Tracheostomy prediction model in neonatal bronchopulmonary dysplasia via lung and airway MRI

Abstract

Rationale:

Clinical management of neonatal bronchopulmonary dysplasia (BPD) is often imprecise and can vary widely between different institutions and providers, due to limited objective measurements of disease pathology severity. There is critical need to improve guidance on the application and timing of interventional treatments, such as tracheostomy.

Objectives:

To generate an imaging-based clinical tool for early identification of those patients with BPD who are likely to require later tracheostomy and long-term mechanical ventilation.

Methods:

We conducted a prospective cohort study of n=61 infants (55 BPD, 6 preterm non-BPD). MRI scores of lung parenchymal disease were used to create a binomial logistic regression model for predicting tracheostomy requirement. This model was further investigated using clinical variables and MRI-quantified tracheomalacia.

Measurements and Main Results:

A model for predicting tracheostomy requirement was created using MRI parenchymal score. This model had 89% accuracy, 100% positive predictive value, and 85% negative predictive value, compared with 84%, 60%, and 83%, respectively, when using only relevant clinical variables. In a subset of patients with airway MRI (n=36), a model including lung and tracheomalacia measurements had 83% accuracy, 92% positive predictive value, and 78% negative predictive value.

Conclusions:

MRI-based measurements of parenchymal disease and tracheomalacia can be used to predict need for tracheostomy in infants with BPD, more accurately than clinical factors alone. This prediction model has strong potential as a clinical tool for physicians and families for early determination of tracheostomy requirement.

Introduction

Bronchopulmonary dysplasia (BPD) is a severe and chronic pulmonary disease that occurs in premature infants. Severity levels are currently defined by extent of oxygen requirement, as determined by National Institute of Child Health and Human Development (NICHD)/National Heart Lung and Blood Institute (NHLBI) consensus.¹ However, there is much debate regarding the utility of this definition, with a clear need for better prognostic markers in early BPD that relate to relevant long-term clinical outcomes, such as the need for tracheostomy.^2-9

Tracheostomy can provide prolonged positive pressure or bypass segments of central airway obstruction to improve respiratory mechanics in patients with chronic respiratory failure and who struggle to wean from ventilatory support. Approximately 20% of infants with severe BPD will require tracheostomy (or die) prior to discharge.¹⁰ In extremely preterm infants, tracheostomy is associated with improved rate of weight gain and increased length, less daily sedation requirement, and improved ability to participate in developmental therapy.¹¹ Further, in a large BPD cohort cared for within the Neonatal Research Network, DeMauro et al. showed tracheostomy before 120 days of life was associated with lower odds of death or neurodevelopmental impairment compared to tracheostomy after 120 days of life.¹²

However, the decision to perform neonatal tracheostomy is typically not taken lightly by clinicians or families, due to risks associated with accidental decannulation, feeding difficulties, wound complications, speech-language development, and high burden on caregivers and the home healthcare system.^{13, 14} In addition, there is little consensus on the recommended timing of tracheostomy in infants, but unnecessary delay in those patients likely to benefit from tracheostomy may be detrimental to their long-term outcomes.^{15, 16} It remains very difficult for clinicians and families to discern whether an individual infant will benefit from tracheostomy, and, if so, at what time-point the procedure should be performed. While certain clinical data are known to be associated with need for tracheostomy, these factors are not sufficiently predictive to guide clinical decision-making, and in the absence of well-supported objective prediction tools, the implementation of tracheostomy is largely reliant on clinician experience.^{10, 17-19} Therefore, patients, clinicians, and families are likely to greatly benefit from a reliable tool that can accurately predict at an early age which patients will require prolonged mechanical ventilation, to guide decisions about need for tracheostomy.

