Acute vital signs changes are underrepresented by a conventional electronic health record when compared with automatically acquired data in a single-center tertiary pediatric cardiac intensive care unit

Abstract

Objective

We sought to evaluate the fidelity with which the patient’s clinical state is represented by the electronic health record (EHR) flow sheet vital signs data compared to a commercially available automated data aggregation platform in a pediatric cardiac intensive care unit (CICU)

Methods

This is a retrospective observational study of heart rate (HR), systolic blood pressure (SBP), respiratory rate (RR), and pulse oximetry (SpO₂) data archived in a conventional EHR and an automated data platform for 857 pediatric patients admitted postoperatively to a tertiary pediatric CICU. Automated data captured for 72 h after admission were analyzed for significant HR, SBP, RR, and SpO₂ deviations from baseline (events). Missed events were identified when the EHR failed to reflect the events reflected in the automated platform

Results

Analysis of 132 054 622 data entries, including 264 966 (0.2%) EHR entries and 131 789 656 (99.8%) automated entries, identified 15 839 HR events, 5851 SBP events, 9648 RR events, and 2768 SpO₂ events lasting 3–60 min; these events were missing in the EHR 48%, 58%, 50%, and 54% of the time, respectively. Subanalysis identified 329 physiologically implausible events (eg, likely operator or device error), of which 104 (32%) were nonetheless documented in the EHR

Conclusion

In this single-center retrospective study of CICU patients, EHR vital sign documentation was incomplete compared to an automated data aggregation platform. Significant events were underrepresented by the conventional EHR, regardless of event duration. Enrichment of the EHR with automated data aggregation capabilities may improve representation of patient condition

INTRODUCTION

Close monitoring of vital signs in the cardiac intensive care unit (CICU) is paramount to detecting short- and long-term changes in a patient’s clinical course. Excursions of vital signs values may be brief and transient, and important long-term changes may be subtle. Identifying meaningful, and possibly critical, changes in the patient’s clinical condition requires incorporation of the clinical assessment and multiple sources of high-volume graphic and numeric data that accumulate at high velocity. In the current era, different electronic health record (EHR) platforms are utilized as the primary data source for clinicians. Currently, these traditional systems rely on static data entry and are not designed to acquire nor display dynamic data streams generated by key medical devices (eg, mechanical ventilator, bedside monitor). While studies on implementation and use of an EHR have demonstrated improvements in certain aspects of intensive care unit (ICU) care (eg, reduced physician documentation time, improved laboratory result turnaround time, and improved satisfaction with workflow), other reports demonstrate decreased efficiency, frustration with EHR workflows, and longer documentation times.^1–7 Whether EHR use improves overall efficiency for physicians in the ICU remains controversial.^1,8,9

With specific regard to the ICU setting, the EHR has been reported to poorly represent the clinical condition of critically ill patients,^10,11 and may portend an increased risk of error.^12,13 At the bedside, one of the foundational components of care is the documentation for permanent capture of accurate vital signs and other clinical characteristics. It is important to emphasize that in the ICU, the documented physiologic data are not only important for retrospective review of the patient’s clinical state, but also for real-time assessments, which occur frequently in the critically ill patient. Specifically, when a provider is called to the bedside to evaluate a concern, availability of high-fidelity data is important to understand the circumstances preceding the event. Keene et al¹¹ previously demonstrated how poor recording of vital signs may lead to underestimation of a patient’s decompensation or inability to detect deteriorating patients¹⁴; this is of critical importance in the pediatric CICU, where management of each patient’s dynamic, individual physiology requires immediate recognition of changes in one or more vital sign measures. Often, these changes are subtle, and an effective clinician will analyze not only the obvious, large variations in any specific measure, but also the subtle constellation of data trends that may act as an early harbinger of clinical destabilization.

The central monitoring stations available in the state-of-the-art contemporary ICU provide high-frequency data trends. However, access to these stations is usually limited to a central work area rather than at each patient’s bedside, and the time window of rolling data available for review is often limited to 24–48 h. Moreover, based on the authors’ experiences, the various central monitoring data streams are difficult to manipulate and modify for review, thus presenting another obstacle to the optimal understanding of each patient’s clinical state. Lastly, a significant limitation of the bedside monitor is the static nature of the alarm limits setup. Alarms are set based on the patient's underlying physiology and clinical state at the time of admission and may be updated on occasion throughout the day. However, the dynamic nature of the critically ill patient renders these static alarm limits inaccurate. Often, the alarm may sound due to events not associated with clinical decompensation (eg, elevated HR due to pain, monitor artifact, or patient care such as endotracheal tube suctioning or obtaining an arterial blood sample), while the critical yet subtle changes in a constellation of vital sign trends remain completely within the static preset alarm limits. When setting static alarm limits in these patients, the resulting multiple or persistent alarms may unfortunately lead to the well-known phenomenon of alarm fatigue,^15,16 and at times may result in lack of recognition of true changes in the patient’s condition.

