Using electronic health record audit logs to identify pediatric intensive care unit teams Authors: Liem M Nguyen, Stefanie S Sebok-Syer, Dane Jacobson, Karley Mariano, John Grunyk, Shelby Burk, Yasaman Nourkhalaj, Kimberley Kirk, Jochen Profit, Christian Rose, A Jay Holmgren, Thomas Kannampallil, Daniel Tawfik Published: 2026 ABSTRACT AbstractBackgroundTeamwork influences patient outcomes, but electronic health records (EHRs) do not reliably capture patient care team members, limiting its study. EHR audit logs provide insights into healthcare workers’ workflows and may help identify patients’ care team members.ObjectiveTo develop, validate, and compare algorithms utilizing EHR audit logs to identify patient-centric primary teams (bedside nurse, frontline clinician, and attending physician) in pediatric intensive care units (ICUs).MethodsWe observed rounds for 1931 patient days (development = 678; validation = 1253) across pediatric (PICU), neonatal (NICU) and cardiovascular (CVICU) ICUs at a quaternary children’s hospital and documented each patient’s daytime care team. We developed 2 algorithms that leveraged EHR audit logs to identify team members: (1) clinically informed heuristics, and (2) a Longitudinal Contribution Score (LCS). Accuracy was computed for each role and algorithm. 95% confidence intervals (CIs) and algorithm comparisons (bootstrap P values) were computed via patient-day bootstrap resampling.ResultsIn the development cohort (PICU), the LCS demonstrated greater accuracy than the heuristics for all roles: nurse (LCS 92.9% [95% CI, 90.9-94.8]; heuristic 67.6% [64.0-71.1]), frontline (83.3% [80.5-86.1]; 77.9% [74.6-81.0]), and attending (66.5% [62.7-70.2]; 61.9% [58.3-65.6]); all P < .005. In the validation cohort (PICU, NICU, and CVICU), only the nurse advantage was statistically replicated (LCS 91.2% [89.5-92.8], heuristic 74.2% [71.7-76.4]; P < .001); with frontline (83.0% [80.9-85.1], 81.2% [79.1-83.4]; P = .110) and attending (69.0 [66.5-71.6] vs 70.0% [67.4-72.4]; P = .326) roles showing comparable accuracy.ConclusionEHR audit log data can accurately identify patients’ bedside nurses in pediatric ICUs. For frontline clinicians and attending physicians whose clinical coproduction and interdependence blurs individual EHR signals, identification is less accurate and supplementary data sources may be needed. Both algorithms provide a scalable foundation for team dynamics research, although the LCS offering greater generalizability than predefined heuristics. Background Intensive care units (ICUs) deliver care through large, interprofessional teams whose membership and roles constantly shift as patient needs change. For instance, the same nurse might serve as a primary nurse for one patient while helping in a secondary role for another. and a patient may receive care from a single physician team for several days before subspecialists are added to the patient’s care team. These dynamics create a uniquely fluid and complex care environment.1 The importance of high-quality teams is increasingly recognized.2–5 Teams with greater stability and collective experience are associated with better patient outcomes.6,7 Therefore, accurate identification of patients’ team members and their associated outcomes is critically important to inform improvement efforts. Although individual- and system-level factors are often quantified at scale, our ability to directly measure how team factors shape care quality remains limited, leaving their impact largely understood only through indirect measures.8 In 2011, Vawdrey et al found that accurate and comprehensive lists of care team members were not readily available in the electronic health record (EHR).9 Despite widespread EHR adoption, healthcare workers (HCWs) are still often neither explicitly nor consistently documented by their role in the EHR.10,11 Alternative methods for identifying team members such as direct observations or surveys are labor-intensive and not scalable. As a result, team structure and dynamics are not reliably and accurately measured or studied at scale, making it difficult to fully assess the associations between interprofessional team contributions and outcomes. Daytime pediatric ICU care in academic settings often revolves around a patient-centric primary team consisting of a bedside nurse, a frontline clinician (resident, hospitalist, or advanced practice provider (APP) such as a nurse practitioner or physician assistant), and an attending physician. However, the complex nature of