PIP – ½ (Pplat-PEEP)) [27]. The ventilatory ratio was computed as [minute ventilation (ml/min) × PaCO2 (mmHg)]/(predicted body weight × 100 × 37.5) [28].

Table 1 Supine position characteristics of the overall population and of the three clusters
Full size table

The transpulmonary driving pressure, lung compliance, chest wall compliance, and esophageal expiratory pressure were extracted for the subcohort of patients with available esophageal manometry measurements before, during, and after prone position, and were analyzed as part of the full cohort and as a standalone cohort.

Subphenotyping by clustering analysis

To identify unobserved clusters of patients, based on respiratory parameters, that may have differential prognosis and response to prone positioning, we used unsupervised machine learning.

We used factorial analysis of mixed data, an extension of principal component analysis suited to mixed (quantitative and qualitative) data. This approach allows to summarize the information by reducing the dimensionality of the data while retaining important information.

The hierarchical clustering method was then used to define three clusters and was built by computing the Euclidean distance between individuals and using the Ward criterion to minimize the variance at each step of the construction of the algorithm and homogenize the clusters [29] (E-Fig. 1).

Fig. 1
figure 1

Flowchart

Full size image

The clusters are described with a v-test or a Pearson’s chi-squared test when appropriate to assess the difference between the cluster and the overall population for the variables of interest.

Association of subphenotypes with mortality and prone positioning response

The primary outcome was 28-day mortality. Secondary outcomes included response to prone positioning, defined as any improvement during prone positioning compared to the supine position, specifically a change > 0 for respiratory system compliance and PaO2/FiO2 ratio, or a change < 0 for driving pressure, ventilatory ratio, and mechanical power. [18]. Mortality was assessed through the national death registry through hospital network if death occurred outside the hospital [30].

Mortality and the proportion of prone position responders were compared between clusters using Pearson’s chi-squared tests with Holm-Bonferroni correction for multiple testing. Threshold for significance was set to 0.05 unless stated otherwise. A Kaplan–Meier survival curve was used to evaluate the survival of the three clusters up until day 28 and a log-rank test was performed.

Analyses were repeated for the subcohort of patients who had complete data on esophageal manometry measurements.

As a sensitivity analysis, prone position responders were defined as those with an improvement of at least 10% in the main respiratory parameters studied. We also assessed the evolution of main respiratory parameters after returning to the supine position.

As a secondary sensitivity analysis, to account for patients with missing secondary outcome, for each secondary outcome and each patient, a weight was computed as the inverse of the probability of having missing data for the specific outcome and weighted Pearson’s chi-squared tests were computed.

Additionally, a post-hoc power analysis was performed to confirm that the study was adequately powered (80%) to detect a difference of at least 10% in responder rates, if such a difference existed.

This study was approved by the Institutional Review Board of Beth Israel Deaconess Medical Center (2024P000813), and the requirement for informed consent was waived. This study is compliant with the STROBE guidelines for reporting observational studies [31].

Data extraction was performed using Stata (version MP 16.0, StataCorp LLC, College Station, TX, USA) and analyses were performed using R Statistical Software (version 4.2.2, R Foundation for Statistical Computing, Vienna, Austria) with the FactomineR package [32].

Results

Study population

Four hundred and sixteen patients fulfilled inclusion criteria and 63 patients were excluded due to missing data on ventilator parameters needed to compute mechanical power (Fig. 1). Baseline characteristics of the 353 patients included in analyses are shown in Table 1. The median age was 62 years (Interquartile IQR: 52–69), patients were mostly male (60%), with a positive RT-PCR COVID-19 test (71%) and a median body mass index (BMI) of 31 (IQR: 26–36) kg/m2. Mean tidal volume in supine position was 6.1 (IQR: 5.8–6.6) ml/kg predicted body weight, with a median PEEP of 12 (IQR: 10–14) cmH2O. One hundred forty-nine patients died by day 28 (42.2%).