To overcome these limitations, several high-definition data acquisition platforms have been developed. One of these platforms, originally developed for the pediatric cardiac critical care setting, is the T3 Data Aggregation & Visualization system (T3) (Etiometry, Inc., Boston, MA, USA). T3 interfaces directly with patient monitors and medical devices, and captures data at the monitor’s native output frequency, typically 0.2 Hz (one data point for each sampled variable every 5 s). In contrast, physiologic data entry by bedside clinical staff into the EHR platform in our pediatric CICU routinely occurs at a frequency of approximately 0.00022 Hz (1 data point every hour), although considerable variability exists and depends on institutional practice.

We hypothesized that vital signs data entered into commercial EHR systems do not capture a complete set of vital signs changes in postoperative pediatric patients with cardiac disease admitted to the ICU. To test this hypothesis, we selected patients immediately following open-heart surgical procedures during the first 72 postoperative hours; this period is often physiologically tenuous and highly dynamic and is considered to represent the highest risk phase of postoperative recovery.¹⁷ For each patient, we established dynamic, rolling patient vital signs ranges that we consider to be acceptable for a given patient’s expected postoperative physiologic course and phase of recovery. We compared vital signs data as entered into the EHR against data acquired by the T3 platform with focus on events outside of our established ranges for each variable.

METHODS

This is a retrospective study of archived vital signs data in single-center tertiary children’s hospital of postoperative pediatric cardiac patients less than 18 years of age admitted to the CICU from January 2017 to November 2019. Institutional Review Board approval was obtained. Patients admitted to the CICU directly from the cardiac operating room after cardiac surgery were included; readmissions to the CICU after a planned or unplanned procedure in the cardiac operating room were counted as the same patient, but separate encounters. Partial encounters (encounters lasting less than 72 h) were included in the analysis. Specifically, if reoperation or return to the cardiac operating room occurred within the first 72 h after the initial operation, data collection for the first encounter would conclude upon leaving the CICU, and a new 72-h encounter window would begin upon return to the CICU. Data for patients who expired in the CICU during the 72-h data collection period were included in the study.

Heart rate (HR), systolic blood pressure (SBP), respiratory rate (RR), pulse oximetry (SpO₂), and end tidal CO₂ (ETCO₂) measurements during the first 72 h after admission from the cardiac operating room were obtained from the EHR (Cerner, Inc.; Kansas City, MO, USA) and from the T3 Platform (Etiometry, Inc.; Boston, MA, USA) from author CAF’s institution. The EHR measurements were manually captured and entered by the bedside care team per routine practice. Concurrent vital sign values were automatically captured from the bedside monitor every 5 s, aggregated, and archived by the T3 Platform. HR values were obtained from continuous cardiac telemetry monitoring (surface electrogram), SBP values were obtained from invasive continuous arterial blood pressure monitoring, RR values were obtained from thoracic impedance changes between 2 adjacent surface electrogram leads, SpO₂ was obtained via noninvasive continuous oximetry, and ETCO₂ values were measured via infrared absorption spectroscopy (inline capnography). Pulse oximetry sensors were distributed by Masimo Corp., Irvine, CA, USA, and all other sensors were from Philips Medical Systems Nederland B.V.; Eindhoven, Netherlands.

The T3 Platform is a data acquisition and analytics system that automatically acquires data captured from the patient, clinical devices (eg, mechanical ventilator, near-infrared spectroscopy), and laboratory result systems. It is not approved as a stand-alone monitor. Patient and equipment data streams are sampled and stored every 5 s (0.2 Hz), and laboratory data values are archived at time of laboratory reporting. All data are aggregated and archived continuously without any input from the clinical care team. Data are displayed in real-time on a web-based application accessible via a standard web browser, and can be displayed on a continuously updated bedside LCD screen. However, during the entirety of our study period, high-definition data streams were displayed only at a central station, and not at the bedside. Hence, bedside providers’ EHR data entry could not have been influenced by the bedside availability of this dataset.

