Development of Computable Phenotype to Identify and Characterize Transitions in Acuity Status in Intensive Care Unit
Yuanfeng Ren, Tyler J. Loftus, Rahul Sai Kasula, Prudhvee Narasimha Sadha, Parisa Rashidi, Azra Bihorac, Tezcan Ozrazgat-Baslanti
11 Development of Computable Phenotype to Identify and Characterize Transitions in Acuity Status in Intensive Care Unit
Yuanfeng Ren, PhD ,4 , Tyler J. Loftus, MD ,4 , Rahul Sai Kasula , Prudhvee Narasimha Sadha , Parisa Rashidi , PhD, Azra Bihorac, MD MS, Tezcan Ozrazgat-Baslanti, PhD , Department of Medicine, College of Medicine, University of Florida, Gainesville, FL, USA; Department of
Surgery , College of Medicine, University of Florida, Gainesville, FL, USA; Crayton Pruitt Family Department of Biomedical Engineering, University of Florida, Gainesville, FL; Precision and Intelligent Systems in Medicine (PrismaP), University of Florida, Gainesville, FL, USA Correspondence to: Azra Bihorac MD MS, Department of Medicine, Precision and Intelligent Systems in Medicine (Prisma P ), Division of Nephrology, Hypertension, and Renal Transplantation, PO Box 100224, Gainesville, FL 32610-0224. Telephone: (352) 294-8580; Fax: (352) Short title: Computable phenotype for transition in acuity status Conflict of Interest Disclosures: None reported. Funding/Support: Conflicts of Interest and Source of Funding: A.B. and T.O.B. were supported by R01 GM110240 from the National Institute of General Medical Sciences. A.B. and T.O.B.. were supported by Sepsis and Critical Illness Research Center Award P50 GM-111152 from the National Institute of General Medical Sciences. A.B. was supported by a W. Martin Smith Interdisciplinary Patient Quality and Safety Award (IPQSA). T.O.B. has received grant from Clinical and Translational Science Institute (97071) and Gatorade Trust, University of Florida. R.I. was supported by the University of Florida Medical Student Summer Research fellowship.
This work was supported in part by the NIH/NCATS Clinical and Translational Sciences Award o the University of Florida UL1 TR000064. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. A.B., and T.O.B. had full access to all of the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis. bstract Background:
In the United States, 5.7 million patients are admitted annually to intensive care units (ICU), with costs exceeding $82 billion. Although close monitoring and dynamic assessment of patient acuity are key aspects of ICU care, both are limited by the time constraints imposed on healthcare providers.
Methods:
Using the University of Florida Health (UFH) Integrated Data Repository as Honest Broker, we created a database with electronic health records data from a retrospective study cohort of 38,749 adult patients admitted to ICU at UF Health between 06/01/2014 and 08/22/2019. This repository includes demographic information, comorbidities, vital signs, laboratory values, medications with date and timestamps, and diagnoses and procedure codes for all index admission encounters as well as encounters within 12 months prior to index admission and 12 months follow-up. We developed algorithms to identify acuity status of the patient every four hours during each ICU stay.
Results:
We had 383,193 encounters (121,800 unique patients) admitted to the hospital, and 51,073 encounters (38,749 unique patients) with at least one ICU stay that lasted more than four hours. These patients requiring ICU admission had longer median hospital stay (7 days vs. 1 day) and higher in-hospital mortality (9.6% vs. 0.4%) compared with those not admitted to the ICU. Among patients who were admitted to the ICU and expired during hospital admission, more deaths occurred in the ICU than on general hospital wards (7.4% vs. 0.8%, respectively).
Conclusions:
We developed phenotyping algorithms that determined patient acuity status every four hours while admitted to the ICU. This approach may be useful in developing prognostic and clinical decision-support tools to aid patients, caregivers, and providers in shared decision-making processes regarding resource use and escalation of care.
