Pilot Study for Postoperative Machine Learning

Although surgery and anesthesia have become much safer on average, many patients still experience complications after surgery. Some of these complications are likely to be avoided or less severe with early detection and treatment. Barnes-Jewish Hospital has recently started using an Anesthesia Control Tower (ACT), which is a remote group lead by an anesthesiologist who reviews live data from BJH operating rooms and calls the anesthesia provider with concerns to improve reaction times and improve use of best-practices treatments. The ACT also uses machine learning (ML) to calculate patient risks during surgery as a way of measuring when the patient is doing better or worse.

The study team suspects that two mechanisms may allow risk prediction to improve postoperative care. First, is that it may make some data more actionable to clinicians. Although intraoperative data is extremely rich with many monitors, drug-response events, and surgical stress reactions to reveal the physiological state of the patient, that data is also extremely specialized and difficult to access. The study team thinks that many times the right interpretation of intraoperative data or the right treatment to give isn't clear until the surgery is nearly finished. The medical team in the recovery room (post-anesthesia care unit, PACU) and surgical wards is responsible for deciding the treatment strategy, but they don't have access to the information from the intraoperative monitors and events. Those providers also lack the familiarity to directly interpret that information and time to review it in detail. Even preoperative information may be less than fully available because the patient may still be too sedated or confused from the anesthesia to explain much about their history. By summarizing these diverse sources of information into a risk profile, machine learning outputs may directly improve the understanding of postoperative providers or improve the identification of patients at elevated risk for postoperative adverse outcomes.

A second mechanism derives from behavior changes which may occur in providers in reaction to machine-generated risk profiles. The study team has observed many handoffs from the operating room and PACU include lists of "important" data, but it is common for the handoff-giver to provide no interpretation (what problem is this information related to) or anticipatory guidance (having identified a potential or actual problem, what should the handoff receiver do). The study team has also observed than once a major risk has been clearly identified along the chain of handoff it tends to be propagated forward with connection to the underlying data, any changes noticed by the current provider, and the current plan. The study team suspects that in the subset of patients with substantially elevated predictions on their risk profile, handoff communication and team coordination for the identified problems may improve.

The larger goal is to deploy a "report card" for each patient that summarizes the preoperative assessment and intraoperative data in a way that is useful for postoperative providers. In this study these ML reports will be integrated into the clinical workflow and determine if it does affect handoff behavior. The study team will also evaluate the information-effect and test the report card for safety by determining if clinicians identify any major inaccuracies related to the implementation.

This study is a substudy of a randomized trial of ACT-intraoperative contact (TECTONICS IRB# 201903026), and only patients in the contact (treatment) group will be eligible. The screened patients will be all adults having surgery at BJH with the division of Acute and Critical Care Surgery. Exclusion criteria are a planned ICU admission. For each included patient, the ACT clinician will review the report card information, and the postoperative providers will either be directly contacted or receive an Epic Best Practices Advisory. Our study will be a before-after quasi-experiment, meaning that after a fixed date, all eligible patients will receive the intervention, and the outcome measures will be compared to patients before that date. The outcome measure we will study is handoff effectiveness from the recovery room to wards. Providers will be surveyed on information value, any inaccurate items, or major omissions.

The ML report card will not recommend specific treatments, and decisions will remain the hands of the physician in the PACU or wards. The postoperative provider will also be given information about the report card and its limitations.

Modifications during piloting We originally intended the report data to be included in the electronic medical record hand off workflow for the pacu receiving nurse, however, due to changes in institutional priorities this was not activated.

The study was originally designed as a pre- versus post-intervention comparison. However, it was apparent during pre-intervention data collection (June 3 2021 - July 22 2021) that differences in individual research-assistant's evaluation of hand-off content and changes in other behavior would make this a low validity method. After 2 months, the design was changed to a non-randomized parallel group study with alternating days with the report card turned on or off in a 2:1 ratio. We also took this opportunity to expand the inclusion criteria to include patients with vascular surgery, as they were frequently identified as high-risk.

Due to a lack of staff availability, no further enrollment was attempted until October 2022. Enrollment continued from Oct 11 2022 to May 11 2023, with no enrollment in Jan or Feb 2023 due to staff availability.