Magnetic resonance imaging (MRI) is a promising modality in pediatrics, as it can achieve strong, regionally sensitive visualization of lung structures without requiring ionizing radiation, invasive procedures (e.g. intubation), intravenous contrast, or sedation/anesthesia.²⁰ Our research center has recently characterized lung parenchyma by ultrashort echo-time (UTE) MRI in neonates with structural visualization comparable to that of chest x-ray computed tomography (CT).²⁰ Using UTE MRI, we also found that a reader-based MRI score for BPD-related lung abnormalities was capable of predicting short-term clinical outcomes in preterm neonates, yielding a stronger correlation with duration and level of respiratory support at NICU discharge than any current standard clinical measure. ^21-23 Further, our center has previously developed novel methods for quantifying regional dynamic tracheal collapse (tracheomalacia, TM) in neonates, with good correlation between MRI and bronchoscopic assessments, and a strong relationship between clinically-diagnosed TM and increased work of breathing as measured by MRI-based computational fluid dynamics.^24-26 While we previously have correlated BPD-related lung disease from neonatal chest MRI with observational measures of respiratory support duration and level at discharge, neonatal respiratory MRI has not yet been applied in early identification of patients who are likely to require prolonged mechanical ventilation and therefore benefit from tracheostomy, which has strong potential for clinical impact.²³ Used in conjunction for the first time, these novel MRI techniques for assessing neonatal lung and airway abnormalities may provide a more comprehensive evaluation of BPD-related respiratory disease.

In the current study, we investigate a novel combination of lung parenchymal and airway disease quantification in preterm infants with BPD from cutting-edge respiratory neonatal MRI, which for the first time provides an imaging-based prediction model to determine the probability that BPD patients will require later tracheostomy and long term mechanical ventilation. This tool has unique potential to guide clinical decision-making on an individual patient basis and shorten the time to tracheostomy in infants who are most likely to benefit from this intervention. Further, this tool has the potential to reduce clinical practice variation and improve outcomes in these vulnerable infants who may face long-term respiratory impairment.

Methods

Study Subjects

Infants were recruited from the neonatal intensive care unit (NICU) at Cincinnati Children’s Hospital Medical Center with Institutional Review Board approval and informed parental consent. Infants were included if they had a history of prematurity (gestational age (GA) <37 weeks) and did not meet the following exclusion criteria: tracheostomy prior to MRI exam; suspected muscular dystrophy or other neuromuscular disorder; suspected neurologic disorder other than intra-ventricular hemorrhage, or significant congenital or genetic abnormalities that may affect lung development; evidence of any respiratory infection at time of imaging; and standard MRI exclusionary criteria. Infants were assigned to the BPD and no BPD groups if clinically diagnosed with or without BPD, respectively, using the NICHD/NHLBI criteria.¹

MRI Acquisition

All MRI exams were performed for research purposes on a neonatal-sized 1.5T MRI system within our institution’s NICU.^27-30 Patients were fed, swaddled, and equipped with ear protection before imaging. Subjects were tidal-breathing and nonsedated, unless respiratory support or sedation was part of ongoing NICU care. Contrast, sedation, and change in respiratory support were not implemented for imaging.

Two 3D axial MRI sequences were used for structural pulmonary imaging with full lung coverage: gradient-recalled echo (GRE) and radial UTE. ^{20, 31-33} GRE parameters were: echo time/repetition time = 1.9/7 ms; flip angle = 4°; field of view = 18 cm; pixel resolution = 0.70–0.86 mm; slice thickness = 3 mm; averages = 5–10; and duration = ~5 min. UTE parameters were: echo time/repetition time ≈ 0.2/5 ms; flip-angle = 5°; field of view = 18 cm; 3D resolution = 0.70–0.86 mm (equal in-plane resolution and slice thickness); duration = ~10–16 minutes; radial projections = ~120,000–200,000; and proton-density weighting. For some subjects, only GRE (n = 6) or only UTE images (n = 1) were acquired, due to subject agitation.