ICU care means that many additional HCWs (eg, social workers, consultants, clinical fellows, medical students, other residents, and extracorporeal membrane oxygenation (ECMO) nurses) contribute to a patient’s care, often leaving digital traces in the EHR. Importantly, EHR role labels do not uniquely identify these team members: for example, attendings, ICU hospitalists (who have completed residency but often do not function as attendings in the ICU setting), and consulting subspecialists may share the “physician” role, and ECMO nurses may share the same “registered nurse” designation as bedside nurses. Because such role ambiguity is common across disciplines, identifying primary team members requires behavior-based criteria rather than reliance on role labels alone. Kahn et al demonstrated that EHR metadata can be repurposed to identify ICU rounding teams at scale and validated a nurse identification algorithm using retrospective chart review.12,13 Building upon this work, we conducted a prospective observational study using EHR audit logs from 3 ICUs within a quaternary children’s hospital to determine whether other, non-nurse team members can accurately be identified. EHR audit logs are passively collected, timestamped, and patient-attributed records of practically all HCW activity in the EHR.14 These data may enable automated, scalable identification of patient-centric teams and, in turn, systematic research on how team composition influences care quality. Objective To determine if primary team members (ie, nurse, frontline, and attending), associated with a patient in the ICU, can be identified accurately using only audit log data. This study reports the development, validation, and comparison of 2 algorithms informed by: (1) clinically informed heuristics and (2) each HCW’s relative contribution to the patient’s EHR over time termed the Longitudinal Contribution Score (LCS) against the current gold standard of direct observation. Methods Study design and setting We conducted a prospective observational study to develop, validate, and compare algorithms for identifying patient-centric ICU primary teams using EHR audit log data. This study examined teams in 3 ICUs in our children’s hospital: a 36-bed pediatric ICU (PICU), a 40-bed neonatal ICU (NICU), and a 36-bed pediatric cardiovascular ICU (CVICU). We defined each patient’s primary team as the daytime nurse, daytime frontline clinician (resident, hospitalist, or APP), and daytime attending physician—the core triad around which daytime ICU care revolves at our hospital. The study protocol was approved by the Stanford University Institutional Review Board with waiver of informed consent. All audit log data were used exclusively for research on team identification and were not used for individual performance evaluation or monitoring. Direct observation data collection We conducted 2 phases of direct observations to serve as the gold standard for algorithm validation. For each observation day, a trained member of the research team, which comprised both clinician and non-clinician research staff supervised by clinical investigators, visited each occupied bed during a consistent window between 9:00 am and 12:00 pm. Researchers observed clinical workflows and team interactions, after which they approached each patient’s bedside nurse directly and recorded: (1) the patient’s primary attending physician during rounds, (2) the patient’s primary frontline clinician during rounds, and (3) the patient’s primary nurse during rounds. When multiple HCWs were present for a given role, the nurse was asked to identify the HCW for that day. All observations were recorded using standardized REDCap data collection forms with embedded data quality checks. A single observer was assigned per unit per day; therefore, formal interobserver reliability was not assessed. Development phase Data collected for 29 nonconsecutive days in the PICU over a 2-month period (678 patient days) were used to develop our algorithms. Validation phase An independent dataset collected for 15 consecutive days in the PICU, NICU, and CVICU (1253 patient days) was used to validate and compare algorithm performance across different ICU environments. Audit log extraction pipeline All EHR audit log data and HCW data were extracted from Clarity and stored in a PHI-secure Google Cloud Platform BigQuery Instance. We developed a SQL script that aggregated raw audit logs at the HCW patient day level, integrating HCW role and degree title information from