The combined dataset was stored on a secure institutional server. A deidentified copy of the dataset was created using the Safe Harbor method.¹⁸ All patient identifiers were removed and replaced with a unique study ID and a random time shift applied to all timestamps related to each patient. The deidentified dataset was then analyzed while the master key relating patient MRN to study ID and time shift remained stored on the secured server.

The physiologic state of each critically ill ICU patient is highly complex and dynamic. At baseline and even moreso in the early postoperative phase, vital signs of children with severe congenital or acquired cardiac disease deviate from the expected normal pediatric ranges. Moreover, significant departures from population-level normal vital sign values are expected in children with single ventricle physiology and children with intracardiac and/or anatomic intrapulmonary shunting. For example, the systemic SpO₂ values of a child with single ventricle physiology following their first stage surgical palliation are expected to range between 75% and 85%; each child’s particular physiology and surgical considerations must be taken into account, and one child’s baseline may be 70%–80%, whereas another with similar physiology and surgical intervention may be 80%–90%. Thus, considering the wide distribution of expected vital signs values and the particularly dynamic nature of these values in the acute postoperative phase, we employed an individualized approach to characterize each vital sign’s T3 data stream. We sought not to characterize normal versus abnormal vital signs values or the clinical significance of specific changes; rather, we sought to describe discrepancies between data represented in the 2 systems. This approach is meant to represent the dynamic state of the postoperative pediatric cardiac surgery patient against which each patient’s own clinical course can be compared.

For the purposes of this analysis, we first established the baseline values for each of the vital signs against which subsequent vital sign values could be compared to identify acute changes (events). We calculated mean values of each vital sign every 5 s based on time-series data of rolling (previous) 12-h intervals of HR, SBP, RR, and SpO₂ values. We defined each patient’s unique baseline as follows: HR, SBP, and RR values within 2 standard deviations above and below the rolling mean, and SpO₂ values no more than 10% below the rolling mean.

In order to avoid the inclusion of artifacts or temporary changes due to provision of clinical care (ie, elevated SBP when drawing an arterial blood sample, decrease of SpO₂ value when noninvasive blood pressure cuff interrogates the extremity), we excluded all events that were sustained for less than 3 min. We also excluded events longer than 60 min since the likelihood of missing these events would be low. Thus, our analysis identified instances where a vital sign deviated from baseline for a period lasting 3–60 min.

Once events were identified, in order to compare the EHR to the automated T3 platform, we evaluated the frequency with which significant events were missed (not entered) in the EHR while still captured in T3. An event was considered missed if there was no contemporaneous data entry in the EHR that corresponded to a significant event identified in the T3 dataset. For all missed events ≥3 and ≤60 min in duration, EHR data were evaluated against T3 data to determine the length of time before values became concordant. EHR and T3 values were considered concordant when (1) the event resolved and the T3 value returned to the baseline range as defined by the rolling mean in T3 or (2) the subsequent EHR entry reflected the departure from the baseline as reflected in T3. Figure 1 provides an illustration of this approach.

Figure 1.

Identification of missed events. T3 data are reflected by the red line; EHR data points are reflected by blue circles. Panel A depicts a heart rate event (increase in HR above the +2σ threshold as calculated by the rolling 12-h time-series data in the T3 data stream), which is also reflected in the EHR. Panel B depicts a missed heart rate event, wherein the heart rate event is reflected in the T3 data stream yet not reflected in the EHR.

Finally, we sought to exclude events due to perceived charting error or device data acquisition failure by identifying and investigating extreme physiological events. We identified all subsets of events, as recorded in either T3 or the EHR, which included apnea (RR = 0), asystole (HR = 0), or profound hypotension (SBP = 0). If a patient suffers a cardiac arrest, one would expect to encounter all 3 of these values concurrently; however, records of isolated apnea, asystole, or profound hypotension are questionable and warrant careful evaluation to determine their accuracy. In our study population of critically ill intubated and mechanically ventilated patients, we further characterized these events by assessing contemporaneous end tidal CO₂ (ETCO₂) data, if available (Figure 2). We compared minimum and mean ETCO₂ levels during the time period 30 min prior to the event and during the event. Physiologically, during apnea, asystole, or SBP = 0 events, ETCO₂ tracing should not be detected (due to lack of ventilation during apnea and lack of pulmonary blood flow during asystole or SBP = 0 events). Hence, identification of stable ETCO₂ levels prior to and during such events would indicate that they do not represent a true change in clinical state but rather EHR charting error or erroneous automatic data capture secondary to equipment failure such as cable disconnection or sensor dislodgement.

