Physical activity trends as predictors of postoperative complications in oncology patients: A machine learning approach

Comprehensive Summary

This study was conducted by Miguel Lorente et al. where they explored artificial intelligence (AI) and machine learning (ML) to assess whether complications may arise after oncological surgery through an accelerometer to collect data for perioperative physical activity. One wearing day would consist of patients doing physical activity for at least five minutes; if this value was below five minutes, they were not considered a wearing day. The participants were selected due to factors such as LR (interpretable data), RF (robust nonparametric ensemble), and XGB (high-performance gradient boosting model). From 965 screened patients, only 189 met strict accelerometer data quality criteria, where 19.8% experienced postoperative complications; patients with and without complications of whom were comparable in baseline demographics and co-existing conditions. Those with complications had fewer activity days and a longer hospital length of stay, but no differences in early postoperative physical activity levels during the first seven days. Predictive modeling using accelerometer-derived data showed only small discrimination, with improved recall for random forest and XGBoost at the cost of low precision and F1-scores. Overall, the study shows that postoperative physical activity trends alone provide only modest predictive value for complications after major oncological surgery. While moderate undersampling improved sensitivity in some models, this came at the cost of low precision and needing to use multimodal approaches.

Outcomes and Implications

This research is important because it finds out whether routinely collected postoperative activity data can help identify patients at risk for complications after oncological surgery. Clinically, the findings show that activity data alone is not sufficient but could contribute to future decision-support systems when combined with other clinical variables.