Epic’s Clarity database and computing algorithm features for downstream Python processing. All code is available at: https://github.com/cdailab/Using-Electronic-Health-Record-Audit-Logs-to-Identify-Pediatric-Intensive-Care-Unit-Teams. Algorithms developed Clinically informed heuristics Based on clinical expertise and previous work by Kahn et al, we developed an algorithm based on the following heuristics for identifying each primary team member type for a given patient day.12Nurse—Epic user with role of nurse with the greatest number of modification-type and barcode-related actions logged for the patient between 7 AM and 11 AM indicating bedside presence during the typical times of morning rounds and aligned with the nurse shift change schedules at our institution.Frontline—Epic user with role of resident, APP, or physician with the most note-signing-related modify-type actions and order-related modify-type actions for the patient on the calendar day of interest. This accounted for frontlines that may have placed orders pre-rounds or signed notes after work hours.Attending—Epic user with role of physician who had the last sign (or cosign) action on the most patient notes where the selected frontline clinician had the first sign action on the calendar date of interest, even if the signatures occurred later. This accounted for attending physicians who used cosigning functions or finalized documentation after the date of service. Nurse—Epic user with role of nurse with the greatest number of modification-type and barcode-related actions logged for the patient between 7 AM and 11 AM indicating bedside presence during the typical times of morning rounds and aligned with the nurse shift change schedules at our institution. Frontline—Epic user with role of resident, APP, or physician with the most note-signing-related modify-type actions and order-related modify-type actions for the patient on the calendar day of interest. This accounted for frontlines that may have placed orders pre-rounds or signed notes after work hours. Attending—Epic user with role of physician who had the last sign (or cosign) action on the most patient notes where the selected frontline clinician had the first sign action on the calendar date of interest, even if the signatures occurred later. This accounted for attending physicians who used cosigning functions or finalized documentation after the date of service. We used Epic’s E1M Encyclopedia to categorize relevant audit log metric IDs (Table S1). As a refinement step, to ensure each metric ID matched its intended activity type, we reviewed the initial mapping and corrected two discrepancies: 6 metric IDs in the original mapping capturing non-modify-type actions were removed, and 7 metric IDs for order-related modify-type actions that were previously omitted were added. This refinement preceded the final analyses for both the development and validation phases; the final metric ID mapping is provided in Table S1. EHR longitudinal contribution score (LCS) Because the clinically-informed heuristics relied on a finite set of predefined actions, we reasoned: (i) other chart actions not captured in that list could still indicate clinically meaningful engagement for these roles, and (ii) the ICU’s highly dynamic workflow allows transient contributors (eg, consulting specialists or additional nurses responding to an acute event) to accumulate many such actions in a short burst, inflating their apparent involvement compared to HCWs who engage consistently throughout the day. To address both limitations and develop a more generalizable solution, we developed an EHR LCS that quantifies the temporal distribution of a HCW’s chart activity. For each time interval within a defined window, LCS computes the proportion of a specified set of EHR actions on a patient’s chart that a given HCW performed, relative to all HCWs active on that chart during that interval. These per-interval fractions are then summed across all intervals in the window. The framework is parameterized by: (1) the set of EHR actions to include—all recorded action types or a specific subset (eg, order-related actions only); (2) the time window, a contiguous range partitioned into uniform intervals of a chosen granularity (eg, 1-h or 15-min bins); (3) this interval granularity—the uniform bin size partitioning the time window (eg, 1-h or 15-min bins); and (4) the eligible role—the EHR role(s) a candidate must hold, applied when selecting the highest-scoring HCW. For each interval, LCS