Machine Learning Successfully Identifies Subphenotypes of Pediatric ARDS

While subphenotyping has been investigated in adults using electronic health record (EHR) data,^1,2 a retrospective exploratory analysis published in the May issue of Pediatric Critical Care Medicine is the first to investigate its use in children.³

“There are a lot of data in the EHR that could be used to identify these biomarker-based subphenotypes more cost effectively and much more quickly,” said Daniel Balcarcel, MD, a pediatric critical care specialist in the pediatric intensive care unit (PICU) at Children’s Hospital of Philadelphia (CHOP). Dr. Balcarcel was first author of the article.³

During his pediatric critical care fellowship at CHOP, Dr. Balcarcel said he struggled with being unable to provide targeted treatment for children with PARDS. “Children with PARDS are among the sickest we care for, yet management remains almost entirely supportive, which is what drove me to study this condition,” he said.

Respiratory failure is one of the most common causes of death for children admitted to PICUs.⁴ PARDS, a rare acute inflammatory lung injury with lung edema, is estimated to be responsible for about 1% to 10% of admissions.⁵ Mortality rates are about 24%.⁶ PARDS’ heterogeneity makes developing targeted therapies challenging. Until pediatric-specific consensus conference guidelines were published in 2015, treatments were based on adult ARDS therapies. The pediatric guidelines were updated in 2023.⁷

“The criteria for PARDS are quite broad,” Dr. Balcarcel said. “We end up grouping together infants and older children with vastly different comorbidities, risk factors, and underlying pathophysiology. As a result, interventions we try may help some patients, harm others, and do nothing for the rest.”

The Findings
Previous researchers have discovered two ARDS subphenotypes—hyperinflammatory and hypoinflammatory—that appear to predict different responses to different treatments.^8,9 In adult trials, the hyperinflammatory subphenotype responded better to simvastatin, high positive end-expiratory pressure, and liberal fluid management.^10-12

Dr. Balcarcel wanted to combine that knowledge with his research interest in biomedical informatics to investigate whether machine learning could help lead to improved treatment for children with PARDS. As machine learning tools are improving workflows in various industries, Dr. Balcarcel realized there was a gap in healthcare and ICUs. “I thought it would be beneficial to incorporate some of these newer, advanced techniques into the research and the care that we’re providing in the ICU,” he said. 

Phenotyping and subphenotyping PARDS is not currently happening clinically, but research has shown that blood-based biomarkers can identify subphenotypes.^8,9 But biomarker testing at the bedside is not currently practical, so Dr. Balcarcel and the research team sought to determine whether an EHR-based classifier could accurately identify biomarker-based PARDS subphenotypes using only routinely collected clinical data within 24 hours of diagnosis. 

The team performed retrospective exploratory analyses of two cohort studies managed at CHOP. They used the open-source XGBoost algorithm, which performed well in studies of adult ARDS. The model was trained on a 2014-2019 cohort of children aged one month to 18 years who were admitted to the PICU, required mechanical ventilation, and met the 2012 Berlin Criteria for ARDS in adult patients. The validation cohort included patients between 2019 and 2022, had the same parameters as the 2014-2019 cohort, and met the PARDS criteria of the 2015 Pediatric Acute Lung Injury Consensus Conference guidelines.

“One of the advantages of a machine learning approach is that you can include a large number of variables and let the model identify which ones are most predictive,” Dr. Balcarcel said. “We went into the project a little agnostic as to what variables would be most predictive of the inflammatory subphenotypes.”

The team gave the model more than 100 variables and selected the top five as the most predictive, all of which are routinely collected for laboratory testing. They are:
 

1. Mean INR
2. Minimum lactate level
3. Minimum platelet count
4. Minimum absolute neutrophil count
5. Mean conjugated bilirubin level

“I was surprised by how few variables were actually needed to make accurate predictions. That is one of the advantages of machine learning—you can put in over 100 variables, identify the five most predictive, and then have a very simple, streamlined model that could be more easily implemented by clinicians,” Dr. Balcarcel said. Both the full model and the sim¬plified five-variable model performed well when tested on the validation cohort. 

The study found that INR and lactate were consistently among the top predictive variables for PARDS subphenotypes. These two laboratory results have not yet been used in ARDS subphenotype classification models. The study also found that the top variables were unrelated to respiratory failure (respiratory rate was highest at 24th).

