In-hospital physical activity trends as an early warning sign for postsurgical complications after major oncological surgery.

Rapid digitalization and advances in data analysis open new windows for the use of real-time data and therefore data driven decision-making from descriptive and diagnostic analyses towards predictive analyses. Using physical activity data measured with an accelerometer would allow researchers and clinicians to use trends of physical activity levels after surgery as a potential tool for predicting complications. We evaluated whether physical activity trends can be used as an early warning sign for predicting complications during hospital stay after major abdominal oncological surgery.

Retrospective clinical data was collected of participants after surgery until discharge, including physical activity levels (PAM AM400 accelerometer), complications information, and surgical and demographic data. The patients included underwent major oncological surgery (liver, esophagus, stomach, colon, and bladder) in the University Medical Center Utrecht, the Netherlands between January 2021 to August 2023. Primary analyses included metrics of three machine learning models (1. Raw data; 2. Interpolation for non-wearing days; 3. Balanced class adjustment) using the trend features from the patient's total activity to predict a complication on the next day. Secondary analyses included a distinction between progressive and immediate complications. Area Under the Receiver Operating Characteristics (AUC ROC), precision, sensitivity and specificity of predicting a complication on the next day were calculated for both primary and secondary analyses.

A total of 254 participants were included for analysis, of which in 86 participants a complication was recorded. Participants wore the PAM accelerometer for an average of 10.1 days and were admitted for an average of 14.3 days. Complication participants had higher ASA score and higher BMI, but not overall less levels of physical activity. For the raw data model, the changes in the physical trends were not a good predictor on complication the next day on the (ROC AUC: 0.525). Analysis of the data after linear interpolation of the missing dates showed similar prediction outcomes (ROC AUC: 0.536) while analysis of a model trained in balanced complication -non complication dataset and tested in the raw dataset showed also similar prediction outcomes (ROC AUC: 0.510). Secondary analyses showed better prediction outcomes when the complication was not expected (ROC AUC: 0.605) compared to complications that were suspected by clinicians (ROC AUC; 0.546).

This is the first study that evaluates the use of in-hospital physical activity trends from usual care to predict complications. Our results show low predictive strength of these features in predicting all complications after major oncological surgery. Improved monitoring and complication description could provide better predictive data to improve clinical usefulness.

Further research is necessary to investigate the potential of physical activity trends as an added value to current EWS after surgery. This is the first study to use physical activity trends data from usual care data to predict post-surgical complications.