NTRODUCTION
In the United States, 5.7 million patients are admitted annually to intensive care units (ICUs), with costs exceeding $82 billion. The annual cost of care delivered to these patients exceeds 4.1% of our national health expenditures, while ICU mortality ranges from 10% to 29%. Close monitoring and dynamic assessment of patient acuity are key aspects of ICU care; both are limited by the time constraints imposed on healthcare providers. Assessment of patient acuity in the ICU relies almost exclusively on physicians’ clinical judgment and vigilance. There is a critical unmet need for assessment of the patient with continuous physiologic measurement and clinical data. ICU physicians spend only 9.4% of their clinical time in direct patient contact. Similarly, most ICU nurses spend only 10% of their time on direct patient assessments of pain and mobility. Patients may not be directly observed by physicians or nurses for 80% of their stay in an ICU. Both self-report and manual observations suffer from subjectivity, poor recall, and limited number of administrations per day, and may lead to missed opportunities for timely interventions.
Manual and repetitive patient assessments result in personnel shortages and burnout. Critical care teams are under significant work pressure. Almost a third of ICU nursing teams suffer from burnout. High nursing workload is one factor in the occurrence of life-threatening adverse events in the ICU.
Given this burden, there is an urgent need for automation of routine tasks. Autonomous assessments can enhance critical care workflow efficiency by facilitating routine nursing assessments in the ICU and allow nurses to spend time on more critical tasks. In addition, assessments that are associated with prognosis and clinical trajectory have the potential to augment prognostication and decisions regarding escalation of care and resource use.
In this study, we developed computable phenotypes for acuity status that will be used to determine clinical trajectories in the ICU, which will be used in the future to develop a precise clinical trajectory prediction tool by utilizing high-resolution physiological signals and digital EHR data.
ETHODS
Study design
This study was approved by the University of Florida Institutional Review Board and Privacy Office (
Data Source
Using the University of Florida Health Integrated Data Repository as Honest Broker, we created a single center longitudinal dataset extracted directly from electronic health records of all patients who underwent surgery at the University of Florida Health in Gainesville, Florida between June 1, 2014 and May 1, 2019. This repository includes demographic information, comorbidities, vital signs, laboratory values, medications with date and timestamps, and diagnoses and procedure codes for all index admission encounters as well as encounters within 12 months prior to the index admission and 12 months after admission. All electronic health records were de-identified, except that dates of service were maintained. The dataset includes structured and unstructured clinical data, demographic information, vital signs, laboratory values, medications, diagnoses, and procedures.
Participants
We included 383,193 hospital encounters from 121,800 patients with age 18 years or older that were admitted to UF Health between June 1, 2014 and August 22, 2019. Among these hospital encounters, there were 51,073 encounters (38,749 unique patients) with at least one ICU stay that lasted more than four hours.
Definition of Acuity Status
At the end of each four-hour interval, for patients who have not died or discharged alive, acuity status was determined as unstable vs stable. Patient was considered as unstable if patient required at least one of the life supportive therapies: vasopressors, mechanical ventilation, ontinuous renal replacement therapy, or massive blood transfusion (defined as at least 10 units in past 24 hours). Otherwise, the patient was labeled as stable. (Figure 1)
Figure 1.
Flowchart for logic for phenotyping acuity status in ICU ata elements and rules used for identification of instability
Data elements used for identification of acuity status are provided in Table S1. Computable phenotyping algorithms used rule-based approach as detailed below:
Identification of ICU stay
We identified date and time during which patient is an ICU room which were adjudicated by clinicians. ICU stays which are within 24 hours of each other were merged together.
Identification of Vasopressor Use
Use of vasopressors was identified by checking existence of at least one of the RxNorms listed in Table 1 for epinephrine, vasopressin, phenylephrine, norepinephrine, droxidopa, or ephedrine during the time interval of interest.
