Can Computer Decision Support Help Us Follow Our Own Rules in Pediatric ARDS?

Article Tweet:

Computerized decision support in pediatric #ARDS may improve guideline compliance. Can #MachineLearning improve on this further? @drjonpelly and @cmhorvat discuss a new phase one clinical trial in this month’s @PedCritCareMed.

Pediatric acute respiratory distress syndrome (ARDS) is defined by the Pediatric Acute Lung Injury Consensus Conference (PALICC) as acute respiratory failure and new radiographic infiltrate within 7 days of a clinical insult with an oxygenation index ≥4 or oxygen saturation index ≥5 not fully explained by heart failure or fluid overload.(1) The most recent PALICC management recommendations were released in 2015, and are largely in alignment with adult recommendations released by ARDSnet.(1, 2) After decades of study, open-lung ventilation, high positive end-expiratory pressure (PEEP), low tidal volumes (V_T), and low plateau pressure remain the cornerstones of management.(1–5) This strategy has been associated with an increase in ventilator-free days and decreased mortality.(1–5) Unfortunately, clinical recognition of ARDS is challenging; the LUNG SAFE study of 459 adult ICUs in 2014 found that only 60% of ARDS cases were recognized by clinicians.(6) Adherence to management algorithms is also variable; pediatric intensivists tend to use lower PEEP, higher fractional inspired oxygen, higher V_T, and higher ventilator rates than recommended by guidelines.(7, 8) Deviation from guidelines is associated with increased pediatric ARDS mortality.(8)

In order to assist clinicians, ARDSnet began considering development of computerized decision support (CDS) tools over 20 years ago.(9, 10) In this issue of Pediatric Critical Care Medicine, Hotz et al(11) report the promising results of a phase one clinical trial evaluating the feasibility of using a CDS-driven management protocol, known as Real-time Effort Driven Ventilator Management (REDvent) in a convenience sample of 32 patients at a single center over two years. The study achieved its primary feasibility outcome, reporting 76% overall adherence with protocol recommendations. As might be expected, protocol adherence increased to 92% by the end of the study period as iterative enhancements were made and clinicians gained familiarity with the REDvent tool. Given that guideline-adherence has been associated with decreased mortality in pediatric ARDS,(8) this is laudable in its own right. The proposed REDvent tool is remarkable for its simplicity. In an era where machine learning is capable of analyzing millions of nodes to predict an outcome,(12) REDvent is essentially the combination of two decision matrices, one each for oxygenation and ventilation, both of which attempt to better align the clinician with previously-established guidelines.(1, 2) On failure analysis, clinicians most often rejected REDvent suggestions due to concerns regarding pulmonary hypertension, desaturation, elevated carbon dioxide levels, or previous algorithm intolerance. Examining these reasons for disagreement, we wonder whether the inclusion of additional hemodynamic variables (e.g. including evidence of pulmonary hypertension with right ventricular strain in the oxygenation matrix) and ventilator calculations (e.g. including rate calculations derived from expiratory time constant in the ventilation matrix) might help the algorithm to better align with clinician’s gestalt.

As a next step, how might we harness technology to transform REDvent from CDS to a learning algorithm? First, background-incorporation into an electronic health record would allow the algorithm to be silently screening for patients meeting ARDS criteria, to enhance clinical recognition and promote early guideline-adoption. This would be especially useful given the seasonality of pediatric ARDS, as patient volume is inversely correlated with clinicians’ recognition of ARDS.(6) Second, the use of time-series data could yield substantially improved performance. Thirteen percent of the time, clinicians rejected the REDvent’s suggestion because the patient had already failed a similar recommendation just six hours prior. Incorporation of patient characteristics over time, rather than single-point data, would allow the algorithm to learn from its “mistakes,” and tailor recommendations to individual patient response. Third, our current understanding of ARDS management is based upon guidelines applied to the common pathophysiologic pathway of multiple disease states. Could machine learning help us better distinguish phenotypes of ARDS that respond differently to treatment?(13)

Hotz et al(11) also found a significant increase in 28-day ventilator-free days compared to matched historical controls. While these results are encouraging, the authors astutely note that it is impossible to separate the effects of their CDS from other confounders when using historical controls from 2009–2012 (such as the release of the 2015 PALICC recommendations). The noted differences in V_T noted between controls and REDvent patients are promising, but were present on study enrollment and did not substantially change after the protocol was initiated. The differences in PEEP between REDvent patients and historical controls are also encouraging, particularly in light of the known association between insufficient PEEP and pediatric ARDS mortality.(8) However, it is unclear whether these differences represent the effect of the protocol per-se, clinicians’ increased recognition of the importance of PEEP over the past decade, or both. Another way of evaluating the effect of REDvent might have been to compare these parameters with patients admitted during the study period, but not enrolled in the trial. Presumably, the authors’ ability to 1:4 match their study population with controls over a three-year period implies that their center had enough patients to facilitate at least a 1:1 contemporaneous match.

Another notable aspect of the study by Hotz et al(11) is the incorporation of pressure-rate-product and airway occlusion pressure during maximal inspiration to understand diaphragmatic dysfunction and guide weaning. While not yet included in PALICC and ARDSnet guidelines, there is a strong biologic plausibility that measurements of diaphragmatic work may be useful in predicting extubation success in ARDS patients.(14) Additionally, the striking (though not statistically significant) difference in reintubation rates between REDvent and control patients in the study by Hotz et al(11) suggests that this approach may have substantial benefit in predicting which ARDS patients are ready for extubation. This finding is especially remarkable when coupled with statistically significant reductions in days to first spontaneous breathing trial and length of mechanical ventilation in survivors.

In guiding clinician decision-making according to established best practices, CDS holds great promise for improving outcomes related to a wide range of diseases. While the question of whether CDS will revolutionize management of ARDS remains to be determined, we commend Hotz and colleagues for the development of a smart, sensible tool to assist clinicians with ventilator management for pediatric ARDS. Early clinical results of this trial are inspiring, and we look forward to the results of their ongoing phase two trial.(15)