Clinicians were blinded to the research MRI results at the time of clinical tracheostomy decision. At our institution, infants less than 34 weeks PMA are supported with mechanical ventilation until the infant reaches extubation criteria, and then they are supported with CPAP until 32-34 weeks PMA. If infants are not stable with CPAP by one month of age, they usually receive steroids at that time. A second course of steroids may be used to facilitate extubation after an additional 1-2 months of mechanical ventilation support. For ventilation, our institution primarily utilizes volume modes of ventilation but will use pressure modes on a case-by-case basis. Tracheostomy is determined by failed extubation attempts and level of respiratory support in the setting of gestational age. Our institution has no specific criteria for tracheostomy and the decision is largely provider dependent.

MRI Scoring of Lung Disease

Lung disease on MRI was independently assessed by two radiologists (R.J.F. and A.H.S.) blinded to patients’ clinical status using a modified Ochiai scoring system that assesses seven categories, each scored 0-2, for a total range of 0-14 (hyperexpansion, mosaic lung attenuation, emphysema (number of cysts/regions), emphysema (size of cysts/regions), fibrous/interstitial triangular subpleural opacities, distortion of bronchovascular bundles, and subjective impression). ²¹ A mean MRI lung score was created by averaging total scores from each radiologist. The two readers in this study had good inter-reader reliability (r = 0.86, 95% confidence interval = 0.76-0.92) when using this identical scoring system in a prior study.²³

MRI Evaluation of Dynamic Tracheal Collapse

Using previously published methods for generating respiratory-gated airway UTE MRI in tidal-breathing neonates, the degree of tracheal collapse was quantified for a cohort subset (n = 36); early study subjects did not have full coverage of tracheal anatomy and were not included in airway analysis.^{25, 33} The degree of tracheal collapse (i.e. tracheal eccentricity) was defined as the ratio of the minimum-to-maximum tracheal lumen diameter at end-expiration, multiplied by 100 (range 0-100).²⁵ MRI tracheal eccentricity was separated into categories based on the most extreme eccentricity value in each patient: mild (>75), moderate (50-75), and severe (<50).

Statistical Analysis

Requirement for tracheostomy or death were combined as a single outcome (T/D), as mortality is a competing outcome with long-term complications such as severe BPD and tracheostomy. A variety of binomial logistic regression analyses were performed in SPSS (IBM Corp. Released 2010. IBM SPSS Statistics for Windows, Version 19.0. Armonk, NY: IBM Corp.) using several combinations of input variables, including clinical factors, MRI parenchymal score, and MRI tracheal eccentricity. First, a model was developed to determine the effect of MRI parenchymal score on the likelihood of T/D. Further models were explored using relevant clinical variables and MRI tracheal eccentricity.

Receiver operating characteristic (ROC) curve analysis was used to determine the area under the curve (AUC), a measure of the overall discriminatory ability of the model independent of a specific threshold value. Depending on the variable that was evaluated (clinical factors, MRI parenchymal score, or MRI tracheal eccentricity), ROC curves designated a different optimal cutoff; therefore, to allow comparison of different variations of the model, a clinically meaningful threshold of 75% probability was used to assign T/D for further testing of the specificity, sensitivity, positive predictive value (PPV), and negative predictive value (NPV) of the models.

For evaluation of the influence of gestational age at the time of MRI, chi-square goodness of fit test was used to determine if there were significant differences between outcomes predicted by the model and actual outcomes.

Some of the data from this work have been previously reported in other manuscripts with a distinctly unique aim.²³ Some of the results of this work have been previously reported in the form of abstracts.

Results

The study cohort was made up of 61 neonates for whom rMRI of the lung parenchyma was obtained. Within that larger cohort, 36 infants also received both MRI of lung parenchyma and trachea (Table 1 and Figure 1). Eighteen infants required tracheostomy and long term mechanical ventilation and an additional 3 infants did not survive. Among the surviving infants, none of the infants in room air at the time of the MRI required tracheostomy, while 21% (n=4) of the infants supported with nasal cannula, 33% (n=1) of the infants supported with CPAP and 68% (n=13) of the infants supported with invasive mechanical ventilation at the time of MRI went on to tracheostomy and long term mechanical ventilation.