In our cohort, 248 (70.2%) of patients underwent only one session of prone positioning, while 105 (29.8%) received two or more sessions. Median duration of first session of proning was 16 (IQR: 6–22) hours, with a time to proning of 54 (IQR:18–128) hours.

Prone position responders were 41% for respiratory system compliance, 41% for driving pressure, 49% for ventilatory ratio and 42% for mechanical power. Seventy nine percent of patients had an improved PF ratio during prone position.

An esophageal balloon was placed and transpulmonary pressure measurement were available in 173 (49%) patients.

Factorial analysis of mixed data

In the factorial analysis of mixed data, the most important variables for the first component included ventilatory ratio, plateau pressure, respiratory rate, and for the second component respiratory system compliance, driving pressure, and positive end-expiratory pressure. The number of components to be retained for the factor analysis of mixed data was determined to be 17, based on the optimal number identified through the scree plot (E-Fig. 1).

The results of the visual analysis indicated that the optimal number of clusters was three, as a higher number of clusters was accompanied by only a minimal increase in the explained variance. The percentages of missing values imputed through the factorial analysis of mixed data are shown in E-Fig. 8.

Hierarchical ascending clustering

Eighty-one patients were assigned to Cluster 1 (22.9%), 181 to Cluster 2 (51.3%), and 91 to Cluster 3 (25.8%) (Table 1).

Regarding the description of the clusters, the most important qualitative variables in characterizing Cluster 1 were the COVID-19 status, with a lower proportion of positive COVID-19 status (46.9% vs. 71.3% in the overall population, p < 0.001) and male gender (72.8% vs. 60.0% in the overall population, p < 0.001). Cluster 2 was defined by a higher proportion of positive COVID-19 patients (84.5% vs 71.3% in overall population, p < 0.001) and Cluster 3 by a higher proportion of patients with restrictive pulmonary disease (24.2% vs 9.9%, p < 0.001).

For continuous variables, Cluster 1 was defined by a higher mean PaO2/FiO2 ratio (184 mmHg vs. 135 mmHg in the overall population) and a lower PEEP (8 cmH2O vs. 12 cmH2O in the overall population). Cluster 2 was characterized by a lower driving pressure (12 cmH2O vs. 13 cmH2O in overall population), higher PEEP (13 cmH2O vs. 12 cmH2O in the overall population) and a higher respiratory system compliance (36 ml/cmH2O vs. 33 ml/cmH2O).

Finally, Cluster 3 was characterized by a lower pH (7.26 vs 7.32), higher PaCO2 (61 mmHg vs. 51 mmHg), and higher ventilatory ratio (2.9 vs 2.2) (Fig. 2, E-Table 2–4).

Fig. 2
figure 2

Profiles of different phenotypes according to the most important variables used to determine clusters. Data were normalized to have a mean of 0 and a standard deviation of 1 across all phenotypes. When multiple values were present for a given variable, they were summarized by its median value. Abbreviations: VR ventilatory ratio; Vt tidal volume; BMI body mass index; Crs respiratory system compliance, PF PaO2/FiO2 ratio; DP driving pressure; pH arterial pH; PaCO2 arterial pressure in CO2; PEEP positive end-expiratory pressure; MP mechanical power; Pplat plateau pressure; RR respiratory rate

Full size image
Table 2 Clinical outcome and response rate to prone position. Responders are defined by a strictly positive increase or decrease in the parameter of interest
Full size table

Similarly, three clusters were defined for the esophageal pressure subcohort using transpulmonary pressure measurements, with similar covariates differentiating the clusters.

The results for the esophageal pressure subcohort are shown in the supplementary material (E-Table 5–6). Considering the data provided using esophageal pressure, the three clusters shared similar lung and chest wall mechanics characteristics when compared to those of the full population.

Primary and secondary outcomes

Only 28-day mortality was significantly different across the three clusters (p = 0.037), as shown in Table 2. The highest mortality was in cluster 3 (56%), the lowest mortality was in Cluster 2 (35.9%), and Cluster 1 was intermediate (40.7%). There were no significant differences in the proportions of compliance, driving pressure, PaO2/FiO2 ratio, ventilatory ratio, or mechanical power responders.