computes the fraction of included actions that a given HCW performed relative to all eligible HCWs active on the same patient’s chart, then sums these fractions across the window: Because each interval’s fraction is bounded between 0 and 1, a HCW active in only a single interval can achieve a maximum LCS of 1.0, regardless of how many actions they perform. A HCW consistently active across multiple intervals accumulates higher score: for example, a HCW with a 0.3 share in each of 12 hourly intervals achieves an LCS of 3.6, whereas a transient contributor who performs all actions in a single hour achieves at most 1.0. When multiple HCWs are active during the same interval, shares are proportional to action count and sum to 1.0 across HCWs. The theoretical maximum LCS equals the number of intervals in the window. The framework can be adapted to different research questions by varying the parameters. In this study, we set the interval granularity to 1 h and included all audit log action types, reflecting the premise that sustained chart engagement of any kind, not only specific predefined actions, signals primary team membership. We set hourly intervals a priori to reflect the approximate cadence of ICU rounding and documentation workflows rather than tuning this parameter empirically. Using this instantiation, we identified the primary team members for a given patient with the algorithm as follows:Nurse—Epic user with role of nurse with the highest LCS across all action types on the given patient from 7 am to 11 am.Frontline—Epic user with role of resident, APP, or physician with the highest LCS across all action types from 6 AM-6 PM on the given patient.Attending—Epic user with role of physician who had the last sign (or cosign) action on the most patient notes where the selected frontline clinician had the first sign action on the calendar date of interest, even if the signatures occurred later. Nurse—Epic user with role of nurse with the highest LCS across all action types on the given patient from 7 am to 11 am. Frontline—Epic user with role of resident, APP, or physician with the highest LCS across all action types from 6 AM-6 PM on the given patient. Attending—Epic user with role of physician who had the last sign (or cosign) action on the most patient notes where the selected frontline clinician had the first sign action on the calendar date of interest, even if the signatures occurred later. Statistical analysis Performance metric Algorithm performance was assessed using accuracy, defined as the proportion of patient-days on which the algorithm selected the correct HCW for a given role, computed independently for each team member role (nurse, frontline clinician, and attending physician). Non-identifications were counted as accuracy errors but occur only when no clinician of that role recorded relevant chart activity, which is virtually impossible in our ICUs. Accuracy is reported at 4 levels of granularity: (1) per role, per ICU unit (eg, nurse accuracy in the PICU), providing the most granular view of performance; (2) per unit averaged across roles (eg, mean accuracy across nurse, frontline, and attending in the PICU), summarizing unit-specific performance; (3) per role averaged across units (eg, nurse accuracy across all ICUs), summarizing role-specific performance; and (4) averaged across roles and units, summarizing overall algorithm performance. Gold standard data from direct observations contained only a given nurse’s first name and last initial, introducing potential ambiguity; however, across both the development and validation phases, there were no instances in which more than 1 nurse with the same first name and last initial had charting activity on the same patient-day, making this a non-issue in our context. Bootstrap methodology We used patient day clustered bootstrap resampling (2500 iterations) to generate 95% confidence intervals (CIs). For each analysis subset (eg, PICU only, NICU only, all units pooled), patient-days within that subset were resampled with replacement, with the number of resampled patient-days equal to the number of patient-days in that subset (ie, 387 for the PICU, 465 for the NICU, 401 for the CVICU, and 1,253 for all units pooled). 