Table 1 . Identification of vasopressors using RxNorms
Vasopressor Name RxNorm
Epinephrine 1001079, 1001082, 1001086, 1001089, 1010677, 1010683, 1010688, 1010745, 1010751, 1010759, 1012377, 1012384, 1012391, 1012707, 107602, 107606, 1305268, 141848, 203180, 214547, 237187, 24255, 245317, 284622, 310115, 310116, 310132, 310133, 310134, 313963, 313967, 314610, 362, 3992, 66887, 691245, 727345, 727347, 727373, 727374, 727386 Vasopressin 11149, 374283, 313578 Phenylephrine 8163, 8164, 373369, 373370, 374570, 373372, 379042 Norepinephrine 242969, 7508, 7512, 1745276 Droxidopa 1489913, 1490026, 1490034, 1490038 Ephedrine 248717, 1115910, 214538, 310110, 1116191, 91165, 1116195, 1116294, 198918, 1116146, 91166, 1115991, 310109, 387570, 3966, 5032, 991423 Identification of continuous renal replacement therapy
We identified continuous renal replacement therapy (CRRT) using crrt file provided by IDR which includes identifiers, date and time stamp, and treatment type. We identify timeframes during which treatment type is continuous veno-venous hemofiltration (CVVH), continuous veno-venous hemodialysis (CVVHD), or, Continuous veno-venous hemodiafiltration (CVVHDF).
Identification of
Massive
Blood Transfusion
For each 4 hour interval, we determined whether the total amount of blood transfusion in the last 24 hours exceeded 10 units. We identified whether there was a massive blood transfusion using transfusion file, provided by IDR, which includes procedure description, amount of transfusion, date and time stamp for order, start, and end of transfusion. We only considered red blood cell transfusion based on procedure description. If there was at least one of the start or end date and time, amount was imputed as 1 unit. If amount was not missing, but either start or end date and time was missing, the missing one was calculated assuming the duration between start and end is equal to the median duration between start and end date time in the cohort which is 1.5 hours. If both start and end date time were missing, start date and time was imputed assuming duration between order date and time and start date and time is the median duration in the cohort which is 1.58 hours, and end date time was imputed by adding 1.5 hours to imputed start date and time. (Figure 2)
Identification of Mechanical Ventilation Use
We identified use of mechanical ventilation using respiratory file, provided by IDR, which has identifiers, date and time, respiratory device, ventilation mode, measured values for respiratory vitals that includes; oxygen flow rate, tidal volume, end-tidal carbondioxide (etCO2), and peep. Patient is assumed to be on mechanical ventilation if at least one of the tidal volume, end-tidal carbondioxide (etCO2), peep, mechanic respiratory rate or ventilation mode is not missing or respiratory device is ventilator or endotracheal tube (ETT). Device is assumed to be same until there is a change in device type. (Figure 3) Figure 2.
Flowchart for massive blood transfusion indication Figure 3.
Identification of Mechanical Ventilation Use Acuity Status patterns derivation
To understand how the acuity status evolve in ICU, once we determined the acuity status every four hours, we used the k-means clustering method to derive the potential patterns. Acuity status in first 3 days were included. For those admissions with less than three days, we pad the missing acuity status as stable if the final status is discharge from ICU; if not, the missing acuity status are labeled as unstable. Thus for each ICU admission, we had 18 (3 * 24 / 4 = 18) acuity status as clustering features feeding to clustering method. Patterns of acuity states were visualized by line plots which illustrate the average value of acuity status across phenotypes over time along with 95% confidence interval for each estimate.
RESULTS Patients
We had 383,193 encounters (121,800 unique patients) admitted to the hospital, and 51,073 encounters (38,749 unique patients) with at least one ICU stay that lasted more than four hours (Table 2). These patients requiring ICU admission had longer median hospital stay (7 days vs. 1 day) and higher in-hospital mortality (9.6% vs. 0.4%) compared with those not admitted to the ICU. Among patients who were admitted to the ICU and expired during hospital admission, more deaths occurred in the ICU than on general hospital wards (7.4% vs. 0.8%, respectively). Table 2.
