A Continuously-Improving Decision Support Tool for Interpreting Postoperative Temperature

Background: Electronic health records (EHRs) create unique opportunities for research and practice.  Questions requiring large-scale data collection can be addressed, and application of research into practice can be facilitated.  We have developed an example of a real-time large-scale data analytic tool that translates research inquiries directly into practice guidance. Physiological postoperative temperature elevations are common.  Little is known about the normal patterns of these temperature elevations for different procedure and patient types.  This knowledge may reduce unnecessary testing and antimicrobial use.

Objective: To characterize the postoperative temperature elevations in a large patient population.

Methods: Step 1 – Multivariable modeling.  NorthShore University HealthSystem is a 4-hospital medical center with a common EHR (EpicCare).  ICD-9 codes were used to identify all patients who underwent any of the following procedure types from January 1, 2006 through January 31, 2010: colectomy (1024), CV surgery (1079), hysterectomy (1455), total knee arthroplasty (2589), total hip arthroplasty (2735), lap-cholecystectomy (2450), craniotomy (840), prostatectomy (1467), spine surgery (3272) and vascular surgery (922).  For all patients, data were collected from the EHR regarding clinical characteristics and maximum daily oral temperature (Tmax) for each postoperative day (POD).  A stepwise selection process was used to prepare 40 multiple linear regression models (one for each procedure and each of 4 postoperative days) describing Tmax; SAS 9.2 was used.  Step 2 – Decision support tool for clinicians.  A web interface was created that allows a clinician to enter variables of interest regarding a real patient.  A graphing tool then uses the appropriate regression equation (i.e. pertaining to the given procedure and postoperative day) to plot a patient-specific Tmax percentile nomogram (Figure).  Step 3 – Continuous model improvement.  Using an extract-transform-load process, all data on new patients undergoing the procedures of interest at our hospitals are collected nightly in a data mart which is used to automatically recalculate the regression models.

Results:   Data were available on 17833 patients (see breakdown by procedure above); in all, 444762 temperature measurements were analyzed.  Tmax differed by procedure and postoperative day (see two examples in Figure).  In some of the 40 multivarible models, additional predictors of Tmax on a given postoperative day included patient age, gender, ethnicity, BMI, mechanical ventilation, ASA class, surgical duration, preoperative Tmax, transfusions and use of antipyretics.

Conclusions: We have characterized postoperative temperature patterns for 10 common procedures and created a decision support tool that incorporates continuously updated clinical data.  Many further applications for this technology exist.