95% CIs were calculated using the 2.5th and 97.5th percentiles of the bootstrap distributions. Algorithm comparison To compare algorithm performance, we computed Δ, the difference in accuracy between LCS and heuristics, on each of the 2,500 bootstrap resamples, using the same resampled patient-days for both algorithms. From this distribution, we derived two-tailed bootstrap P-values and 95% CIs on Δ, testing whether accuracy differed between algorithms for each role within each unit and for each role pooled across all units. Mean accuracy across roles (reported per unit and overall) were computed as the arithmetic mean of the three role-specific point estimates. Results Development phase The development phase encompassed 678 patient days in the PICU, we extracted 65 532 unique HCW-patient days of interactions, representing 198 unique patient encounters and 3328 unique HCWs with 52 different labeled roles. The median daily patient count was 24 (IQR, 21-26) patients. Accuracy by role In the development phase, the LCS algorithm outperformed heuristics for all 3 roles (Table 1). Nurse identification saw the largest absolute improvement (LCS 92.9% [95% CI, 90.9-94.8%] vs heuristic 67.6% [95% CI, 64.0-71.1%]), the highest accuracy observed for any role in either algorithm. Frontline clinician identification improved modestly (LCS 83.3% [95% CI, 80.5-86.1%] vs heuristic 77.9% [95% CI, 74.6-81.0%]). Attending physician identification proved most challenging for both algorithms, with the lowest absolute accuracy for both LCS (66.5% [95% CI, 62.7-70.2%]) and heuristics (61.9% [95% CI, 58.3-65.6%]). Table 1.Development phase algorithm performance by role (678 PICU patient-days).HCW roleClinically-informed heuristics accuracy, % (95% CI)Longitudinal Contribution Score accuracy, % (95% CI) Δ a , percentage point (pp) (95% CI) P value a Nurse 67.6 (64.0, 71.1)92.9 (90.9, 94.8)+25.3 (+21.8, +28.8)<.001 Frontline 77.9 (74.6, 81.0)83.3 (80.5, 86.1)+5.5 (+1.9‚ +9.0).002 Attending 61.9 (58.3, 65.6)66.5 (62.7, 70.2)+4.6 (+1.5‚ +7.7).004 Mean across roles 69.180.9+11.8–a Δ = LCS − Heuristic in percentage points. P = 2-tailed bootstrap P value. 95% CIs from 2500 patient-day bootstrap resamples. Development phase algorithm performance by role (678 PICU patient-days). Δ = LCS − Heuristic in percentage points. P = 2-tailed bootstrap P value. 95% CIs from 2500 patient-day bootstrap resamples. Accuracy averaged across all roles Mean accuracy across all 3 roles was 80.9% for LCS compared to 69.1% for heuristics, driven primarily by the large nurse identification improvement. Validation phase The validation phase encompassed 1253 patient days from 15 consecutive days, yielding 104 267 unique HCW-patient days of interactions, representing 226 unique patient encounters and 4145 unique HCWs with 55 different roles. The median daily patient count was 83 (IQR: 79.5-88) patients per day in all units combined, with 26 (IQR: 24-27.5) patients per day in the PICU, 32 (IQR: 29-34) patients per day in the NICU, and 26 (IQR: 26-28) patients per day in the CVICU. Accuracy per-role, per unit The LCS algorithm outperformed heuristics for nurse identification in all 3 units (all P < .001), with LCS nurse accuracy ranging from 89.5% [86.7-92.3] (NICU) to 92.8% [90.1-95.1] (PICU) (Table 2, rows 1-3, 5-7, 9-11). The algorithms did not differ significantly in accuracy for frontline or attending identification in any unit (Table 2). Table 2.Validation phase algorithm performance by role and unit.RowUnitRole n Clinically-Informed Heuristics Accuracy, % (95% CI)Longitudinal Contribution Score Accuracy, % (95% CI) Δ a , percentage point (pp) (95% CI) P value a 1PICUNurse38780.2 (76.3, 83.8)92.8 (90.1, 95.1)+12.6 (9.0, 16.5)<.0012PICUFrontline38780.9 (76.9, 84.8)80.7 (76.7, 84.3)−0.2 (−4.4, 3.9).943PICUAttending38757.7 (52.8, 62.6)55.2 (50.1, 60.1)−2.5 (−5.9, +0.5).134PICUMean (across roles)38772.976.23.3–5NICUNurse46568.1 (63.6, 72.4)89.5 (86.7, 92.3)+21.4 (+17.7, +25.5)<.0016NICUFrontline46584.2 (80.7, 87.4)87.4 (84.2, 90.4)+3.2 (+0.2, +6.4).0487NICUAttending46587.8 (84.8, 90.8)87.4 (84.2, 90.4)−0.4 (−3.0, +1.9).7998NICUMean (across roles)46580.088.108.10–9CVICUNurse40175.6 (71.3, 79.6)91.5 (88.5, 94.0)+16.0 (+12.1, +20.0)<.00110CVICUFrontline40178.1 (73.8, 82.0)80.3 (76.3, 84.0)+2.2 (−2.2, +6.7).34311CVICUAttending40161.1 (56.4, 65.8)61.1 (56.4, 65.8)+0.0 (−3.7, +3.7)1.000b12CVICUMean (across roles)40171.677.66.1–13Mean (across units)Nurse125374.2 (71.7, 76.4)91.2 (89.5, 92.8)+17.0 (+14.8, +19.3)<.00114Mean (across units)Frontline125381.2 (79.1, 83.4)83.0 (80.9, 85.1)+1.8 (−0.4, +4.1).1115Mean (across units)Attending125370.0 (67.4, 72.4)69.0 (66.5, 71.6)−1.0 (−2.8, +0.9).32616Mean (across units)Mean (across roles)125375.181.15.9–a Δ = LCS − Heuristic in percentage points. P = 2-tailed bootstrap P value. 