Figure 2.

Identification of a physiologically implausible event by comparing ETCO₂ and respiratory rate trends. Panel A depicts ETCO₂ values archived in the T3 platform, and Panel B depicts contemporaneous respiratory rate values archived in the T3 platform. Identification of stable ETCO₂ levels prior to and during such events indicates that the event is very unlikely to reflect a true change in clinical state. These physiologically implausible events (consisting of RR, SBP, or HR of 0 in the presence of adequate ETCO₂ values) are likely due to operator error, equipment failure, or patient sensor disconnection/dislodgement.

RESULTS

Eight hundred fifty-seven patients were included in our analysis, which resulted in 1250 postoperative encounters and 57 447 h of continuous monitoring. Characteristics of the patient population are described in Table 1.

Table 1.

Characteristics of the study population

Gender
Male	481 (56%)
Female	376 (44%)
Age
Neonates (<28 days)	233 (27%)
Infants (1–12 months)	270 (32%)
Children (1–18 years)	354 (41%)

The dataset included 132 054 622 discrete time-stamped data entries in total, comprised of 264 966 (0.2% of total) entries from the EHR dataset, and 131 789656 (99.8% of total) entries from the T3 dataset (Table 2).

Table 2.

Discrete data points comprising the final dataset

	Electronic health record	Etiometry T3 platform
Heart rate	81 862	41 361 904
Systolic blood pressure	27 625	15 550 834
Respiratory rate	81 325	40 631 655
SpO2	74 154	34 245 263
Total	264 966	131 789 656

Overall, there were approximately 500 T3 entries for each entry in the EHR. T3 data points for each vital sign were acquired every 5 s, while EHR data points were recorded considerably less often (Table 3).

Table 3.

Observed intervals between consecutive provider-entered vital sign values in the EHR

Vital sign parameter	Time from admission	Mean	Median	IQR
	(h)	(min)	(min)	(min)
Heart rate	0–72 h	57 min	55 min	48–72 min
	0–24	49	51	45–57
	24–48	61	57	49–77
	48–72	69	57	51–160
Systolic blood pressure	0–72 h	58 min	58 min	52–66 min
	0–24	57	57	49–65
	24–48	74	60	57–180
	48–72	67	60	55–75
Respiratory rate	0–72 h	58 min	60 min	30–60 min
	0–24	49	60	30–60
	24–48	61	60	60–60
	48–72	69	60	60–60
SpO2	0–72 h	63 min	60 min	52–81 min
	0–24	56	55	49–62
	24–48	65	60	53–90
	48–72	74	60	53–160

Missed events

Per our definition, there were no significant events logged into the EHR that were missed in the T3 platform. Of 15 839 total HR events, 48% were missed in the EHR (Table 4). There were 5851 SBP events in total, of which 58% were missed in the EHR. Of 9648 RR events in total, 50% were missed in the EHR, and among 2768 total desaturation events, 54% were not reflected in the EHR (Table 4). The frequency of missed EHR events decreased as the duration of the event increased for all 4 variables (Figure 3).

Table 4.

Duration and number of events missed by the electronic health record

	Heart rate		Systolic blood pressure		Respiratory rate		Pulse oximetry
Event duration	Total events (low, high)	Missed by EHR (low [%], high [%])	Total events (low, high)	Missed by EHR (low [%], high [%])	Total events (low, high)	Missed by EHR (low [%], high [%])	Total events (low)	Missed by EHR (low [%])
3–5 min	6138 (1384, 4754)	3684 (628 [45%], 3056 [64%])	2408 (569, 1839)	1641 (332 [58%], 1309 [71%])	5301 (1301, 4000)	2992 (761 [59%], 2231 [56%])	1349	837 [62%]
5–10 min	4841 (1156 , 3685)	2525 (465 [40%] , 2060 [56%])	1986 (493, 1493)	1174 (240 [49%], 934 [63%])	2566 (702, 1864)	1267 (333 [47%], 934 [50%])	844	452 [54%]
10–15 min	1913 (591 , 1322)	760 (189 [32%] , 571 [43%])	681 (189, 492)	337 (68 [36%], 269 [55%])	857 (224, 633)	313 (79 [35%], 234 [37%])	259	121 [47%]
15–25 min	1621 (657 , 964)	438 (153 [23%] , 285 [30%])	496 (166, 330)	182 (47 [28%], 135 [41%])	580 (224, 356)	169 (69 [31%], 100 [28%])	193	62 [32%]
25–40 min	883 (419 , 464)	142 (77 [18%] , 65 [14%])	195 (91, 104)	42 (17 [19%], 25 [24%])	237 (108, 129)	39 (15 [14%], 24 [19%])	84	15 [18%]
40–60 min	443 (204 , 239)	16 (6 [3%] , 10 [4%])	85 (41, 44)	5 (1 [2%], 4 [9%])	107 (60, 47)	11 (7 [12%], 4 [9%])	39	1 [3%]
Total	15 839 (4411 , 11 428)	7565 (1518 [34%], 6047 [53%])	5851 (1549, 4302)	3381 (705 [46%], 2676 [62%])	9648 (2619, 7029)	4791 (1264 [48%], 3527 [50%])	2768	1488 [54%]