Table 1:

Neonatal cohort demographics. DOL, day of life; MRI, magnetic resonance imaging; PMA, post-menstrual age; BPD, bronchopulmonary dysplasia.

N *
Total	61 (100%)
No-BPD	6 (9.8%)
Mild BPD	6 (9.8%)
Moderate BPD	11 (18.0%)
Severe BPD	38 (62%)
Male *	32 (52.5%)
GA (weeks) †	27.4 (23.6-36.1)
Multiparous *	12 (19.7%)
Birth weight (kg) †	1.0 (0.4-2.8)
Birth weight percentile (by GA) †	44.8 (0.1-99)
Small for Gestational Age *	10 (16.4%)
Sildenafil Therapy at discharge *	6 (9.8%)
DOL at MRI (days) †	91 (1-150)
PMA at MRI (weeks) †	40.4 (26.6-49.3)
Tracheostomy *	18 (29.5%)
DOL at Tracheostomy (days) †	137 (95-212)
PMA at Tracheostomy (weeks) †	45.9 (40.0-55.9)
Respiratory support at time of MRI *
Room Air	19 (31.1%)
Nasal Cannula	20 (32.8%)
CPAP	3 (4.9%)
Ventilator	19 (31.1%)
Deceased *	3 (5%)
MRI parenchymal score †
No-BPD	1.1 (0-4.0)
Mild BPD	1.3 (0-5.0)
Moderate BPD	4.2 (0.5-9.0)
Severe BPD	9.1 (1.5-13.5)

^*N (percent of total N)

^†Mean (range)

Figure 1: Respiratory MRI in Neonates with BPD.

Two patients with BPD underwent 3D ultrashort echo-time (UTE) MRI; the patient shown at top did not require later tracheostomy, while the patient shown at bottom did. Axial chest images are shown at left with MRI-based lung parenchymal scores of 2.5 (A) and 12.5 (B), out of a maximum 14 points. Airway images and 3D airway surfaces from UTE MRI of these patients are shown at right, demonstrating no (C) and severe (D, arrows) tracheomalacia.

MRI parenchymal score was modeled as a predictor of T/D in 61 infants. (Table 2). MRI parenchymal score was a significant variable (p < 0.0001) and each one unit increase in MRI parenchymal score had 1.77 times higher odds of T/D (ROC AUC 0.92). Using probability of 75% as the threshold for assigning T/D, the model correctly classified 88.5% of cases with 65.0% sensitivity. This model was presented graphically comparing probability of T/D versus MRI parenchymal score, in order to create a clinically useful tool (Figure 2).

Table 2:

Characteristics of a binary regression model to predict tracheostomy requirement. A: BPD patients with MRI-measured lung parenchymal scores and clinical variables (N = 61). AUC, area under receiver-operating characteristic curve; PPV, positive predictive value; NPV, negative predictive value; MRI, magnetic resonance imaging; GA, gestational age; SGA, small for gestational age. B: BPD patients with MRI-measured lung parenchymal scores and tracheomalacia (N = 36).

