iROBOT: Implementing Real-time Operational dashBOards for Trauma care

Abstract

Background: Trauma care across British Columbia (BC) is
provided by regional health systems. Island Health (ISLH),
which serves Vancouver Island (population of 864, 000), is one of seven health authorities reporting to the BC Trauma Registry (BCTR). The BCTR data are collected by registrars after patient discharge, resulting in a structural lag to data. Our objective is to accelerate access to data by building a real-time dashboard for trauma patients who may be included in the BCTR.

Methods: We matched BCTR cases to ISLH visit-level data, to retrospectively develop a decision-tree classifier that identified patients who will be admitted to trauma services. We then validated this classifier prospectively to determine the precision and recall of the system. We are now linking ISLH electronic data to matched BCTR data elements to develop an automated system for presenting the data needed for registry reporting, quality improvement, and eventual real-time surveillance of trauma.

Results: We retrospectively matched 97.45% (n = 5,989/6,146) of BCTR to ISLH cases between 2018 and 2022. Our classifier was 96% Sensitive, 93% specific in identifying cases who were admitted to trauma services using only three digital attributes (nodes), all of which are available within the first 24 hours of patient care.

The prospective validation of the decision-tree classifier occurred between June and August 2024 and examined 206,509 patients. When compared to the traditional approach, the classifier model identified more true positive (405 vs. 397) and true negative cases (206,031 vs. 201,773) and fewer false positive (52 vs. 4,310) and false negatives (21 vs. 29) than the standard approach. The classifier had higher sensitivity and specificity (95.07%, 99.97%, respectively) than the standard approach (sensitivity = 93.19, specificity = 97.91%) with dramatically higher precision scores, 88.62% versus 8.43%, respectively. Operationally, the model has resulted in a 90.16% reduction in screening time (18 min per week versus 183 min). Clinically, the model identified 8 trauma cases who were, incorrectly, not evaluated by trauma services. These cases were subsequently transferred to the care of trauma services.

We have incorporated our classifier into a real-time reporting dashboard and have successfully linked 32.9% (n = 53) data variables. We are adding additional data elements to the dashboard and anticipate that we will successfully link 80% (n = 76) of the remaining variables that are currently extracted manually.

Implications and lessons learned: Our classifier system identified trauma admissions with a high degree of precision and reduced screening time by more than 90%. We have successfully mapped 33% of registry data elements to a dashboard, we anticipate that 80% of the elements will eventually be mapped, allowing for semi-automatic data collection of most data reporting requirements. This Semi-automation will facilitate a shift from data extraction to data validation and will eventually be used to identify biological trauma cases for surveillance.