In the esophageal pressure subcohort, Cluster 3 also had a higher mortality rate (p = 0.04), but there were no significant differences in respiratory parameter responders, including transpulmonary pressure and lung compliance.

The trajectories of the primary respiratory parameters were plotted in supine, prone, and return-to-supine positions (E-Fig. 2–6).

For each cluster, there were no significant changes upon prone position for the ventilatory ratio, mechanical power, and driving pressure. Compared to Cluster 1 and 2, Cluster 3 had higher supine values for Cluster 3. This cluster also had a lower supine respiratory system compliance.

There were no significant changes in the esophageal pressure subcohort, with lower supine values for lung compliance and chest wall compliance in Cluster 3. Regarding outcomes, only mortality significantly differed among the three clusters. Cluster 3 had the highest mortality rate.

The Kaplan–Meier survival curves demonstrate distinct differences in survival probabilities over time among the study groups, with Cluster 3 showing significantly lower survival rate compared to Cluster 1 and 2 (E-Fig. 7).

For the definition of prone position responders as an improvement of at least 10% in the respiratory parameters studied, there were also no differences between the three clusters (E-Table 7).

Results were similar when accounting for patients with missing data for each of the secondary outcome (E-Table 8).

The results of the post-hoc power analysis (E-Table 9) confirmed that the study was well-powered, with an 80% likelihood of detecting a minimum 10% difference in responder rates, supporting the robustness of the findings. A sensitivity analysis with the pressure-controlled ventilation equation for mechanical power showed no significant difference between responders for the three clusters (E-Table 11).

Discussion

In this retrospective observational study, using a machine learning clustering algorithm we identified 3 subphenotypes of patients with ARDS based on respiratory mechanics, oxygenation parameters, and demographic variables in the supine position. The distinct subphenotypes, which are biologically and physiologically coherent and associated with varying mortality rates, displayed no differences in their response rates to proning for the parameters that were evaluated as part of this study. These parameters, which are often considered predictors of a better outcome, might not be related to the observed difference of mortality in our study. It has been reported that the improvement in oxygenation was not associated with a better outcome if the PaO2/FiO2 ratio was computed over two days after the prone position [33] and also failed to predict improved survival [34]. The benefits of prone positioning through ventilator-induced lung injury (VILI) prevention, lung recruitment, and homogenization of lung aeration [35] might not be captured through punctual bedside measurements, such as in this retrospective study. Transpulmonary pressure measurements did not improve the capacity to discriminate among these cluster subphenotypes of responders but increased the internal validity of the clusters by being reproducible [36]. Adding different sources of data, including longitudinal data or morphological data with lung ultrasound assessment [17], might provide more information and help derive more specific clusters.

Nonetheless, these clusters presented different prognoses and could be clinically relevant for prognostication, particularly Cluster 3. The patients in Cluster 3 presented with a higher disease severity, such as a higher SOFA score and lower PaO2/FiO2 ratio, but their main characteristic was the occurrence of a pH lower than 7.3 with high levels of PCO2 and a median respiratory rate of 30 breaths/min. Increasing the tidal volume in these patients might have been limited by the median plateau pressure of 29.5 cmH2O at 6 ml/kg of tidal volume per predicted body weight.

Cluster 3 might represent a subphenotype of patients with relevant pulmonary dead space, which causes hypercapnia and acidosis and could explain, in part, the worse outcome. Prone position may also increase the ventilation perfusion mismatch in persistent ARDS patients [37], which also presents the highest ventilatory ratio, a surrogate for dead space [28]. A high proportion of patients in our cohort with COVID-19 ARDS, which increased dead-space ventilation by provoking alteration of the pulmonary vascular bed, could explain those results [38].