Clinical characteristics of cohort
UF Health Hospital Admissions (06/01/2014-08/22/2019)
Number of patients 121,800 Number of encounters 383,193 Number of patients with at least one ICU stay 38,749 Number of encounters with at least one ICU stay 51,073 Number of ICU stays 54,178
Hospital admissions, n 383,193
Hospital LOS, days, median (25th, 75th) 2 (1, 4) In-hospital mortality, n (%) 6240 (1.6)
Never to ICU admissions, n (%) 332,120 (86.7)
Hospital LOS, days, median (25th, 75th) 1 (1, 3) In-hospitality mortality, n (%) 1346 (0.4)
ICU admissions, n (%) 51,073 (13.3)
Hospital LOS, days, median (25th, 75th) 7 (4, 13) ICU LOS, days, median (25th, 75th) 4 (2, 7) In-hospitality mortality, n (%) 4894 (9.6) Death within 7-days of ICU admission, n (%) 2984 (5.8) Number of intervals from ICU admission to hospital discharge 35 (18, 65)
ICU admissions died in ICU, n (%) 3775 (7.4)
Hospital LOS, days, median (25th, 75th) 6 (3, 14) ICU LOS, days, median (25th, 75th) 4 (2, 10) Death within 7-days of ICU admission, n (%) 2534 (5.0) Number of intervals from ICU admission to hospital discharge 25 (9, 61)
ICU admissions died in Ward, n (%) 389 (0.8)
Hospital LOS, days, median (25th, 75th) 12 (7, 20) ICU LOS, days, median (25th, 75th) 5 (3, 9) Death within 7-days of ICU admission, n (%) 161 (0.3) Number of intervals from ICU admission to hospital discharge 55 (31, 91) Distribution of acuity status
Distributions of acuity status within first 15 days of ICU admission are illustrated in Figure 4. Four acuity status as stable, unstable, discharge from ICU, and dead in ICU were included. Number of encounters in ICU decreased rapidly from more than 50,000 to 10,000 encounters within first seven days, consistent with the 75th percentile ICU days. Every four hours primarily consisted of stable and unstable acuity status, and very few discharges and deaths. The percentage of stable acuity status first increased for 24-48 hours following ICU admission and then gradually decreased over time for approximately two weeks. The maximum percentages of patients who were discharged or died during each four-hour period were 1.83% and 0.38% respectively.
Figure 4.
Distribution of acuity status within first 15 days in ICU Transition probability matrix of acuity status
The transition probability matrix of acuity status was listed in Table 3 and illustrated in Figure 5 for all transitions from ICU admission to ICU discharge. Stable acuity status transited to stable status within next 4 hours with a high probability (93.0%), and to unstable status with probability (2.2%). Unstable acuity status transited to unstable status within next 4 hours with a high probability (93.0%), and to stable status with probability (6.6%). Patients with stable acuity status are more likely to discharge from ICU compared with unstable acuity status (4.6% vs. 0.0%). Patients with unstable acuity status are more likely to dead in ICU compared with stable acuity status (0.4% vs. 0.1%).
Table 3.
Transition probability matrix of acuity status
Acuity status within next 4 hours
Stable Unstable Discharge Dead
Stable
Unstable
Figure 5.
Acuity state transitions in every 4 hr interval from ICU admission to ICU discharge Phenotypes of acuity status
Three phenotypes of acuity status were derived (Figure 6). Cluster 1 was mainly manifested as stable acuity status. This cluster had the most patients (35844, 70%), lowest mortality (2.5%), shortest median ICU stay (3 days) and hospital stay length (6 days). Cluster 2 presented with persistent unstable acuity status. As a result, this cluster contained the patients (9619, 19%) with highest mortality (31.8%) and longest median ICU stay (8 days) and hospital stay length (12 days). Cluster 3 first presented with unstable acuity state, and then gradually back to stable acuity state. This cluster contained patients (5610, 11%) with second lowest mortality (3.5%), and median ICU stay (4 days) and hospital stay length (8 days).
Figure 6.
Acuity status distribution across phenotypes within first three days of ICU admission DISCUSSION
These findings demonstrate that during the first 24 to 48 hours after ICU admission, there is an increase in the raw number and percentage of ICU patients who are stable. After this point, rates of discharge from the ICU, including death in the ICU, remain constant, and the remaining population of ICU patients shifts toward the unstable phenotype. This trend continues for approximately two weeks after ICU admission, consistent with clinical definitions of chronic critical illness.