Note: Percentages after the number of events missed by the electronic health record represent the percentage of all missed events of that duration for each vital sign (not percentage of all missed events). High and low events are defined as departures of >2 standard deviations above and below the baseline mean, respectively, except in the case of pulse oximetry where low events are defined as a departure more than 10% below the baseline mean (see Methods section).

Figure 3.

Frequency of missed heart rate events, blood pressure events, respiratory rate events, and pulse oximetry events depicted as a function of event duration. Graphs A₁ and A₂ reflect the frequency of low and high (respectively) missed heart rate events as a function of heart rate event duration. B₁ and B₂ reflect the frequency of low and high (respectively) missed blood pressure events as a function of event duration. C₁ and C₂ reflect the frequency of low and high (respectively) missed respiratory rate events as a function of event duration. D reflects the frequency of missed desaturation (low pulse oximetry value) events as a function of event duration. Note. Horizontal error bars correspond to the range of event durations that were included for the calculation of missed event fractions indicated in Table 4. Vertical error bars represent 1σ uncertainties on the missed fractions, assuming Poisson statistics.

Apnea, asystole, and SBP = 0 events

ETCO₂ data were available in 231/295 (78%) automated apnea events and 11/32 (34%) automated asystole events. SBP = 0 events were not included in the analysis since only 2 such events were identified, neither of which had contemporaneous ETCO₂ data. No differences were found between pre, intra, and post apnea and asystole events’ minimum and mean ETCO₂ values (Figure 2). Hence, these events were nonphysiological and likely represent equipment failure. Notably, 31% of the apnea events and 41% of asystole events were logged into the EHR.

DISCUSSION

In this retrospective study comparing vital signs entered by providers into a conventional EHR versus data acquired by a high-frequency automated capture platform in a pediatric CICU, the EHR dataset does not include a high percentage of events, which may be clinically consequential. Identifying these events is necessary to identify changes in patients’ clinical states; thus, lack of identification could have adverse ramifications. In this experience, shorter events were missed more frequently; however, even events longer than 20 minutes in duration were missed at relatively high rates.

Entry of vital signs data into the EHR is the responsibility of clinical providers. Data from the bedside monitor are selected by the staff member and archived to EHR flowsheets. Hourly data entry is routinely ordered in the CICU. However, more frequent data entry may be ordered in certain circumstances, including procedures, recent admission from the operating room or cardiac catheterization laboratory, or during periods of cardiorespiratory lability. Since these situations usually require the immediate care team members at the bedside to be completely focused on the patient, documentation (ie, “charting”) tasks may be delegated to other staff members, delayed, or omitted. Delayed or omitted entry is problematic because it may lead to inaccurate representation of the physiological state. Even during routine circumstances, the requirement that the nurse or respiratory therapist select data for entry into the EHR introduces a considerable potential for bias. Nonetheless, the retrospective nature of this study limits potential bias that may have been anticipated in a prospective analysis of this nature.

To simplify the comparison between the data platforms we included 4 major cardiorespiratory signals that are incorporated into every CICU patient’s basic vital signs profile, regardless of diagnosis, acuity, or complexity. The wide spectrum of underlying diagnoses in our patient population lends to markedly different physiologies (eg, 1-ventricle vs. 2-ventricle physiology), and due to the dynamic nature of the acute postoperative phase, we elected to individualize the analysis by identifying significant deviations from each patient’s calculated mean vital sign value, rather than using an arbitrary cutoff to characterize normal versus abnormal. Furthermore, the duration of each deviation from the patient’s mean was also considered.