This study has several limitations. First, there was a relatively low number of patients undergoing prone positioning prior to the COVID-19 pandemic. However, change in practice during the COVID-19 pandemic have led to a major increase in the use of the prone position in the world, also at our center. This resulted in a high number of patients being included in this study during the COVID-pandemic [12]. Second, the study was monocentric, and the practice of esophageal pressure measurements and prone position might hinder the generalizability of the clusters. Nonetheless, the use standardized proning protocol enhances the internal validity of our findings and offers a framework that can be replicated in other critical care settings.

Furthermore, regarding the generalizability of our results, the observed 42% mortality rate in our cohort is consistent with international studies, such as the LUNG SAFE [1] study and the ROSE [39] trial, which reported similar mortality rates among ARDS patients.

Third, because all patients underwent prone positioning, no heterogeneity in the strategy effect was studied.

Fourth, patients in cluster 1 and 2 had no significant differences in total respiratory compliance, tidal volume, and driving pressure. Distinction between the two clusters arises not from difference in ventilator settings or ARDS severity but from different risk factors at baseline, with higher Elixhauser score, age, and proportion of male patients in the cluster 1.

Fifth, our analysis was limited to the first proning session to ensure independent samples, a requirement for clustering; while this approach does not capture responses to subsequent sessions, it allows for robust identification of subphenotypes with potential clinical implications. The PROSEVA trial demonstrated that the response to prone positioning is not associated with patient outcomes, and prone positioning should be continued regardless of oxygenation improvement. Similarly, our findings should not suggest discontinuation of prone positioning based solely on these early responses to the first session of prone position.

Sixth, no radiologic or inflammatory biomarkers data was used to build the clusters. While our clustering analysis was based on routinely available clinical variables, future research incorporating anthropometric, biological, morphological, radiologic, and longitudinal data could enhance the precision and applicability of proned position patients subphenotyping.

Seventh, while mortality was chosen as the primary outcome due to its clinical relevance and simplicity, we also analyzed lung-oriented outcomes such as respiratory system compliance and mechanical power to provide a comprehensive understanding of the effects of prone positioning. Notably, the clustering analysis in our study was blinded to outcomes, being constructed solely from baseline variables and respiratory parameters during the supine position, ensuring it was independent of the choice of primary or secondary outcomes.

Conclusion

In this study, we successfully identified distinct subphenotypes of ARDS patients undergoing prone positioning using unsupervised machine learning based on respiratory mechanics, transpulmonary pressures, and gas exchange parameters. No association was observed between these subphenotypes and mortality benefit or response to prone positioning. These findings underscore the need for continued efforts in phenotyping ARDS through multimodal data to better understand the heterogeneity of this population. Until further evidence is available, prone positioning should remain a cornerstone of treatment for moderate-to-severe ARDS [4], as its broad mortality benefits have been well-established.

Availability of data and materials

Requests of qualified researchers trained in human subject research and confidentiality to access additional documents and the dataset may be sent to the corresponding author.

References

  1. Bellani G, Laffey JG, Pham T, Fan E, Brochard L, Esteban A, et al. Epidemiology, patterns of care, and mortality for patients with acute respiratory distress syndrome in intensive care units in 50 countries. JAMA. 2016;315:788–800.

    Article  CAS  PubMed  Google Scholar 

  2. Matthay MA, Arabi Y, Arroliga AC, Bernard G, Bersten AD, Brochard LJ, et al. A new global definition of acute respiratory distress syndrome. Am J Respir Crit Care Med. 2024;209:37–47.

    Article  PubMed  Google Scholar 

  3. Thompson BT, Chambers RC, Liu KD. Acute respiratory distress syndrome. N Engl J Med. 2017;377:562–72.

    Article  CAS  PubMed  Google Scholar 

  4. Grasselli G, Calfee CS, Camporota L, Poole D, Amato MBP, Antonelli M, et al. ESICM guidelines on acute respiratory distress syndrome: definition, phenotyping and respiratory support strategies. Intensive Care Med. 2023;49:727–59.