For both stable and unstable acuity types, there was a 93% probability that the following acuity state would be unchanged. The probability of death during the following four-hour period was approximately 4-hold higher among unstable vs. stable patients, suggesting that our methods for classifying patient acuity by vasopressors, massive blood product transfusions, mechanical ventilation, and renal replacement therapy were effective in identifying patients at increased risk for death. Finally, we identified three patient clusters of patients, one that remained stable, one that was persistently unstable, and one that was unstable at the time of ICU admission but transitioned to stability within 48 hours. Identifying these clusters has potential clinical utility for informing discussions among patients, caregivers, and providers regarding prognosis, and informing decisions regarding resource use. This task can be automated, without requiring additional patient assessments by ICU health care workers, who face worsening work force shortages and job-related stress. At the time of ICU admission for an unstable patient, patients and their caregivers may wish to embark on a course of aggressive, life-sustaining treatments if there is a high probability of recovery and transition to stability. Some critically ill patients have previously expressed a desire to forego prolonged life-sustaining treatments. In these cases, providing the patient and their caregivers with information to suggest that the probability of early recovery is low could augment their decision-making process and alleviate the stress associated with the decision to forgo aggressive, resource-intense therapy. Therefore, identifying acuity phenotypes has the potential to augment clinical prognostication and decision-making. This study was limited by a single-institution design, which limits its generalizability to other practice settings. We made binary distinctions between stable and unstable patients because these distinctions facilitate the identification of transitions in acuity states and clustering based on these states. However, we recognize that true patient acuity actually exists on a continuum. It remains unknown whether identifying acuity states and clusters will affect prognostication and decision-making in the clinical setting; this is a target for future research. CONCLUSION
We developed phenotyping algorithms that determined patient acuity status every four hours while admitted to the ICU. This task can be automated, which has the advantage of avoiding additional patient assessments by ICU health care workers, who already face worsening work force shortages and job-related stress. Automated acuity phenotyping has the potential to leverage high-resolution physiological signals and digital EHR data to develop prognostic and clinical decision-support tools that aid patients, caregivers, and providers in shared decision-making processes regarding resource use and escalation of care. Table S1 . Data elements that are used to run acuity phenotyping algorithm
Used for identification of Features Description Format
Identifier patient_deiden_id Deidentified Patient ID Strings Identifier encounter_deiden_id Deidentified Encounter ID Strings Continuous renal replacement therapy meas_value Measured Value Measured Value Continuous renal replacement therapy recorded_time Recorded Time The date and time value was recorded Continuous renal replacement therapy vital_sign_group_name Name of the measured variable Group name of vital sign. Values include: Device Number, Hourly Net Balance, Maintenance, Output (mL), Presctiption, Therapy, or Treatment. Continuous renal replacement therapy vital_sign_measure_name Name of the measured variable Name of the measured variable Mechanical Ventilation respiratory_datetime Respiratory DateTime The date and time when the respiratory device is used Mechanical Ventilation respiratory_device Respiratory Device The device being used to deliver oxygen or move air in and out of the lungs Mechanical Ventilation adult_vent_mode Adult Ventilator Mode The breathing pattern programmed into the mechanical ventilator, which is moving air in and out of the lungs Mechanical Ventilation adult_mech_resp_rate Mechanical Respiratory Rate Breaths per minute for a patient on a mechanical ventialtor/breathing machine Mechanical Ventilation peep End of Expiratory Pressure The pressure in the airways at the end of exhalation (mm Hg or cm H20) Mechanical Ventilation tidal_volume Tidal Volume (mL) The volume of air that moves with each breath (mL) Mechanical Ventilation etco2 End-tidal Carbon Dioxide Amount The amount of CO2 in the air that is moving out of the lungs (mm Hg) for patients on mechanical ventilation and without invasive airway, fio2 Massive blood transfusion blood_end_instant Blood transfusion end date and time Blood transfusion end date and time Massive blood transfusion blood_start_instant Blood transfusion start date and time Blood transfusion start date and time Massive blood transfusion frequency Amount of blood transfusion Amount of blood transfusion as Transfuse X units Massive blood transfusion procedure_desc Description of the procedure Description of the procedure as Transfuse Red Blood Cells/Plasma/Platelets/… Vasopressor Use taken_datetime Action Taken DateTime Time at which MAR action was logged Vasopressor Use med_order_display_name Medication Order Display Name Medication order display name Vasopressor Use rxnorm_concat Concatenated Medication Name The concatenated medical name of the medicine Vasopressor Use mar_action Medical Administration Record Action Taken Medical administration record action taken Vasopressor Use med_order_discrete_dose Medication Order Discrete Dose The dosage at which the medication needs to be administered Vasopressor Use med_order_route Medication Order Route The medium through which the medication is administered Vasopressor Use med_order_discrete_dose_unit Medication Order Discrete Dose Unit The units of the medication dosage Vasopressor Use height_weight_datetime Height and Weight Measured DateTime The date and time that the patient's height and weight are measured. Vasopressor Use weight_kgs Weight (kgs) Patient's weight in kilograms (kg) REFERENCES
1. Wallace DJ, Angus DC, Seymour CW, Barnato AE, Kahn JM. Critical care bed growth in the United States. A comparison of regional and national trends.