A: Model Characteristics for Patients with Lung Scores from MRI (N=61)
Overall Binary Regression Model								Individual Variables
Model	P	Area Under ROC	Accuracy (%)	Specificity (%)	Sensitivity (%)	PPV (%)	NPV (%)	Included	P
Clinical variables	<0.001	0.946	83.6	95.1	60.0	60.0	83.0	Non-singleton GA (weeks) SGA Sildenafil use Pneumonia	0.182 0.113 0.017 0.037 0.003
MRI lung	<0.001	0.921	88.5	100.0	65.0	100.0	85.4	MRI parenchymal score	<0.001
MRI lung + SGA	<0.001	0.928	88.5	100.0	65.0	100.0	85.4	MRI parenchymal score SGA	<0.001 0.106
MRI lung + pneumonia	<0.001	0.959	86.9	97.6	65.0	92.9	85.1	MRI parenchymal score Pneumonia	0.001 0.008
MRI lung + sildenafil	<0.001	0.928	86.9	100.0	60.0	100.0	83.7	MRI parenchymal score Sildenafil use	0.001 0.067
MRI lung + significant clinical variables	<0.001	0.968	90.2	97.6	75.0	93.8	88.9	MRI parenchymal score SGA Pneumonia Sildenafil use	0.006 0.092 0.010 0.266
B: Model Characteristics for Patients with Lung Scores and Tracheal Eccentricity Measurements from MRI (N=36)
MRI lung	<0.001	0.920	83.3	94.7	70.6	92.3	78.3	MRI parenchymal score	0.001
MRI airway	0.001	0.763	66.7	94.7	35.3	85.7	62.1	Tracheal eccentricity	0.019
MRI lung + MRI airway	<0.001	0.950	83.3	94.7	70.6	92.3	78.3	MRI parenchymal score MRI tracheal eccentricity	0.009 0.044

Figure 2: MRI Lung Parenchymal Score as a Predictor of T/D.

Binomial logistic regression (N = 61) was used to predict the percent probability of tracheostomy or death (T/D) for each MRI lung parenchymal score (range 0-14). The shaded area represents the threshold for patients’ assignment of T/D (75% probability).

Our previous report found that GA, birth weight, non-singleton birth, pulmonary arterial hypertension (PAH) requiring systemic pulmonary hypertension medications (i.e. sildenafil), and pneumonia during NICU stay (defined as worsening respiratory status and clinical chest radiograph changes requiring antibiotic therapy for a minimum of 5 days) were clinical variables with a statistically significant relationship to duration of respiratory support.²³ Other factors such as race, sex, IUGR, postnatal growth failure, patent ductus arteriosus, atrial septal defect, diuretics, antenatal steroids, postnatal steroids, surfactant, insurance status, and highest level of respiratory support in the NICU did not have a statistically significant relationship with duration of respiratory support level in our previous work, so these factors were not included in the current study.²³ Using the entire study cohort (N = 61), a new binary regression model was generated, using only the previously significant clinical variables, with some modifications: non-singleton birth, GA, small for gestational age (SGA, as a categorical substitute for birth weight that reduced the multicollinearity between birth weight and GA), pulmonary hypertension requiring sildenafil therapy (at any time in NICU course), and pneumonia during NICU stay (Table 2). Variables that significantly contributed to the model (p < 0.05) included SGA, sildenafil use, and pneumonia. The model based on clinical variables correctly classified 83.6% of cases and had 60.0% sensitivity.

To determine which clinical factors further enhanced the MRI parenchymal model, a model was generated using MRI parenchymal score and the three clinical factors that were significant contributors in the earlier model (SGA, sildenafil therapy, and pneumonia during hospitalization) (Table 2).²³ This model correctly identified 90.2% of cases with 75.0% sensitivity. Each factor was used also independently as a predictor of T/D for comparison (Figure 3).

Figure 3: MRI Lung Parenchymal Score and Clinical Variables as Predictors of T/D.

Binomial logistic regression (N = 61) using MRI lung parenchymal score (range 0-14) and a discrete clinical factor was used to predict probability of tracheostomy or death (T/D) for a given lung score. Presence and absence of each clinical factor are represented by green and gray markers, respectively: A) Small for gestational age (SGA), B) pneumonia during hospital stay, and C) sildenafil requirement. The shaded area represents the threshold for patients’ assignment of T/D (75% probability).

A model was further developed using MRI parenchymal score and MRI tracheal eccentricity as predictors of T/D (cohort subset of n = 36 with both lung and airway MRI). Both MRI parenchymal score and MRI tracheal eccentricity were statistically significant variables. The model was 83.3% accurate and had 70.6% sensitivity using a threshold of 75% (Table 2 and Figure 4).

