CRITICAL CARE SYNDROME SURVEILLANCE USING THE ELECTRONIC MEDICAL RECORD: GIZMO IDOLATRY OR PRACTICAL SOLUTION? WORK IN PROGRESS AT KAISER PERMANENTE: THE EDIP PROJECT Gabriel J. Escobar, MD Vincent Liu, MD, MS B. Alex Dummett, MD Arona Ragins, MA Juan Carlos LaGuardia, MS David Draper, PhD Patricia Kipnis, PhD 1
DISCLOSURES Development work leading to this project has been funded by The Permanente Medical Group, Inc.; Kaiser Foundation Health Plan, Inc.; Kaiser Foundation Hospitals, Inc.; the Sidney Garfield Memorial Fund; the Agency for Healthcare Research and Quality ( Rapid clinical snapshots from the EMR among pneumonia patients, 1 R01HS018480-01), and the Gordon and Betty Moore Foundation funding ( Early detection of impending physiologic deterioration in hospitalized patients: feasibility study for a randomized clinical trial ) Current work is funded by the Gordon and Betty Moore Foundation ( Early (Early detection, prevention, and mitigation of impending physiologic deterioration in hospitalized patients outside intensive care: Phase 3, pilot ) No conflicts of interest to disclose 2
PRESENTATION OBJECTIVES What is the rationale for early detection? Two key tools for early detection developed dat Kaiser Permanente Description of the EDIP project and its mathematical and electronic infrastructure The challenge of evaluation 3
SETTING: KAISER PERMANENTE NORTHERN CALIFORNIA Capitated integrated health care delivery system serving 3.5 M members 8,000 MDs, 21 hospitals, ~60 clinics Epic outpatient and inpatient EMRs fully deployed in all facilities Has been using automated severity scores in risk adjustment for all hospitalized patients (not just ICU patients) since 2008 4
RATIONALE FOR EMR-BASED EARLY DETECTION I Multiple l studies have demonstrated t d that t hospitalized patients t who deteriorate outside the ICU have increased morbidity and mortality Published analyses involving Kaiser Permanente data have found that, even after controlling for admission severity of illness, mortality for these patients is 3-4 times expected A more recent internal KPNC matching study found that, after matching for age, sex, diagnosis, admission severity of illness, comorbidity burden, and care directive status (full code, DNR, etc.) unplanned transfers in our 21 hospital system incurred approximately 1200 excess deaths/year and 48,000 excess patient days/year (half of which were ICU days) 5
RATIONALE FOR EMR-BASED EARLY DETECTION II Certain conditions sepsis, evolving respiratory failure seem to have two attributes: they are over-represented among crashes and also appear amenable to earlier treatment Limited data are available on whether in fact early detection can make a difference; Bapoje et al., in a study of a small cohort of medical unplanned transfers, estimated that ~10% would have been preventable, and that the key driver is inappropriate triage At KPNC, a 10% reduction would translate to a 2-4% decrease in ICU census No data are available on the possible benefits of mitigation (e.g., soft landing ) but this also seems intuitively important Evidence supporting the use of manually assigned severity scores is very limited Deployment of comprehensive EMRs is becoming widespread 6
MODULAR TOOLS FOR EARLY DETECTION & EVALUATION OF ITS POTENTIAL IMPACT Cornerstone of our effort in this area has been to map key inpatient EMR fields (vital signs, neurological status checks, pulse oximetry, care directives) data from these fields then are used for predictive modeling We have developed several modular tools (code packages) LAPS2: comprehensive physiologic score applicable to all hospitalized adults (not just ICU patients) COPS2: longitudinal comorbidity score esaps3: automated version of European score (for ICU admits) Two of these LAPS2 and COPS2 are now available to clinicians in real time 7
LAPS2: Laboratory-based Acute Physiology Score, version 2 Objective severity of illness score based on worst values in preceding 72 hours Vital signs + pulse oximetry + neurological status checks + 16 labs (including lactate) Developed using data from 391,584 KP hospitalizations Continuous variable with theoretical maximum of 414 (scores > 200 very rare) Can be replicated by any entity with a high end EMR 8
LAPS2 9
COPS2: CO-morbidity Point Score, version 2 Longitudinal comorbidity score based on all diagnoses incurred by a patient in preceding 12 months Developed using data from 391,584 KP hospitalizations; now assigned monthly to all KPNC adult members Diagnoses are first grouped into CMS Hierarchical Condition Categories (HCCs); these are then used in a regression model, yielding a continuous variable with theoretical maximum of 1014 (scores > 300 very rare) Can be replicated by any entity with longitudinal data Correlates well with POA comorbidity burden, but has certain mathematical advantages due to its being a continuous variable 10
COPS2 11
OVERVIEW OF THE CURRENT KAISER PERMANENTE PILOT 2KPh hospitals (South thsan Francisco &Sacramento) went tlive in November 2013 and April 2014 Electronic decision support is delivered through the Epic EMR Targets triage process in ED and care in 3 non-icu units (med-surg, transitional care unit, telemetry) Uses data as they become available (no additional instrumentation required); factors in real world constraints (e.g., nurses do not chart instantaneously) Based on providing 12 hours lead time (models calibrated with this time frame in mind) 12