95% CIs from 2500 patient-day bootstrap resamples. “Mean (across roles)” rows = arithmetic mean of nurse, frontline, and attending accuracies within a unit; because the analysis is restricted to complete-team patient-days (all 3 roles observed), n is identical across roles within each unit, and the macro-average across roles equals the micro-average. “Mean (across units)” rows = accuracy computed on the pooled set of patient-days across all 3 units (n = 1256), equivalent to a weighted average of unit-specific accuracies where each unit’s weight is proportional to its number of patient-days (PICU = 387, NICU = 465, CVICU = 401).bThe CVICU attending P-value is capped at 1.00: the two algorithms achieve identical accuracy for this role, so the two-tailed bootstrap finds no evidence of a difference. Validation phase algorithm performance by role and unit. Δ = LCS − Heuristic in percentage points. P = 2-tailed bootstrap P value. 95% CIs from 2500 patient-day bootstrap resamples. “Mean (across roles)” rows = arithmetic mean of nurse, frontline, and attending accuracies within a unit; because the analysis is restricted to complete-team patient-days (all 3 roles observed), n is identical across roles within each unit, and the macro-average across roles equals the micro-average. “Mean (across units)” rows = accuracy computed on the pooled set of patient-days across all 3 units (n = 1256), equivalent to a weighted average of unit-specific accuracies where each unit’s weight is proportional to its number of patient-days (PICU = 387, NICU = 465, CVICU = 401). The CVICU attending P-value is capped at 1.00: the two algorithms achieve identical accuracy for this role, so the two-tailed bootstrap finds no evidence of a difference. Accuracy per-unit averaged across roles Mean accuracy across the 3 roles was highest in the NICU (LCS 88.1%, heuristic 80.0%), followed by the CVICU (77.6%, 71.6%) and PICU (76.2%, 72.9%); within-unit LCS advantages were again driven primarily by nurse identification (Table 2, rows 4, 8, 12). Accuracy per-role, averaged across units When averaging across all 3 units, nurse identification showed a large LCS advantage (LCS 91.2% [89.5-92.8] vs heuristic 74.2% [71.7-76.4], P < .001) (Table 2, rows 13-15). Frontline (LCS 83.0% [80.9-85.1] vs heuristic 81.2% [79.1-83.4], P = .110) and attending (LCS 69.0% [66.5-71.6] vs heuristic 70.0% [67.4-72.4], P = .326) accuracy did not differ significantly between the algorithms. Of the 3 role-specific LCS advantages observed in the development phase, only the nurse advantage was replicated in the independent validation cohort; the frontline and attending advantages did not reach statistical significance in validation, even within the PICU where the algorithms were originally developed. All units and roles Overall mean accuracy was 81.1% for LCS versus 75.1% for heuristics, though again, this difference was driven primarily by accurate nurse identification (Table 2, row 16). Discussion In this prospective observational study, we demonstrated that EHR audit log data can identify daytime patient-centric primary pediatric ICU teams, with performance varying by clinical role. The LCS algorithm, which measures sustained proportional chart engagement, substantially outperformed heuristics for nurse identification across all 3 ICU environments (Δ = +17.0 percentage points (pp) [95% CI on Δ: +14.8 to +19.3], P < .001; pooled across units), but frontline and attending identification showed comparable performance between algorithms in validation (frontline: Δ = +1.8 pp [−0.4 to +4.1], P = .110; attending: Δ = −1.0 pp [−2.8-+0.9], P = .326; pooled across units), suggesting that audit log data may be insufficient at this time for these roles. Error analyses could guide future refinements. These findings provide a beginning foundation for enabling measurement and study of team dynamics and their impact on care quality. Our algorithms use EHR engagement as a proxy for bedside team membership. Because much of ICU care is computer-mediated (eg, medication ordering, documentation, reviewing results, medication administration workflows, etc.), audit log patterns are a reasonable proxy. We found this approach adequate for roles with more computer-mediated work (ie, nurse and frontlines) but weaker for supervisory roles that generate fewer digital traces (ie, attendings).15,11,16 The approach is best suited for research questions about team stability, care coordination, and the relationship between team