Figure 4: MRI Lung Parenchymal Score and Tracheomalacia as Predictors of T/D.

Binomial logistic regression (N = 36) using MRI lung parenchymal score (range 0-14) and MRI-assessed tracheomalacia (as measured by airway eccentricity, range 0-100) to predict probability of tracheostomy or death (T/D). Airway eccentricity was separated into categories of mild (>75, green), moderate (50-75, yellow), and severe (<50, red).

To determine if MRI parenchymal score as a predictor of T/D was accurate in babies who underwent MRI before 40 weeks post-menstrual age (PMA), the cohort (N = 61) was divided into infants who underwent MRI at a PMA of <40 weeks versus ≥40 weeks. In the 27 infants who underwent MRI before 40 weeks PMA (average PMA of 38.0 weeks, range 26.6-39.9 weeks), the model indicated that the percentage of infants predicted to undergo T/D (14.8%) was not significantly different from the percentage that actually underwent T/D (18.5%) (p = 0.370, 1-sided). In the 34 infants who underwent MRI at or after 40 weeks PMA (average PMA of 42.2 weeks, range 40.1-49.7 weeks), the model indicated the percentage of infants predicted to undergo T/D (35.3%) was not significantly different from the percentage that actually underwent T/D (44.1%) (p = 0.184, 1-sided).

Discussion

Infants affected by BPD often require tracheostomy and long-term mechanical ventilation, but clinicians lack objective measures to guide patient selection and timing of tracheostomy. Previous studies have identified a correlation between clinical factors and need for tracheostomy. Indeed, we also found a correlation and were able to generate a tracheostomy prediction model using only clinical factors, and while this model had 84% accuracy, the clinical utility of this model is severely limited by the low sensitivity (60%), PPV (60%), and NPV (60%). These limitations are likely the reason why clinical factors have never been viewed as a reliable tool in clinical decision-making for tracheostomy.

MRI of neonatal lung parenchyma and airway is an objective, reproducible, and safe tool for measuring BPD respiratory disease. It has great potential in prediction modeling of outcomes and need for tracheostomy, providing a novel tool for clinical decision-making. In this study, we created a model to predict T/D using MRI parenchymal score. The model was created based on a cohort of 61 patients. Compared to the 84% accuracy and 60% PPV of the model based on clinical variables alone, predicting T/D based on MRI parenchymal score improves the accuracy and PPV of the model to 89% and 100%, respectively. Figures 2 and 3 demonstrate the simplicity of this model and its clinical utility as a visual tool for communication between clinicians and families, who can plainly see the probability of a patient requiring tracheostomy and long-term mechanical ventilation based on the MRI parenchymal score.

Patients with lower MRI parenchymal scores can be further stratified by assessing the probability of T/D in the presence of other clinical factors. By adding specific clinical factors to the model, the model becomes more sensitive and creates further subgroups of patients. Figure 3 demonstrates how the individual dichotomous clinical factors contribute to probability of T/D when included with MRI parenchymal score. These models are most useful for risk-stratifying infants with moderate MRI parenchymal scores. For example, an infant with an MRI parenchymal score of 8 out of 14 and who never required sildenafil has ~25% chance of T/D, but an infant with the same score who did require sildenafil has ~80% chance of T/D.

While many patients with BPD require tracheostomy because of parenchymal disease, tracheomalacia is also a contributor in some infants with BPD. The combination of MRI parenchymal score and MRI tracheal eccentricity as predictors of T/D shows great promise as a model with high sensitivity and accuracy. Infants with either severe parenchymal score or severe airway eccentricity have a high likelihood of T/D (Figure 4). Perhaps the greatest potential value of the combined parenchymal and airway model is useful for demonstrating the increased risk in patients who have both moderate parenchymal disease and moderate airway disease where each individual component alone may not be severe enough to warrant tracheostomy. These are the infants who traditionally would require a “wait-and-see” approach due to inability to predict outcomes from each individual disease component, but who would likely benefit from early tracheostomy as indicated by the MRI prediction model. A subset of our infants were supported with positive pressure ventilation at the time of the MRI. Although Hysinger et al. demonstrated that MRI under such conditions was very good at evaluating tracheal collapse (or lack thereof) when compared to bronchoscopy under similar conditions, it is possible that this respiratory support influenced the level of tracheal collapse that we measured in our study.²⁴

