Telemedicine Notifications With Machine Learning for Postoperative Care (ODIN-Report)

The ODIN-Report study will be a randomized controlled trial of the effect of providing machine learning risk forecasts to providers caring for patients immediately after surgery on serious complications. The complications studied will be ICU admission or death on wards, acute kidney injury, and hospital length of stay.

This will be a single center, randomized, controlled, pragmatic clinical trial. The investigators will screen surgical patients enrolled in TECTONICS (NCT03923699) and randomized to intraoperative contact. Near the end of the operation, the investigators will calculate the same machine learning risk forecasts of major complications as TECTONICS, and enroll patients if all of the following are true: (1) No ICU admission is intended (2) ML mortality risk forecast is in top 15% of historical PACU patients. Patients will be randomized 1:1:1 to no contact, brief contact, and full contact. The postoperative provider (PACU physician, anesthesiologist, ward clinician) will be notified before arrival of the risk forecast in the contact groups, and in the full contact group an additional set of explanatory ML outputs will be provided. The intention-to-treat principle will be followed for all analyses.

Surgery--Complications, Perioperative/Postoperative Complications, Acute Kidney Injury, Hospital Mortality

Telemedicine, Anesthesia Control Tower, Machine Learning, Forecasting Algorithms, Randomized Controlled Trial, PACU

1:1:1 randomization between standard of care (no contact), postoperative contact (brief), postoperative contact (long).

Participants in the non-contact group will be monitored by anesthesia control tower clinicians who will utilize AlertWatch and integrating machine-learning forecasting algorithms for adverse outcomes predictions, but who will not contact the postoperative provider unless it is clinically necessary for patient safety purposes.

PACU and ward providers caring for participants in the brief contact group will be notified by Anesthesia Control Tower clinicians before arrival if the patient's forecast for mortality is in the top 15% of historical PACU patients. The notification will contain a brief summary of the patient's forecast risk of major adverse events.

PACU and ward providers caring for participants in the full contact group will be notified by Anesthesia Control Tower clinicians before arrival if the patient's forecast for mortality is in the top 15% of historical PACU patients. The notification will contain a report card of the patient's forecast risk of major adverse events, explanatory machine-learning outputs, most influential pre- and intraoperative data, and predicted treatments.

Real-time data will be monitored through the AlertWatch system as well as the electronic health record. Risk forecasts of adverse events (30 day mortality, acute kidney injury, postoperative delirium, respiratory failure), PACU length of stay, and hospital length of stay will be generated by a machine learning algorithm. Additional outputs identifying the most important predictors and their effects will be combined with risk forecasts to form a report card.

Admission to a "critical care" bed regardless of rationale or duration at any point in the follow up time frame. Patients who expire without transfer to ICU will be marked as positive.

Postoperative laboratory values and urine output will be used to calculate Kidney Disease Improving Global Outcomes grades of acute kidney injury. Where unavailable, baseline Glomerular filtration rate will be assumed to be age, sex, and body size normal.

Inclusion Criteria: Enrolled in TECTONICS Study (ID 201903026, NCT03923699), in OR randomized to contact workweek hours preoperative assessment completed estimated risk of mortality in top 15% of historical PACU patients Exclusion Criteria: Not enrolled in TECTONICS Study Operating room randomized to non-contact in TECTONICS Planned ICU admission