Predicting physiologic response to changes in positive end-expiratory pressure in mechanically ventilated children: a computable phenotype and machine learning approach

Positive end-expiratory pressure (PEEP) is often increased to improve ventilation efficiency and gas exchange during pediatric mechanical ventilation. Although it is clinically important to optimize PEEP in this population, there is a paucity of literature to guide the clinician at the bedside. Increasingly, time-series physiologic data are available for mechanically ventilated subjects in the intensive care unit. However, these data have not been adequately explored in the literature. Therefore, we sought to apply time-series computable phenotyping on time-series physiologic data and develop a model to predict PEEP response in mechanically ventilated children. We conducted a retrospective analysis of continuous data in a academic hospital multidisciplinary intensive care unit. Patients were elgible for inclusion in the study if they received mechanically ventilation for > 25 hours and were < 18 years of age. Time-series data from the patient monitor and mechanical ventilator were abstracted 1-hour preceding and 1-hour following a PEEP change. PEEP increase (PEEPincrease), a responderwas defined as anyone who exhibited an improved dead-space fraction (Vd/Vt); non-respondersdemonstrated a worsening Vd/Vtin the hour following the PEEP change. Features from continuous mechanical ventilation variables were extracted and used to train a support vector machine model in order to predict Vd/Vt response to changes in PEEP. The performance of the model was assessed by calculating the area under the receiver operator characteristic curve (AUROC) and computing measures of diagnostic accuracy.

computable phenotypes were identified and incorporated into the model. The AUROC was 0.82 and 0.90 for classifying response to PEEP increases and decreases respectively. The overall diagnostic accuracy was 0.75 for PEEP increases and 0.84 for PEEP decreases.

Conclusions
The model classified responders to increases and decreases in PEEP with reasonable accuracy. The model performed better for those cases when PEEP was decreases. In the future, these methods may play an important role in optimizing care of the mechanically ventilated pediatric patients, especially if they can be tailored to individual institutions.

Background
Positive end-expiratory pressure (PEEP) is often increased to improve ventilation efficiency and gas exchange during pediatric mechanical ventilation. However, PEEP can ameliorate or exacerbate lung injury.(1) Although it is clinically important to optimize PEEP in this population, there is a paucity of literature to guide clinicians at the bedside.
The physiologic rationale for increasing PEEP is often to reduce physiologic dead-space fraction, improve oxygenation and lung mechanics as well as improve shunt fraction and ventilation/perfusion mismatch. (2)(3)(4) Ratio of dead-space to tidal volume ratio (Vd/Vt) has been associated with pediatric disease severity and success of ventilator weaning. (5,6) Although the use of increased levels of PEEP has been shown to be safe in this population (7)(8)(9)(10), widespread application within the pediatric intensive care unit has not been recommended and further work in this area is needed. (11,12) Recently, we describe the baseline performance of bedside clinicians at selecting patients who will respond to increases or decreases in PEEP. (13) Of note, only 56% of PEEP increase cases demonstrated a subsequent improvement in oxygenation and 54% demonstrated an improvement in Vd/Vt.
Methods and techniques are needed to individualize care, reduce morbidity and duration of ventilation in the pediatric population. (11) Traditionally, gleaning robust representations of pediatric pathophysiology is especially difficult because the underlying causes of disease span body systems and physiologic processes, creating complex nonlinear relationships among observed measurements. This presents an important problem when designing experiments and transcribing medical data in a traditional sense.
However, the increased connectivity of bedside medical devices as well as the availability of sophisticated time-series data analytics and machine learning offer an important opportunity in the modern intensive care unit. Therefore, we sought to apply time-series computable phenotyping on time-series physiologic data and develop a model to predict PEEP response in mechanically ventilated children. A number of computable phenotypes are extracted from continuous physiologic data, important phenotypes are identified and used to train a machine learning model to identify those cases where a subject is likely to see an improvement in gas exchange following a manipulation of PEEP.

Subject selection
Subjects were eligible for inclusion in the study if they received mechanical ventilation for > 24 hours in the pediatric intensive care unit (PICU), age was less than 18 years, continuous mechanical ventilation data were recorded during that time period and they exhibited hypoxic respiratory failure defined as an oxygen saturation index ³5.(11)