Study limitations included lack of validation of the model outside the CHOP PICU, the possibility of misclassification of PARDS because the current analysis was unplanned and secondary, the possibility that model performance may be case-mix dependent, and smaller derivation and validation cohorts than those used in adult ARDS research. 

What’s Next 
Future studies should include multiple institutions and diverse cohorts, assessment of other models, and analysis of whether specific treatments differ across subphenotypes based on the models. 

These inflammatory subphenotypes have also been identified in adults with sepsis and respiratory failure without ARDS, Dr. Balcarcel said. “We’re starting to see these inflammatory subphenotypes emerge across a growing number of critical illness syndromes,” he said. “They may not be unique to PARDS. It’s possible that children with sepsis, traumatic brain injury, or post-cardiac arrest syndrome in the ICU also exhibit similar inflammatory patterns.”

The goal is to provide better healthcare for children, Dr. Balcarcel said. “Until we’re able to better define critical illness in pediatrics, we won’t be able to find targeted therapies. One possible next step is to study these inflammatory subphenotypes, see what kind of patients in our ICU have them, and then, from there, start to look into what therapies could be beneficial to one phenotype over the other.”

References

1. Sinha P, Churpek MM, Calfee CS. Machine learning classifier models can identify acute respiratory distress syndrome phenotypes using readily available clinical data. Am J Respir Crit Care Med. 2020 Oct 1;202(7):996-1004.
2. Maddali MV, Churpek M, Pham T, et al; LUNG SAFE Investigators and the ESICM Trials Group. Validation and utility of ARDS subphenotypes identified by machine-learning models using clinical data: an observational, multicohort, retrospective analysis. Lancet Respir Med. 2022 Apr;10(4):367-377.
3. Balcarcel DR, Mai MV, Mehta SD, et al. Development and validation of an electronic health record-based, pediatric acute respiratory distress syndrome subphenotype classifier model. Pediatr Crit Care Med. 2025 May;26(5):e611-e621.
4. Agra-Tuñas C, Rodriguez-Ruiz E, Rodríguez Merino E; MOdos de Morir en UCI Pediátrica-2 (MOMUCIP-2) study group of the Spanish Society of Paediatric Intensive Care (SECIP). How do children die in PICUs nowadays? A multicenter study from Spain. Pediatr Crit Care Med. 2020 Sep;21(9):e610-e616.
5. Schouten LRA, Veltkamp F, Bos AP, et al. Incidence and mortality of acute respiratory distress syndrome in children: a systematic review and meta-analysis. Crit Care Med. 2016 Apr;44(4):819-829.
6. Orloff KE, Turner DA, Rehder KJ. The current state of pediatric acute respiratory distress syndrome. Pediatr Allergy Immunol Pulmonol. 2019 Jun 1;32(2):35-44.
7. Emeriaud G, López-Fernández YM, Iyer NP, et al; Second Pediatric Acute Lung Injury Consensus Conference (PALICC-2) Group on behalf of the Pediatric Acute Lung Injury and Sepsis Investigators (PALISI) Network. Executive summary of the second international guidelines for the diagnosis and management of pediatric acute respiratory distress syndrome (PALICC-2). Pediatr Crit Care Med. 2023 Feb 1;24(2):143-168.
8. Yehya N, Zinter MS, Thompson JM, et al. Identification of molecular subphenotypes in two cohorts of paediatric ARDS. Thorax. 2024 Jan 18;79(2):128-134.
9. Dahmer MK, Yang G, Zhang M, et al; RESTORE and BALI study investigators; Pediatric Acute Lung Injury and Sepsis Investigators (PALISI) Network. Identification of phenotypes in paediatric patients with acute respiratory distress syndrome: a latent class analysis.
10. Calfee CS, Delucchi K, Parsons PE, Thompson BT, Ware LB, Matthay MA; NHLBI ARDS Network. Subphenotypes in acute respiratory distress syndrome: latent class analysis of data from two randomised controlled trials. Lancet Respir Med. 2014 Aug;2(8):611-620.
11. Famous KR, Delucchi K, Ware LB, et al; ARDS Network. Acute respiratory distress syndrome subphenotypes respond differently to randomized fluid management strategy. Am J Respir Crit Care Med. 2017 Feb 1;195(3):331-338.
12. Calfee CS, Delucchi K, Sinha P, et al; Irish Critical Care Trials Group. Acute respiratory distress syndrome subphenotypes and differential response to simvastatin: secondary analysis of a randomised controlled trial. Lancet Respir Med. 2018 Sep;6(9):691-698.