Development of a predictive model and electronic health record-based probability scoring system and dashboard for postoperative respiratory failure - PROJECT SUMMARY/ABSTRACT
The objective of this research proposal project is to identify modifiable factors associated with different
postoperative respiratory failure (PRF) phenotypes in adults following elective surgery and to utilize this
information to develop and deploy a predictive model and electronic health record-based probability scoring
system and dashboard for PRF. PRF, defined as the prolonged inability to wean from mechanical ventilation or
inadequate oxygenation and/or ventilation, has an incidence of up to 7.5% and has been associated with a
risk-adjusted $53,000 increase in hospital charges, 9 extra days of hospitalization, and a 22% increase
in-hospital mortality. With the number of elective surgical procedures increasing annually, there is an urgent
and unmet need to reduce the incidence and burden of this potentially preventable event by elucidating risk,
preventive, and therapeutic factors. These factors, some of which may be modifiable, may differ between
phenotypic presentations. AIM 1: To optimize and validate an automated, EHR-based, clinical prediction model
for PRF. We will automate data collection and model the contributions of pre-and intra-operative factors on full
model discrimination and calibration. Hypotheses: (H1.1) It is possible to automate data curation. (H1.2) A
model including data from 2014-2021 and quantitative risk indices will outperform our previous model that used
data from 2012-2015. AIM 2: To identify unique PRF phenotypes using clinical and biochemical markers that
are readily available in the postoperative phase and determine if these markers predict PRF within 48 hours.
Hypotheses: (H2.1) Readily available clinical and biochemical biomarkers (e.g., mean arterial pressure,
creatinine) previously associated with hypo- and hyper-inflammatory acute respiratory distress syndrome and
acute respiratory failure phenotypes are also present in PRF. (H2.2) These clinical and biochemical markers
can be used to predict the probability of PRF within the next 48 hours. AIM 3: To develop and deploy a
single-site, proof-of-concept, EHR-based probability scoring system, and dashboard for PRF. Hypotheses:
(H3.1) Despite the benefits of the OMOP Common Data Model (CDM), data mapping into the CDM may cause
information loss and decrease the predictive performance of a CDM-mapped model compared to the native,
site-specific EHR model. (H3.2) The feasibility of a multisource (e.g., real-time and historic clinical and
biomarker data) probability score, embedded in the EHR, will be demonstrated through successful deployment
in a pre-production environment. Completing these Aims, and the five papers we foresee producing from this
work will enable me to develop preliminary data for a competitive R01 proposal focused on implementing and
evaluating a validated, real-time PRF predictive model in a UC-wide multi-center study. My long-term goal is to
expand my existing program of research to enroll more geographically, epidemiologically, and
socioeconomically diverse centers and conduct a large-scale, multisite intervention study (U grant) to validate
our modeling and facilitate personalized treatment strategies to reduce the risk and burden of PRF.