composition and outcomes—not as a definitive measure of clinical responsibility or decision authority. Consistent with the concept of collaborative interdependence described by Sebok-Syer et al, we frequently observed 2-3 frontline clinicians working simultaneously on a given patient’s chart during rounds (eg, colleagues started notes or placed orders that were verbally discussed between the patient’s primary frontline clinician and attending physician).17 This collaborative workflow makes it difficult to identify the “true” primary frontline clinician using EHR activity alone. For nursing identification, we observed instances where supporting nurses temporarily relieved the primary nurse or where acute events required the primary nurse to remain bedside while a secondary nurse handled EHR documentation. The LCS algorithm’s ability to recognize consistent engagement throughout the day better distinguishes primary from supporting nursing staff. This study highlights some inherent limitations in relying solely on EHR audit log data for frontline identification in academic settings where there tends to be more coproduction and interdependence, though supplementary data sources such as HCW scheduling, ordering metadata, note metadata, or time-resolved unit-level rosters of HCWs eligible for a given role, may be helpful to resolve which frontline clinician is truly “first call.” Attending physician identification is structurally limited. Because attendings often share the “physician” role label in the EHR with hospitalists and consultants, we used a hierarchical logic: attendings cosign notes initiated by frontline clinicians. This approach depends on correct frontline identification, limiting its independent interpretability, and does not apply to nonacademic settings (although in nonacademic settings, the “frontline” clinician is often the same person as the attending, making this less of a concern). Attending identification was the least accurate role (LCS 69.0% [66.5-71.6], heuristic 70.0% [67.4-72.4]; pooled across units) and varied by unit (PICU: LCS 55.2% vs heuristic 57.7%; NICU: LCS 87.4% vs heuristic 87.8%; CVICU: both 61.1%), likely reflecting differences in documentation practices across subspecialty ICUs. Our study extends the prior work of Kahn et al and Riman et al by leveraging audit log data in a meaningful way to refine team member identification.12,13 When Kahn and colleagues developed their algorithms, they identified 3 types of HCWs: intensivist physicians, bedside nurses, and respiratory therapists. Their approach relied on specific electronic signatures, like our clinically-informed heuristics: progress notes for physicians, clinical assessments and medication administrations for nurses, and ventilator settings and respiratory assessments for respiratory therapists. Only the nurse algorithm was formally validated (93% accuracy via retrospective chart review). Our LCS nurse identification achieved 91.2% accuracy (averaged across validation units), indicating consistency in a metadata-driven approach. Together, this body of work demonstrates that over a decade later, limitations identified by Vawdrey et al can be addressed using the “digital breadcrumbs” HCWs leave in the EHR.9,14 Beyond accuracy, the 2 approaches differ in an important practical dimension: generalizability. The clinically-informed heuristics may require more site-specific configuration than the LCS: categorization of specific audit log actions (Table S1) required for the heuristics may differ across clinical settings, EHR systems, and institutions. The LCS algorithm may be more portable than action-based heuristics because it has less parameters, and its parameters may be more intuitive to set for researchers without audit log expertise. Tuning the LCS parameters that we fixed a priori may further improve performance. In both algorithms, the attending identification logic depends on institutional hierarchy and would require modification in nonacademic settings. Although our validation focused on 3 core roles (nurse, frontline clinician, and attending), our analysis revealed over 50 distinct roles accessing patients’ charts (eg, Respiratory Therapist, Chaplain, Pharmacist, etc.) across the 2 study phases. Our methodology provides a framework that can be extended to identify and study other roles within these broader interprofessional teams. Pharmacists, respiratory therapists, consulting specialists, social workers, and other HCWs leave digital footprints in the EHR that