Development of a Respiratory Support Score as a Visualization Tool in Intensive Care

Discussion

We have created a methodology for the automated extraction and presentation of many data elements related to respiratory support within a continuous numerical score. The development of this score was made possible by the presence of an electronic data warehouse, a pipeline that allows end users to access electronic medical record data for a large number of patients. As electronic data warehouses become more prevalent in health care, the extraction, transformation, and use of similarly large and complex datasets will become increasingly possible.

We developed this score with several goals in mind. First and foremost, this tool will allow bedside providers in diverse settings to visualize a patient's complete respiratory history by annotating the RSS curve with relevant events, including surgical procedures or hospital events such as an infection or transplant. Because respiratory support tends to track with illness severity, the RSS is in some ways a broadly useful “illness score” that can be quantified continuously over time. Therefore, it may also be useful as a backdrop against which one might visualize other factors that may either result from (eg, a change in serum creatinine) or cause (eg, degree of valvar regurgitation) critical illness. More immediately, the RSS allows for a visual depiction of a patient's trajectory over time regardless of RST. In this first iteration, we empirically weighted the variables toward those typically weaned in a pediatric cardiac ICU, a population generally absent pulmonary pathology. However, the score can be easily customized for different environments by applying different weightings or even different variables within a similar construct (eg, more heavily weighting peak inspiratory pressure, PEEP, and in the ARDS setting). The score facilitates visualization of changes in RST, such as pressure support (or spontaneous breathing) trials, including the number and duration of those trials. This may be useful to communicate such information in the ICU environment, in which providers change regularly. As illustrated in Figure 1D, hyperlinks to changes in RST might allow for visualization of contemporaneous changes in vital signs or other relevant patient-level responses.

Similar to other composite scores that have been developed for early warning of subject deterioration,^6,7 with further score optimization, the RSS may permit early identification of changes in patient respiratory status (when manifest with increasing respiratory support). The addition of vital signs or measured variables (rather than settings) may improve the power of the RSS to predict extubation readiness.^8–10 It may also permit early identification of trends in a patient's degree of respiratory support that may be represented more subtly in individual variables. The RSS may also identify when a patient's ventilator weaning trajectory has stalled, which may prompt clinical reevaluation.

The second purpose of the RSS was to enable the automated adjudication of meaningful end points, such as the duration of mechanical ventilation, incidence of nonprocedural re-intubation, and the use of noninvasive ventilation. Such end points are commonly used for quality improvement, benchmarking, and central reporting purposes. For some events (eg, extubation), there are discrete fields in which care providers are able to record the event within the electronic medical record. Even when present, though, these fields lacked the time resolution that a more comprehensive approach yields. Creating logic around all available respiratory data allowed us to refine the timing of extubation in 68% of subjects and allowed us to identify the timing of extubation accurately in 1.9% of subjects in whom extubation was not otherwise recorded discretely. This improvement in accuracy resulted from the use of all available data elements related to respiratory support to make a judgment regarding a subject's intubation status. For instance, at times one member of the care team neglected to record an extubation within the extubation field, but others recorded that the subject was on a nasal cannula, and that extubation was only captured accurately using the algorithm. Although the creation of the double-entry rule precluded the algorithm from identifying the fraction of subjects with nearly immediate re-intubations, the use of the RSS combined with a verification field allowed nearly perfect identification of the presence and timing of extubations and re-intubations in automated fashion. As the streaming and capture of continuous data becomes more prevalent, the mapping of data directly from devices to electronic medical record systems will ensure consistency in data entry protocols, remove typographical errors, and increase the detail of data recorded. As this occurs, the utility of such data for quality improvement, benchmarking, and reporting purposes will significantly increase. The importance of scores that amalgamate such data into useful matrices is therefore paramount.

However, the use of data in this way requires several caveats. Ensuring data accuracy (ie, that we are reporting what we intend to report) requires systematic and rigorous validation of data. We found that the most common and impactful error was a systematic misunderstanding of what data are entered into particular fields and what is meant by it. Field names in the electronic medical record are often nonintuitive, confusing, and even misleading. Rules preventing erroneous entry are incomplete. Navigating such errors requires knowledge of clinical workflows, including how data were entered (by whom and at what times), to anticipate common human error patterns (eg, recording room air for of 0.21 in ventilated subjects within a field intended to connote that an extubated subject is breathing spontaneously). The use of a confidence score was an electronic manifestation of such judgments and will likely be an important component for others who choose similar approaches. The validation process also requires a robust reference standard, which will in most cases be manually adjudicated data. However, once data are mapped and verified in this way, reporting may become substantially more efficient and even more accurate, particularly because connectivity between medical devices and the electronic medical record improves and data recording is increasingly consistent. Similarly, the granularity of detail in central repositories may increase when this approach is used, allowing for the exploration of new variables and hypotheses.

The third, and admittedly secondary, purpose of our work was to describe the demographics of mechanical and noninvasive ventilation in a large cohort of subjects at a single center, and to describe associations of each with a surgical complexity score (ie, STAT). A recent registry reported that the extubation failure rate across 15 pediatric cardiac ICUs was 1.1–9.8%, and the only factor significantly associated with re-intubation was duration of mechanical ventilation prior to attempted extubation.¹¹ Similarly, we found that the incidence of nonprocedural re-intubation at 24, 48, and 72 h was significantly associated with both time on mechanical ventilation and the area under the RSS-time curve. Additionally, the incidence of noninvasive ventilation use in this report is lower than that reported in a recent multicenter study of comparable subjects, although we did not include high-flow nasal cannula as a type of noninvasive ventilation.¹² It is likely that the broad use of tools such as RSS may allow centers to compare bedside behaviors that may contribute to this phenomenon, such as the recurrent use of pressure support trials to determine readiness for extubation.