    Article  PubMed  PubMed Central  Google Scholar 

  5. Reddy K, Calfee CS, McAuley DF. Acute respiratory distress syndrome subphenotypes beyond the syndrome: a step toward treatable traits? Am J Respir Crit Care Med. 2021;203:1449–51.

    Article  PubMed  PubMed Central  Google Scholar 

  6. Calfee CS, Delucchi K, Parsons PE, Thompson BT, Ware LB, Matthay MA, et al. Subphenotypes in acute respiratory distress syndrome: latent class analysis of data from two randomised controlled trials. Lancet Respir Med. 2014;2:611–20.

    Article  PubMed  PubMed Central  Google Scholar 

  7. Maddali MV, Churpek M, Pham T, Rezoagli E, Zhuo H, Zhao W, et al. Validation and utility of ARDS subphenotypes identified by machine-learning models using clinical data: an observational, multicohort, retrospective analysis. Lancet Respir Med. 2022;10:367–77.

    Article  PubMed  PubMed Central  Google Scholar 

  8. Famous KR, Delucchi K, Ware LB, Kangelaris KN, Liu KD, Thompson BT, et al. Acute respiratory distress syndrome subphenotypes respond differently to randomized fluid management strategy. Am J Respir Crit Care Med. 2017;195:331–8.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  9. Redaelli S, Von Wedel D, Fosset M, Suleiman A, Chen G, Alingrin J, et al. Inflammatory subphenotypes in patients at risk of ARDS: evidence from the LIPS-A trial. Intensive Care Med. 2023;49:1499–507.

    Article  PubMed  Google Scholar 

  10. Guerin C, Reignier J, Richard J-C, Beuret P, Gacouin A, Boulain T, et al. Prone positioning in severe acute respiratory distress syndrome. N Engl J Med. 2013;368:2159–68.

    Article  CAS  PubMed  Google Scholar 

  11. Sud S, Fan E, Adhikari NKJ, Friedrich JO, Ferguson ND, Combes A, et al. Comparison of venovenous extracorporeal membrane oxygenation, prone position and supine mechanical ventilation for severely hypoxemic acute respiratory distress syndrome: a network meta-analysis. Intensive Care Med. 2024;50:1021–34.

    Article  CAS  PubMed  Google Scholar 

  12. Johnson SW, Garcia MA, Sisson EKQ, Sheldrick CR, Kumar VK, Boman K, et al. Hospital variation in management and outcomes of acute respiratory distress syndrome due to COVID-19. Crit Care Explor. 2022;4:e0638.

    Article  PubMed Central  Google Scholar 

  13. Matthay MA, Arabi YM, Siegel ER, Ware LB, Bos LDJ, Sinha P, et al. Phenotypes and personalized medicine in the acute respiratory distress syndrome. Intensive Care Med. 2020;46:2136–52.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  14. Guerin C, Albert RK, Beitler J, Gattinoni L, Jaber S, Marini JJ, et al. Prone position in ARDS patients: why, when, how and for whom. Intensive Care Med. 2020;46:2385–96.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  15. Constantin J-M, Jabaudon M, Lefrant J-Y, Jaber S, Quenot J-P, Langeron O, et al. Personalised mechanical ventilation tailored to lung morphology versus low positive end-expiratory pressure for patients with acute respiratory distress syndrome in France (the LIVE study): a multicentre, single-blind, randomised controlled trial. Lancet Respir Med. 2019;7:870–80.

    Article  PubMed  Google Scholar 

  16. Papazian L, Munshi L, Guerin C. Prone position in mechanically ventilated patients. Intensive Care Med. 2022;48:1062–5.

    Article  PubMed  PubMed Central  Google Scholar 

  17. Haddam M, Zieleskiewicz L, Perbet S, Baldovini A, Guervilly C, Arbelot C, et al. Lung ultrasonography for assessment of oxygenation response to prone position ventilation in ARDS. Intensive Care Med. 2016;42:1546–56.