Am J Respir Crit Care Med (4): 410-6. 2. Elias KM, Moromizato T, Gibbons FK, Christopher KB. Derivation and validation of the acute organ failure score to predict outcome in critically ill patients: a cohort study.
Crit Care Med (4): 856-64. 3. Mamykina L, Vawdrey DK, Hripcsak G. How Do Residents Spend Their Shift Time? A Time and Motion Study With a Particular Focus on the Use of Computers. Acad Med (6): 827-32. 4. Wong DH, Gallegos Y, Weinger MB, Clack S, Slagle J, Anderson CT. Changes in intensive care unit nurse task activity after installation of a third-generation intensive care unit information system. Crit Care Med (10): 2488-94. 5. Brown H, Terrence J, Vasquez P, Bates DW, Zimlichman E. Continuous monitoring in an inpatient medical-surgical unit: a controlled clinical trial. Am J Med (3): 226-32. 6. Kipnis E, Ramsingh D, Bhargava M, et al. Monitoring in the intensive care.
Crit Care Res Pract : 473507. 7. Wollschlager CM, Conrad AR, Khan FA. Common complications in critically ill patients.
Dis Mon (5): 221-93. 8. Rubins HB, Moskowitz MA. Complications of care in a medical intensive care unit. J Gen Intern Med (2): 104-9. 9. Desai SV, Law TJ, Needham DM. Long-term complications of critical care. Crit Care Med (2): 371-9. 10. Coomber S, Todd C, Park G, Baxter P, Firth-Cozens J, Shore S. Stress in UK intensive care unit doctors. Br J Anaesth (6): 873-81. 11. Verdon M, Merlani P, Perneger T, Ricou B. Burnout in a surgical ICU team. Intensive Care Med (1): 152-6. 12. Bucknall TK. Medical error and decision making: Learning from the past and present in intensive care. Aust Crit Care (3): 150-6. 13. Ksouri H, Balanant PY, Tadie JM, et al. Impact of morbidity and mortality conferences on analysis of mortality and critical events in intensive care practice. Am J Crit Care (2): 135-45; quiz 46. 14. Pagnamenta A, Rabito G, Arosio A, et al. Adverse event reporting in adult intensive care units and the impact of a multifaceted intervention on drug-related adverse events. Ann Intensive Care (1): 47. 15. Valentin A, Capuzzo M, Guidet B, et al. Patient safety in intensive care: results from the multinational Sentinel Events Evaluation (SEE) study. Intensive Care Med (10): 1591-8. 16. Valentin A, Schiffinger M, Steyrer J, Huber C, Strunk G. Safety climate reduces medication and dislodgement errors in routine intensive care practice. Intensive Care Med (3): 391-8. 21 17. Jastremski CA. Retention of critical care nurses: Important to the future of critical care. Crit Care Med (7): 2015. 18. Duke EM. The Critical Care Workforce: A Study of the Supply and Demand for Critical Care Physicians. Health Resouree and Service Administration US Department of Health and Human Services W ashington DC US A 2008
Lancet Respir Med (7): 566-73. 20. Kahn JM, Le T, Angus DC, et al. The epidemiology of chronic critical illness in the United States*. Crit Care Med (2): 282-7.(2): 282-7.