EHR Cognitive Load and Statin Initiation In Primary Care

Key Takeaways

• Longer encounter duration was associated with higher statin initiation, while longer time per EHR event was associated with lower initiation.
• Loop count followed a U-shaped pattern and distinct event count followed an inverse U-shaped pattern, indicating that the relationship was not purely linear.
• Average event time emerged as the strongest predictor, with lab-result review and suggested medication order sets contributing positively and order-list modification plus looping back to that list contributing negatively.

Across 20,376 statin-eligible primary care encounters, an EHR audit-log analysis found that workflow signals used as proxies for cognitive load were associated with statin initiation. Longer total encounter duration was associated with a higher likelihood of starting a statin during the visit, and spending more time on each EHR event was associated with a lower likelihood of initiation.

This retrospective, cross-sectional observational study combined EHR audit-log metadata with clinical encounter data from a large academic health system. It covered primary care encounters from January 1 through December 31, 2024. Adults aged 20 to 75 years were assessed for statin eligibility using the 2013 ACC/AHA 10-year ASCVD risk score where applicable and other guideline-based clinical data. Exclusions included no primary care encounter, statin allergy or contraindication, pregnancy in 2024, and missing ASCVD-risk variables.

Multivariable logistic regression adjusted for patient covariates and provider fixed effects, with encounter-level cluster-robust standard errors. Longer encounter duration was positively associated with initiation, with beta 8.223 x 10^-3 and p<0.001, whereas average time per event was negatively associated, with beta -5.017 x 10^-4 and p<0.001. Loops showed a U-shaped association, with linear beta -0.0124 and quadratic beta 0.597 x 10^-5, both p<0.001. Distinct events showed an inverse U-shaped association, with linear beta 0.0124 and quadratic beta -0.597 x 10^-5, both p<0.001.

In exploratory XGBoost modeling, the full-encounter model had an AUROC of 0.87 and identified average event time as the strongest contributor. The top 10 features accounted for about 60% of summed absolute SHAP values, and average event time contributed about 72% across significant features. Positive contributions included time reviewing lab results and suggested medication order sets, while order-list modification and looping back to that list contributed negatively. Inbox activity, diagnosis-linking actions, and summary-view actions also appeared among important features. A pre-decision sensitivity model reached an AUROC of 0.95.

EHR metadata only approximate cognitive load and do not capture all workload dimensions, including intrinsic, extrinsic, and germane components. The study came from a single academic health system using one EHR platform, which limits generalizability. The authors also cautioned that SHAP can misattribute feature importance when predictors are correlated, and individual actions were not linked to specific medications, tests, or diagnoses.