    Article  PubMed  Google Scholar 

  18. Dam TA, Roggeveen LF, van Diggelen F, Fleuren LM, Jagesar AR, Otten M, et al. Predicting responders to prone positioning in mechanically ventilated patients with COVID-19 using machine learning. Ann Intensive Care. 2022;12:99.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  19. Hannon DM, Syed JDA, McNicholas B, Madden M, Laffey JG. The development of a C5.0 machine learning model in a limited data set to predict early mortality in patients with ARDS undergoing an initial session of prone positioning. Intensive Care Med Exp. 2024;12:103.

    Article  PubMed  PubMed Central  Google Scholar 

  20. Sarge T, Baedorf-Kassis E, Banner-Goodspeed V, Novack V, Loring SH, Gong MN, et al. Effect of esophageal pressure–guided positive end-expiratory pressure on survival from acute respiratory distress syndrome: a risk-based and mechanistic reanalysis of the EPVent-2 trial. Am J Respir Crit Care Med. 2021;204:1153–63.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  21. Goligher EC, Telias I, Sahetya SK, Baedorf-Kassis E, Patel BK, Yehya N, et al. Physiology is vital to precision medicine in acute respiratory distress syndrome and sepsis. Am J Respir Crit Care Med. 2022;206:14–6.

    Article  PubMed  PubMed Central  Google Scholar 

  22. Baedorf Kassis E, Talmor D. Clinical application of esophageal manometry: how I do it. Crit Care. 2021;25:6.

    Article  PubMed  PubMed Central  Google Scholar 

  23. Gattinoni L, Tonetti T, Cressoni M, Cadringher P, Herrmann P, Moerer O, et al. Ventilator-related causes of lung injury: the mechanical power. Intensive Care Med. 2016;42:1567–75.

    Article  CAS  PubMed  Google Scholar 

  24. Ball L, Talmor D, Pelosi P. Transpulmonary pressure monitoring in critically ill patients: pros and cons. Crit Care. 2024;28:177.

    Article  PubMed  PubMed Central  Google Scholar 

  25. von Wedel D, Redaelli S, Suleiman A, Wachtendorf LJ, Fosset M, Santer P, et al. Adjustments of ventilator parameters during operating room-to-ICU transition and 28-day mortality. Am J Respir Crit Care Med. 2024;209:553–62.

    Article  Google Scholar 

  26. Audigier V, Husson F, Josse J. A principal component method to impute missing values for mixed data. Adv Data Anal Classif. 2016;10:5–26.

    Article  Google Scholar 

  27. van Meenen DMP, Serpa Neto A, Paulus F, Merkies C, Schouten LR, Bos LD, et al. The predictive validity for mortality of the driving pressure and the mechanical power of ventilation. Intensive Care Med Exp. 2020;8:60.

    Article  PubMed  PubMed Central  Google Scholar 

  28. Sinha P, Calfee CS, Beitler JR, Soni N, Ho K, Matthay MA, et al. Physiologic analysis and clinical performance of the ventilatory ratio in acute respiratory distress syndrome. Am J Respir Crit Care Med. 2019;199:333–41.

    Article  PubMed  PubMed Central  Google Scholar 

  29. Husson F, Josse J, Pages J. Principal component methods-hierarchical clustering-partitional clustering: why would we need to choose for visualizing data. Appl Math Departm. 2010;17

  30. von Wedel D, Redaelli S, Suleiman A, Wachtendorf LJ, Fosset M, Santer P, et al. Adjustments of ventilator parameters during operating room–to–ICU transition and 28-day mortality. Am J Respir Crit Care Med. 2024;209:553–62.

    Article  Google Scholar 

  31. von Elm E, Altman DG, Egger M, Pocock SJ, Gøtzsche PC, Vandenbroucke JP, et al. The strengthening the reporting of observational studies in epidemiology (STROBE) statement: guidelines for reporting observational studies. Lancet. 2007;370:1453–7.