What characterizes the KPNC EDIP pilot? 1. VERY EARLY WARNING Information is provided to clinicians BEFORE adverse events occur Alerts are calibrated for 12 hours in the future and run every 6 hours; severity scores provided at time of rooming in and (later this year) in the emergency department at the time the triage decision is to be made Gives clinicians time so that they can involve patients in medical decision-making and establishing goals of care 2. PREVENTIVE 13 Prolonged lead time gives clinicians plenty of time to reassess patient Alert system is also supported by multiple workflows, including several automated checklists that can also be generated in real time 3. Not DISEASE-specific Evaluates all patients and all data elements without bias 13
What does a pre-sponse look like? We know what a code blue looks like. 14
What does a pre-sponse look like? Excuse me doctor? I m going to transfer to the ICU and need a ventilator in 13 hours? Dialysis on day 4? An extra 10 days in the hospital? Oh my, I better call my family to put out some extra food for the dog. 15
First EDIP alert on go live date No drama, kind of dull -- that's why we want lots of lead time 2011 Kaiser Foundation Health Plan, Inc. For internal use only. 16
Early detection of impending deterioration in hospitalized patients (EDIP): real-time risk prediction embedded in the electronic record Inpatient EMR Other KPNC servers (e.g., MIA, IDR) Raw data via web service COPS2 Processed results (alerts, other displays) returned to KPHC for direct use by clinicians External Server f(x,y,z) yz) = x 2 +3xy 3xy 17
ED card swipe time 1 Order to admit 0815 3 8/9/13 16:00 2300 8/8/13 0100 8/9/13 CST RIT ED rooming in time Decision to admit 0800 8/9/13 10:30 2 Time HET (Patient roomed in) 1 2 Advanced alert 1: LAPS2 and COPS2 at time of HBS consult initiation to be deployed in summer or fall of 2014 Advanced alert 2: At hospital entry time, aka rooming in time (LAPS2, COPS2, EDIP alert) 3 Advanced alert 3: on 6 hour cycle for all ward and TCU patients (EDIP alert) 18
Upon Opening the Chart, a Pop-Up Alert Displays if Threshold is Met 19
The Pop-Up will display until Accepted. The Link will Display the Report 20
The Report Contains the Scores and Last Time it was Calculated, along with Additional Information 21
New columns in the Patient List activity show the latest Advance Alert (EDIP) scores & the admission LAPS2 & COPS2 22
PREDICTIVE MODEL WARD MODEL Based on regression model that includes age, sex, LAPS2, COPS2, individual vital signs, vital sign trends, individual labs, time of day, elapsed LOS in the hospital, interaction terms Data used: 650,684 hospitalizations with 20,471 adverse outcomes (unplanned transfer to the ICU or death on the ward in a patient who was full code ) - includes ~263M vital signs &l laboratory measurements Alert frequency set to equal ~ 1 alert/day/7000 discharges/year so as to avoid clinician alert fatigue At this frequency, has a sensitivity of 22-25% and a specificity of 98% 23
PREDICTIVE MODEL ED MODEL (in progress) Based on regression model lthat t includes age, sex, LAPS2, COPS2, individual vital signs, vital sign trends, individual labs, time of day, interaction terms Data used: ~ 4.8 M ED visits Information to be provided to clinicians at the time of triage: probability of crash in next 12 hours if patient is admitted to ward; probability of staying 2 midnights; risk of return to ED or hospital in next 72 hours Will also bring up sensitivity of overall system to ~40-45% range without loss of specificity 24
PLANS FOR UPDATING MODELS We are trying a variety of machine learning approaches This will include developing models with the Lawrence Livermore National Laboratory super computers We will also be testing pattern recognition approaches ( SuperAlarms ) as well as fuzzy logic 25
HOW DOES ONE EVALUATE THE IMPACT OF AN EARLY DETECTION SYSTEM? The two obvious metrics transfer to ICU and mortality will not work ICU care can save lives, and a major part of the project aims to get patients into the ICU sooner Not all patients or their families will desire an escalation of care Consequently, our evaluation strategy will need to be more complex, and will have to include multiple l assessments, not just risk adjusted mortality comparisons 26
CONCLUSIONS Increasing availability of fine grain EMR data will have a significant impact on early detection Other EMR components e.g., order sets, smart phrases that can generate automated check lists, various tracking reports will also play a major role in the response arm This current KPNC pilot shows one approach clearly, other approaches are possible All of these face two common challenges From a statistical perspective, although the aggregate impact on hospital mortality is high, the actual event rate is very low From a methodological perspective, it is going to be difficult to employ a simple outcome measure 27
KEY PAPERS Bapoje SR., Gaudiani JL et al. (2011). Unplanned transfers to a medical intensive care unit: causes and relationship to preventable errors in care. J Hosp Med 6(2): 68-72. Escobar GJ., Greene JD, et al. (2011). Intra-hospital transfers to a higher level of care: contribution to total hospital and intensive care unit (ICU) mortality and length of stay (LOS). J Hosp Med 6(2): 74-80. Escobar, GJ., LaGuardia JC, et al. (2012). Early detection of impending physiologic deterioration among patients who are not in intensive care: development of predictive models using data from an automated electronic medical record. J Hosp Med 7(5): 388-395. Liu, V., Kipnis P, et al. (2012). Adverse outcomes associated with delayed intensive care unit transfers in an integrated healthcare system. J Hosp Med 7(3): 224-230. 230 Delgado MK, Liu V, et al. (2013). Risk factors for unplanned transfer to intensive care within 24 hours of admission from the emergency department in an integrated healthcare system. J Hosp Med 8(1):13-9. Escobar GJ, Gardner MN, et al. (2013) Risk-adjusting hospital mortality using a comprehensive electronic record in an integrated health care delivery system. Medical Care 51(5):446-453 28