Quantifying Electronic Health Record Data: A Potential Risk for Cognitive Overload

We performed a retrospective, single-center, descriptive study in a tertiary PCICU. Patients admitted to PCICU for a minimum of 24 hours from February 1, 2020, to February 15, 2020, were included. Demographic features and surgical complexity by using the Society of Thoracic Surgeons–European Association for Cardiothoracic Surgery (STAT) mortality category, when appropriate, were recorded.¹²  New data elements from the EHR generated over each 24-hour period for each patient were collected. Partial PCICU days were not included.

Data elements were extracted from deidentified flowsheets (hemodynamics, vital signs, and respiratory data, including mechanical support, mechanical circulatory support, nursing documentation, and pharmacy), laboratory data, and clinical reports (radiology, echocardiography, and consult notes). Each flowsheet cell was considered. Generally, cells contained single data points (eg, pulse or mechanical ventilator rate); however, some contained multiple data points for which each element was counted as a single data point. For example, the systolic, diastolic, and mean arterial pressure were each counted, resulting in 3 data points per blood pressure measurement. Because of variability in clinical documentation, a blood pressure may or may not have been accompanied by documentation of where it was checked. When these data were available, they were included. Pharmacy data were extracted from flowsheets and included medication, dose, timing of administration, and rate of infusion when appropriate. Similarly, nursing documentation was extracted from flowsheets (eg, pain assessments). Although clinical reports consistently contain multiple data points and a summary analysis, each instance of a clinical report was treated as a single data point. Patient-reported symptoms, continuous physiologic waveform data, and telemetry data were not included. Generated data were compared to the typical PCICU provider workflow. The study was approved by the medical center Institutional Review Board.

From February 1, 2020, to February 15, 2020, 41 patients were admitted to the PCICU. Thirty patients (50% female) had at least 1 24-hour period in the PCICU and were included. Collectively, 126 24-hour periods were analyzed. Approximately 63% of patients underwent congenital heart surgery, with a median STAT score of 2 (interquartile range [IQR] 2–3), whereas the remainder were medical patients.

Patients generated an average of 1460 (SD 509) data points per 24-hour period. There was no difference between the amount of data generated by medical versus surgical patients (P =.24) (Table 1). Approximately half (51%) of data generated were hemodynamic monitoring, vital signs, and fluid balance documentation. Nearly 30% of generated data included a combination of nursing and respiratory data, whereas laboratory and imaging results accounted for 10% of data (Fig 1).

Frontline providers typically cared for 3 patients (range 2–6) during the day shift (7:00 am to 7:00 pm) and 11 patients (range 8–13) overnight (7:00 pm to 7:00 am), resulting in average exposure to 4380 (range 2920–8760) and 16 060 (range 11 680–18 980) data points during day and night, respectively. PCICU attending physicians cared for 11 (range 8–13) patients during the day and 22 (range 16–26) patients overnight, resulting in exposure to 16 060 (range 11 680–18 980) and 32 120 (23 360–37 960) data points during day and night, respectively.

In this retrospective, descriptive study, we found that clinicians are presented with a significant amount of data through the EHR. The amount of data did not differ significantly between the types of patients, suggesting that most data are a byproduct of routine clinical care.

Clinicians are not quantitative by nature and operate in nonlinear workflows. Yet, with the fundamental transformation of clinical practice into a digital domain, capturing clinical data has been prioritized over interpreting clinical data. Our study supports the idea that clinicians’ abilities to cognitively process disparate data lakes are challenged by voluminous data generation, putting them at risk for cognitive overload. This risk is particularly acute for trainees.¹³  The acquisition of clinical decision-making skills relies on experiential heuristics to choose between relatively few options. The scaffoldings of such heuristics are developed through repetitive pattern recognition. However, the trend to capture increasing amounts of data threatens to dampen clinical patterns, potentially leaving trainees unable to distinguish pertinent signals from irrelevant or erroneous ones.^13–15 

Much of the cognitive work associated with data processing lies in interpretation or contextualization (filtering, trending, and alerting). Contextualizing data transforms it into information. Ackoff,¹⁶  a systems and management scientist, first described a hierarchy in which data optimization generates information, which subsequently generates knowledge and wisdom. This schema reveals a stepwise progression of increasing sophistication and utility. As a field, health care has not developed and widely deployed systems to transform raw data into information, instead requiring individual clinicians to perform this task. Our study reinforces the notion that clinicians are faced with the accumulating task of decoding meaningful information amid a significant amount of unfiltered data.

Much of these data and the risk they generate for cognitive burden existed before EHRs. However, the transition to an electronic platform (with purported benefits of up-to-date complete patient information, intersystem operability, and efficiency gains) was not associated with reductions in cognitive burden. Subsequent reports have revealed concerning divisions between patients and providers due to real-time charting, redundancy in data review, and clinical documentation with regulatory demands. Arguably, the EHR replicated previous challenges and generated new challenges.^3,6  Patient dissatisfaction and physician burnout have been notable and persistent themes associated with the EHR, and there remains a desire to improve data management.

The solution is not to exclusively eliminate data but to optimize it and the systems by which it is delivered to clinicians. Over the past decade, there has been considerable interest in artificial intelligence (AI) as a means of leveraging EHR data for operational and clinical predictions. The health care system’s enthusiasm is partly based on the impact of AI on other industries. However, before implementation of contemporary AI solutions, those industries invested in data optimization.¹⁷  Optimization refers to systems and algorithms that convert data into actionable information for improved efficiency and effectiveness. Airlines use algorithms to create efficient flight schedules. Algorithms enable retail to identify cost-effective delivery routes. AI offered these industries productivity gains for problems in which optimization was already thoroughly exploited. That is, these industries had already taken the transformative step of contextualizing data into information. Lessons from these industries have revealed that similar investment in clinical data optimization is necessary before deployment of more sophisticated systems, including AI.

For health care, the journey to optimize data is a necessary first step in shifting the clinician’s cognitive task to more efficient knowledge generation. Such examples have already revealed value. Scoring systems, including automated early warning systems, amalgamate structured data into meaningful information related to patient decompensation risk.^18,19  Algorithms integrating disparate data pools from pharmacies and laboratories are used to identify patients at risk for acute kidney injury.²⁰  EHR-enabled checklists are associated with a reduction in resource use and health care costs.²¹  Dashboards designed to interact with EHRs to improve clinicians’ data display reduce prerounding time.²²  These examples of data optimization hint at opportunities to employ AI to further transform information into more sophisticated prediction models.²³  Following in the footsteps of other industries, hospitals must invest in technical foundations to optimize data into information. Subsequently, we can envision a future in which implementation of AI results in clinical precision and actionable predictions.

Our study is potentially limited by our inclusion of only patients in the PCICU, among the sickest and most closely monitored of patients. Patients in other settings may generate fewer data points. However, patient-to-provider ratios would also likely be higher, so providers might still face important cognitive burden related to data generation. Additionally, although our findings suggest EHR data contributes to cognitive burden, we did not measure cognitive overload specifically, nor did the project scope allow us to link data generation directly to provider experience or patient outcomes. However, our findings are consistent with previous reports in which it was found that EHR interactions by clinicians are associated with cognitive burden and burnout.^3,7–10  Specifically studying the impact of data generation on clinician experience and patient outcomes would be an important future direction to consider.

Given the extent of demonstrable data generation in our study, we report a potential and important risk for cognitive overload for health care providers (particularly in ICUs). Our findings highlight a potential opportunity to improve data management within our health care system. Integrative systems and data optimization are requisite platforms for AI solutions to develop more informative EHR systems.