    Article  Google Scholar 

  32. Lê S, Josse J, Husson F. FactoMineR: an R package for multivariate analysis. J Stat Softw. 2008;18(25):1–8.

    Google Scholar 

  33. Lemasson S, Ayzac L, Girard R, Gaillard S, Pavaday K, Guerin C. Does gas exchange response to prone position predict mortality in hypoxemic acute respiratory failure? Intensive Care Med. 2006;32:1987–93.

    Article  PubMed  Google Scholar 

  34. Albert RK, Keniston A, Baboi L, Ayzac L, Guerin C. Prone position–induced improvement in gas exchange does not predict improved survival in the acute respiratory distress syndrome. Am J Respir Crit Care Med. 2014;189:494–6.

    Article  PubMed  Google Scholar 

  35. Guerin C, Baboi L, Richard JC. Mechanisms of the effects of prone positioning in acute respiratory distress syndrome. Intensive Care Med. 2014;40:1634–42.

    Article  CAS  PubMed  Google Scholar 

  36. Castela Forte J, Perner A, van der Horst ICC. The use of clustering algorithms in critical care research to unravel patient heterogeneity. Intensive Care Med. 2019;45:1025–8.

    Article  PubMed  Google Scholar 

  37. Yuan X, Zhao Z, Chao Y, Chen D, Chen H, Zhang R, et al. Effects of early versus delayed application of prone position on ventilation-perfusion mismatch in patients with acute respiratory distress syndrome: a prospective observational study. Crit Care. 2023;27:462.

    Article  PubMed  PubMed Central  Google Scholar 

  38. Bertelli M, Fusina F, Prezioso C, Cavallo E, Nencini N, Crisci S, et al. COVID-19 ARDS is characterized by increased dead space ventilation compared with Non-COVID ARDS. Respir Care. 2021;66:1406–15.

    Article  PubMed  Google Scholar 

  39. The National Heart, Lung, and Blood Institute PETAL Clinical Trials Network (2019) Early Neuromuscular Blockade in the Acute Respiratory Distress Syndrome. N Engl J Med. 380: 1997–2008.

  40. Vincent J-L, Moreno R, Takala J, Willatts S, De Mendonça A, Bruining H, et al. The SOFA (sepsis-related organ failure assessment) score to describe organ dysfunction/failure. Intensive Care Med. 1996;22:707–10.

    Article  CAS  PubMed  Google Scholar 

Download references

Acknowledgements

Maxime Fosset thanks SRLF (Société de Réanimation de Langue Française) and Fulbright France Foundation for their help. The authors thank Joris Pensier (Department of Anesthesia and Intensive Care Unit, Regional University Hospital of Montpellier, Montpellier, France) for his helpful advice on this study.

Funding

No financial support was specifically received for this study. Maxime Fosset has been funded through grants provided by the MUSE (Montpellier University School of Excellence Young Investigators PhD award 2022) foundation, SRLF (Société de Réanimation de Langue Française) and Fulbright France foundation. This work was publicly funded through ANR (the French National Research Agency) under the "Investissements d’avenir" program with the reference ANR-16-IDEX-0006»). All researchers, both external and internal, had full access to all of this study's data, statistical reports, and tables, and therefore take responsibility for the integrity of the data as well as the accuracy of data analysis and reporting.

Author information

Author notes

  1. Maxime Fosset, Dario von Wedel and Simone Redaelli have contributed equally to this work.

Authors and Affiliations

  1. Department of Anesthesia, Critical Care and Pain Medicine, Beth Israel Deaconess Medical Center, Harvard Medical School, Boston, MA, USA

    Maxime Fosset, Dario von Wedel, Simone Redaelli, Daniel Talmor, Maximilian S. Schaefer & Boris Jung

  2. Center for Anesthesia Research Excellence (CARE), Beth Israel Deaconess Medical Center, Harvard Medical School, Boston, MA, USA

    Maxime Fosset, Dario von Wedel, Simone Redaelli, Daniel Talmor, Maximilian S. Schaefer & Boris Jung

  3. Medical Intensive Care Unit and PhyMedExp, Lapeyronie Montpellier University Hospital, Lapeyronie Teaching Hospital, University Montpellier, 1; 371 Avenue Du Doyen Gaston Giraud, 34090, Montpellier, CEDEX 5, France

