Continuous Physiologic Markers of Heart Rate Variability Derived From Bedside Electrocardiogram Precede Onset of Acute Respiratory Distress Syndrome: A Physiologic Modeling Study.

OBJECTIVE: Acute respiratory distress syndrome (ARDS) is estimated to be prevalent in 10% of ICU patients and results in high mortality rates of up to 45%. The recognition of ARDS can be complex and is often delayed or missed entirely. Recognition of increased ARDS risk among critically ill patients may prompt judicious care management strategies and initiation of preventative therapies known to improve survival. DESIGN: Retrospective observational cohort study. SETTING: In-patient tertiary hospital. PATIENTS: Among 1160 patients (2017-2018), 761 had adequate duration and quality of monitoring waveform data for analysis. INTERVENTIONS: None. MEASUREMENTS AND MAIN RESULTS: This is an observational, retrospective, institutional review board-approved study of patients admitted to ICUs at a tertiary hospital system. Physiologic data were captured among critically ill patients who developed ARDS (n = 62) and matched controls (n = 699) during their hospitalization. Machine learning algorithms were evaluated against statistical features from continuous electrocardiogram (ECG) and sparse clinical data. Waveform-derived cardiorespiratory features, namely measures relating to heart rate variability were found to be robust and reliable features that predicted ARDS up to 2 days before onset. The combined model consisting of waveform features and clinical data with 12-hour prediction horizon achieved an area under the receiver operating characteristic curve and positive predictive value of 0.92 (95% CI, 0.91-0.93) and 0.58 (95% CI, 0.55-0.62), surpassing a model with the clinical data removed (0.86 [95% CI, 0.85-0.88] and 0.49 [95% CI, 0.46-0.52]) and the Lung Injury Prediction Score's maximum of 0.88 and 0.18. CONCLUSIONS: Waveform markers can combine with Electronic Medical Records (EMR) data to improve predictability of ARDS before onset. The markers appear to modulate the sparser EMR data. They also provide, in and of themselves, sufficient dynamical information for comparable results to models with EMR data. Further prospective validation is needed to evaluate the robustness of the model and potential clinical utility.

Duke Scholars

Author Rishi Kamaleswaran Trauma, Acute, and Critical Care Surgery