can be analyzed using small modifications to our framework. Some roles may be more difficult to identify; digitally silent roles like clinical fellows, or communication-focused roles like chaplains may require additional methodological considerations.11 Formal validation of extended identifications presents challenges given the complexity of direct observation, but audit log-based approaches offer a scalable, non-intrusive foundation for studying broader care team structure and dynamics. The ability to identify patient-centric teams at scale enables the computation of team level contextual factors that may influence patient outcomes. For example, factors like team stability or team familiarity could be quantified and studied as a potential determinant of care quality. This could inform more context-aware team design and scheduling approaches that intentionally pair team members based on complementary skills, shared experience, patient familiarity, or trainee education opportunities, potentially improving patient outcomes through optimized team composition rather than relying solely on individual HCW characteristics. Limitations Our study should be viewed in light of its design. Our study was limited to a single health system using a specific EHR platform (ie, Epic). Future studies ought to assess the generalizability of our findings to other settings (eg, adult ICUs, pediatric acute care, emergency care) and EHR systems (eg, Cerner). We anticipate that deploying this approach to other institutions would require relatively minor changes, as audit logs are passively collected and stored at all institutions as mandated under the Health Insurance Portability and Accountability Act (HIPAA) and Meaningful Use regulations.14 Still, metric ID mappings and role label conventions may require site-specific configuration. Our analysis was also limited to daytime teams and may not generalize to nighttime coverage patterns given differences in workflows and team structures. In addition, our attending physician identification depends heavily on note-writing patterns, and generalizability to such institutions may require other mechanisms. Our gold standard was based on direct observation performed by different researchers; there always exists the possibility that this could have introduced observer bias or errors. Further, interobserver reliability was not assessed (single observer per unit per day), so individual differences in resolving ambiguous situations may have introduced variability. Several constraints arise from the underlying design of the algorithms and their reliance on EHR audit logs as a proxy for clinical team membership. The algorithm is based on interaction with the EHR and does not provide context about in-person HCW behavior. Similarly, our algorithm cannot distinguish between HCWs who are physically present at the bedside versus those who may engage remotely with the chart (eg, providing backstage oversight).16 This may impact the accuracy of identifying team members who have minimal direct chart interaction despite playing crucial roles in patient care, such as clinical fellows, who are often digitally silent.11 Lastly, bootstrap confidence intervals treat patient-days as independent, but patient-days within encounters have correlated team structures in practice, potentially underestimating variance. Further, accuracy does not adjust for task difficulty: patient-days with few potential candidates inflate accuracy estimates relative to those with larger pools. Conclusion EHR audit log data can be leveraged to identify patient-centric primary teams in the ICU setting with accuracy varying by role. A LCS algorithm that measures the temporal distribution of each HCW’s chart activity across hourly intervals improves nurse identification compared with a clinically-informed heuristics algorithm. While both algorithms did not differ significantly in performance for frontline and attending roles, identification accuracy was consistently higher for frontline clinicians than for attending physicians. These findings suggest that sustained patterns of chart engagement may be useful signal of bedside team membership, but that audit log data alone may have limits for roles that generate fewer digital traces. Nonetheless, this approach provides a scalable method for team identification, which can support future research on how team factors influence quality of care and contribute to patient outcomes in both critical care settings and beyond. Supplementary Material