Our findings shed light on the marked differences in fidelity between high-definition data acquisition platforms (represented here by T3) and commercially available EHR platforms in the highly dynamic pediatric CICU environment. We suspect that similar differences would be found when comparing automated data capture to a conventional EHR in many other care settings as well, including pediatric and adult medical and surgical ICUs. As we expected, the fraction of lost (missed) events was inversely proportional to the duration of the events. However, we were surprised to find that even events longer than 20 min were underrepresented in the EHR at relatively high rates. In clinical practice, the ramifications are two-fold: missed events of short duration may reflect a missed opportunity to intervene early in the stage of an evolving process (ie, prior to deterioration), and missed events of longer duration may represent lack of recognition of a patient’s evolving clinical state (ie, failure to recognize evolving cardiogenic shock or hypovolemia).

Notably, our analysis exposed another concerning finding related to EHR data. EHR data entry is performed by experienced staff members who screen the data critically and should refrain from entering artifactual or nonphysiological data. However, our analysis revealed that questionable physiological data are, in fact, logged into the EHR. These data are captured by the automatic platform as well. While it is possible that these entries were simply entered in error, the possibility of erroneous data entry due to a provider knowledge deficit remains. The retrospective nature of our study precludes us from objectively elucidating the circumstances and possible rationale for charting these data, and from identifying any knowledge deficits that may have contributed.

In contrast to the EHR, the T3 interface allows a clinician to align several high-resolution data streams to determine the validity of the data. As we have demonstrated, aligning apnea and asystole intervals with ETCO₂ data in our study population provided more objective understanding of the clinical state of the patients during these episodes. Presently, platforms such as T3 allow manual analysis of aligned physiologic data streams to identify and eliminate artifacts and form an accurate clinical picture. For example, loss of contact between ECG leads and the skin may lead to erroneous numeric HR data capture. The T3 platform provides 3 HR data sources (derived from surface electrocardiography, pulse oximetry, and invasive arterial manometry), and alignment of these HR data streams provides a continuous safety net and cross-validation tool that increases user confidence in the veracity of the captured data.

To our knowledge, this is the first pediatric study comparing an automated data aggregation platform to a commercial EHR. However, similar studies conducted in adults^19–22 have revealed findings consistent with our own. The main limitations of our study design include the retrospective analysis and single-center data analysis. Retrospective study design precludes identification of reasons for the lack of data capture in the EHR, and the single-center dataset limits generalization of our findings to other institutions and care settings. Moreover, this study analyzes vital signs’ perturbations in isolation, yet in clinical practice, a provider synthesizes vital signs and many other clinical data (including, but not limited to, changes in physical examination findings, biochemical markers of perfusion, and diagnostic studies). Particular attention is paid to changes in data over time. Gradual changes are at times subtle and difficult to determine, especially when data resolution is low, such as that entered into the EHR flowsheets. Hence, comparison of high-resolution trends to EHR trends is of significant importance but beyond the scope of this study. We intend to use the current data to conduct this important analysis in the near future.

Furthermore, we chose an empirical statistical approach to characterize normal from abnormal using an inferred baseline (rolling averaged 12-h window). This approach focuses on an individual patient’s dynamic physiological state and avoids the use of absolute thresholds or arbitrary clinical definitions of abnormal versus normal. One must note that this retrospective numerical/statistical definition lacks any clinical validation of the vital signs that were identified as abnormal (with the exception of identification and exclusion of implausible events).

Finally, we were unable to correlate the missed EHR events with clinical interventions, as the study design is agnostic of any interventions taken in response to perturbations in a patient’s clinical state.

CONCLUSION

In critically ill pediatric CICU patients, vital signs documentation by a commercially available EHR platform is incomplete when compared with a higher frequency, automated data acquisition and display platform. The frequency of events that are underrepresented or omitted from the EHR is substantial across a wide range of event durations. Incorporation of automated high-definition data into the EHR rather than relying on manual entry may decrease bedside staff workload and provide a more complete clinical picture. Furthermore, more complete clinical data may also improve predictive clinical decision support using EHR data and secondary uses of such data, including machine learning computational approaches for hypothesis generation and scientific discovery.

AUTHOR CONTRIBUTIONS

CAF obtained institutional review board approval, compiled the dataset, deidentified and archived the data, oversaw execution of data analysis, conceptualized and designed the study, drafted the initial manuscript, and reviewed/revised all iterations including the final manuscript. AWL and AZG conceptualized and designed the study, drafted the initial manuscript, and reviewed/revised all iterations including the final manuscript. All authors approved the final manuscript as submitted and agree to be accountable for all aspects of the work.