Feasibility of Real-Time Cognitive Load Assessment Using Electronic Health Record Data and Wearable Sensors in Emergency Medicine: A Pilot Study
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Emergency Department (ED) physicians experience high cognitive workload that can impair patient safety and physician well-being. Despite this critical challenge, robust methods for real-time cognitive load measurement remain elusive. Cognitive Load Theory (CLT) provides a framework distinguishing intrinsic load (IL: inherent task complexity) from extraneous load (EL: factors that interfere with task completion). Building on this framework, the ALERT (Adaptive Load Estimation in Real-Time) project aims to develop dynamic cognitive load models using passively collected Electronic Health Record (EHR) data, validated through biometric and subjective measures.
Objective
This pilot study aimed to (1) demonstrate the feasibility of collecting, synchronizing, and processing multimodal data (EHR interactions, patient encounter data, wearable biometrics, subjective cognitive load reports) from ED physicians during clinical shifts, and (2) explore preliminary relationships between these data modalities to identify promising signals of cognitive load, including the utility of biometrics in predicting high subjective load.
Methods
Two ED physicians (Physician A: 11 shifts; Physician B: 7 shifts) from an academic medical center were monitored. Data included: (a) EHR interaction logs (15-minute aggregations) linked with (b) patient encounter data (ESI scores mapped to an inverted numeric acuity scale) to derive IL and EL proxies; (c) continuous wearable biometric data (EDA, HRV features processed in 5-minute windows, aggregated to 15 minutes); and (d) subjective cognitive load ratings (Paas Scale), entered when physicians experienced notable high or low load. Data were merged onto a synchronized 15-minute timeline. Pearson correlations explored relationships. Logistic regression using selected biometric features (LF/HF, heart rate, RMSSD, skin conductance level) plus a physician dummy variable was used to predict high subjective load (Paas ≥ 7).
Results
Multimodal data collection and synchronization were successfully achieved across 864 15-minute intervals. For Physician A, EHR activity diversity (EL proxy) showed a moderate positive correlation with Paas scores (r = 0.38), while maximum patient acuity (IL proxy) showed a smaller positive association (r = 0.10). For Physician B, several EL proxies demonstrated moderate positive correlations with Paas scores (r = 0.30–0.32), while the number of critical patients (IL proxy) showed a small positive association (r = 0.05). Physician A exhibited a moderate positive correlation between LF/HF ratio and Paas (r = 0.50). A logistic regression model using four selected biometric features and a physician identifier achieved an Area Under the ROC Curve (AUC) of 0.781 in discriminating high vs. low Paas scores (N = 60 observations with Paas data).
Conclusion
This pilot study demonstrates the feasibility of a multimodal approach for cognitive load assessment in the ED. Preliminary findings suggest EHR-derived load proxies correlate with subjective effort, and importantly, a small set of objective biometric features can discriminate high subjective cognitive load with good accuracy. These results strongly support the potential for the larger ALERT project.