Given the historical difficulty in devising a definition of BPD that captures the heterogeneity of the disease, this model also demonstrates the value of the MRI parenchymal score as a continuous measurement of severity, which could be used as a measure of efficacy in clinical trials and as a longitudinal metric for tracking long-term disease progression. It is also conceivable that quantitative respiratory imaging may eventually be used in lieu of the categories of BPD severity, where the diagnosis of the severity of an infant’s chronic lung disease is based on objective and sensitive radiological findings.

Because this model may be used to determine tracheostomy need at an earlier age, it is important to ensure that the MRI parenchymal score is valid in younger infants. When we evaluated younger infants (less than 40 weeks PMA), the strength of the model predictions were not significantly different from the infants who received MRI at or after 40 weeks PMA. Therefore, the model is statistically reliable in infants with an average gestational age as young as 38 weeks PMA, which is well within the 120 day window suggested by DeMauro et al.¹² Although our cohort did include MRI evaluation of infants as young as 26 weeks PMA, the validity of our model at younger PMA requires further study. Some data suggest benefit to early tracheostomy, but ultimately further investigation is needed to evaluate the benefits and risks of early tracheostomy based on imaging.^{11, 12}

A shortcoming of this work is the limited sample size. Further, this is a single-center study and will need to be validated externally. Since our study population was biased towards the more severe patients, the tracheostomy rate was higher than rates reported in other studies; however, our study still maintained a substantial representation from all levels of BPD severity and indeed is reflective of the severity distribution in our institution’s NICU. This work is also limited by the radiologists' subjective reading of the MRI parenchymal score; however, inter-reader reliability has previously been reassuring, suggesting that this approach can be readily adapted by other institutions.²³ Finally, UTE MRI sequences are currently limited to research applications. However, all major MRI system vendors are currently developing product UTE sequences, making UTE MRI a realistic clinical tool in the near future. Of note, modalities such as CT are currently more accessible, and high resolution chest CT has the advantage of less susceptibility to motion than chest MRI, since the acquisition times are shorter. The lung parenchymal structure from proton-density MRI (such as our UTE MRI) is comparable to that of CT, with high 3D resolutions that approach CT resolution.^{20, 25} However, our chest MRI technique has clear advantages over chest CT in neonates with BPD: no ionizing radiation, performed during quiet tidal-breathing and on clinical supports without requiring sedation, intubation, or specific breathing maneuvers. As such, the information gained from minimal-risk serial evaluations via chest MRI may provide more value and reduce both risk and cost of care in the long term (e.g., sedated bronchoscopies, prolonged hospitalization, etc.), compared with an ionizing tomographic modality like chest CT.

In conclusion, quantitative MRI of the neonatal lung and airway can be used to generate clinically useful models for predicting T/D. This imaging-based prediction modeling is a novel and objective tool that significantly advances the current approach to tracheostomy decisions and importantly has strong potential as an early marker of long-term clinical outcomes.

Abbreviations

AUC

area under the curve

BPD

Bronchopulmonary dysplasia

GA

gestational age

NHLBI

National Heart Lung and Blood Institute

NICHD

National Institute of Child Health and Human Development

NICU

neonatal intensive care unit

NPV

negative predictive value

PAH

pulmonary arterial hypertension

PMA

post-menstrual age

PPV

positive predictive value

ROC

Receiver operating characteristic

SGA

small for gestational age

T/D

tracheostomy or death

TM

tracheomalacia (dynamic tracheal collapse)

UTE

ultrashort echo-time