Data collection
Mechanical ventilation was applied using the Servo-I (Getinge AB-Maquet, Gothenburg, Sweden) and connected to a physiologic monitor (IntelliVue MP90, Philips Healthcare, Andover, MA). The mechanical ventilator was interfaced to the monitor using the IntelliBridge medical device-interfacing module (model EC10, Philips Healthcare, Andover).
Data was recorded at a frequency of 0.2Hz for the duration of stay in the ICU. The variables included were peak inspiratory pressure (P peak ), positive end expiratory pressure (PEEP), total respiratory rate (RR), respiratory system compliance (C RS ), spontaneous respiratory rate (RR spontaneous ), fraction of inspired oxygen (FiO 2 ), expired minute ventilation (Ve), inspired minute ventilation (Vi), spontaneous minute ventilation (Ve spontaneous ), mean airway pressure (P m ean ), end-tidal CO 2 concentration (PetCO 2 ), volumetric CO2 elimination (VCO 2 ), expired tidal volume (V te ), inspired tidal volume (V ti ), estimate of the pressure in the first 100ms of the breath (P 100 ), end-expiratory flow rate (V ee ), work of breathing of the ventilator (WOB vent ), barometric pressure (Pb), heart rate (HR), oxygen saturation (SpO 2 ) and dead-space fraction (Vd/Vt).
Demographic and outcome data were abstracted from the medical record for each subject and the diagnosis was recorded according to the International Classification of Diseases published by the World Health Organization (Revisions 9 and 10, Clinical Modification) and binned to either primary respiratory, surgical procedure, neurologic, sepsis or other.(14)

Data preprocessing
Because the procedures included in computable phenotype extraction often require the input data to be completely intact (no gaps in data), 1-dimensional linear interpolation was implemented for each variable. Further, the physiologic monitor and mechanical ventilator offer built-in preprocessing but signals can still be corrupted by noise and artifact.(15) Band-pass and low-pass filters were applied to filter out data that was not physiologically plausible according to established methods.(16) A Savitzky-Golay filter was applied to mechanical ventilation data in order to remove noise and artifact but preserve local data phenomenon. (17) Further, normalization of individual parameters is important in a pediatric population since signals are expected to change as the child grows. Data were normalized to either body weight (for respiratory parameters tidal volume, minute ventilation, carbon dioxide elimination, end-expiratory flow rate) and Z-scores were computed for heart rate and respiratory rate. (16,18)

Case identification
For an individual subject, a case was defined as a 2-hour period, to include a 1-hour period preceding and 1-hour period following a change in PEEP. We have previously demonstrated that a period ~60 minutes is necessary to observe physiologic effects from modest changes in PEEP. (19) A quality function was built to ensure that only 'clean' cases were analyzed. A clean PEEP case was defined as one where no ventilator changes were made (other than PEEP and FiO 2 ); the PEEP change was sustained for > 1-hour. For cases where the PEEP was increased, a responder was defined as a case that exhibited any

Statistical analyses
The D'Agostino and Pearson omnibus test was applied to test the normality of the data.
Since the data were not normally distributed, continuous variables are presented as median (interquartile range). To assess model performance, the area under the receiver operator characteristic curve was calculated as well as the diagnostic accuracy,   Table 1. An example of the time-series signals obtained during the present study are shown in Fig. 1.

Discussion
In mechanically ventilated children with hypoxic respiratory failure, the prediction models demonstrated diagnostic accuracy of 75% and 84% for PEEP increases and decreases respectively. The performance of the present models are superior to the empirical probability of improved condition following clinician directed alterations in PEEP. (13) The empirical probability of improved pulmonary condition defined as an improvement in deadspace fraction was 53.9% when increasing PEEP and 46.3% when decreasing PEEP.
The majority of pediatric investigations involving PEEP titration has been done in combination with a recruitment maneuver. (8,(26)(27)(28) In the pediatric literature, few studies have assessed the titration of PEEP without a recruitment maneuver. However, typical management of the child during mechanical ventilation includes the modest titration of PEEP (in increments ranging from 1-3 cmH 2 O). (29) In a large multicenter, randomized controlled trial, a strategy included lung recruitment maneuvers and C RS guided PEEP titration versus low PEEP demonstrated an increase in 28-day mortality. (30) In this study, the authors note the findings did not support routine lung recruitment combined with PEEP titration. Although this seems to suggest, at least in adult subjects

Ethics approval and consent to participate
The Boston Children's Hospital Institutional Review Board approved the present study and need for informed consent was waived since the analysis utilized retrospectively collected data.

Consent for publication
No individual participant data is reported that would require consent to publish from the participant.

Availability of data and materials
The datasets used and/or analyzed during the present study are available from the corresponding author upon reasonable request.

Competing interests
The authors have no competing interests to disclose.

Funding
No external funding was used for the present study

Author contributions
CDS, AG and JHA were responsible for the conception and study design. CDS was responsible for data collection and analysis. All authors were responsible for interpreting the study results. CDS was responsible for drafting the manuscript. All authors read and approved the final manuscript.