    Maxime Fosset & Boris Jung

  4. Desbrest Institute of Epidemiology and Public Health, University of Montpellier, INRIA, Montpellier, France

    Maxime Fosset, Nicolas Molinari & Julie Josse

  5. Institute of Medical Informatics, Charité-Universitätsmedizin Berlin, Berlin, Germany

    Dario von Wedel

  6. School of Medicine and Surgery, University of Milano-Bicocca, Milan, Italy

    Simone Redaelli

  7. Department of Anesthesiology, Perioperative and Pain Medicine, Lahey Hospital and Medical Center, Burlington, MA, USA

    Simone Redaelli

  8. Department of Pulmonary, Critical Care and Sleep Medicine, Beth Israel Deaconess Medical Center, Harvard Medical School, Boston, MA, USA

    Elias N. Baedorf-Kassis & Boris Jung

  9. Department of Anesthesiology, Duesseldorf University Hospital, Duesseldorf, Germany

    Maximilian S. Schaefer

Authors
  1. Maxime Fosset
    View author publications

    Search author on:PubMed Google Scholar

  2. Dario von Wedel
    View author publications

    Search author on:PubMed Google Scholar

  3. Simone Redaelli
    View author publications

    Search author on:PubMed Google Scholar

  4. Daniel Talmor
    View author publications

    Search author on:PubMed Google Scholar

  5. Nicolas Molinari
    View author publications

    Search author on:PubMed Google Scholar

  6. Julie Josse
    View author publications

    Search author on:PubMed Google Scholar

  7. Elias N. Baedorf-Kassis
    View author publications

    Search author on:PubMed Google Scholar

  8. Maximilian S. Schaefer
    View author publications

    Search author on:PubMed Google Scholar

  9. Boris Jung
    View author publications

    Search author on:PubMed Google Scholar

Contributions

MF, DvW, SR contributed equally to this work. MF, DvW, SR and BJ contributed to the conception and design of the study; to the analysis and interpretation of data; to drafting the submitted article and provided final approval of the version to be published. MF, DvW, SR contributed to data acquisition, drafting the submitted article, and provided final approval of the version to be published. MF, DvW, SR, NM, JJ contributed to data analysis, drafting the submitted article, and provided final approval of the version to be published. MF, DvW, SR, DT, ENBK, MSS, BJ contributed to data interpretation, drafting the submitted article, and provided final approval of the version to be published. All authors provided agreement to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Corresponding author

Correspondence to Boris Jung.

Ethics declarations

Ethical approval and consent to participate

This study was approved by the Institutional Review Board of Beth Israel Deaconess Medical Center (2024P000813), and the requirement for informed consent was waived.

Consent for publication

Not applicable.

Competing interests

M.S.S. received funding for investigator-initiated studies from Merck & Co., which do not pertain to this manuscript. He is an associate editor for BMC Anesthesiology. He received honoraria for lectures from Fisher & Paykel Healthcare and Mindray Medical International Limited. He received an unrestricted philanthropic grant from Jeffrey and Judith Buzen. E.N.B-K. has received lecturing fees from Hamilton Medical Inc. outside the submitted work and has received a KL2 award from Harvard Catalyst; The Harvard Clinical and Translational Science Center (National Center for Advancing Translational Sciences, National Institutes of Health award No. KL2 TR002542). The funders had no role in the design and conduct of the study, the collection, management, analysis, and interpretation of the data, the preparation, review, or approval of the manuscript; or the decision to submit the manuscript for publication. All authors declare no conflicts of interest related to this publication.

Additional information

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Information

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Fosset, M., von Wedel, D., Redaelli, S. et al. Subphenotyping prone position responders with machine learning. Crit Care 29, 116 (2025). https://doi.org/10.1186/s13054-025-05340-8

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doi.org/10.1186/s13054-025-05340-8

Keywords

Critical Care

ISSN: 1364-8535