Purpose
Postoperative flap monitoring has always been a critical part of microvascular reconstructive surgery. However, the need for specially trained flap-monitoring medical staff and care unit limits its wide application of this treatment option. In addition to the traditional approach, many studies have proposed new monitoring methods based on novel technologies, but it is still difficult to replace the traditional ways because of the cost and learning curve. Through this study, we aimed to use machine learning (ML) to create a new tool for postoperative monitoring.

Material and Method
We collected observation data from the patient receiving free flap transfer from Micro ICU in CGMH from 2020/12 to 2021/7. Each observation data contained objective measurements, such as flap photos and surface temperature, and subjective evaluation, which was obtained with the traditional approach, such as turgor of skin paddle of the flap, capillary refilling time, puncture bleeding time, and puncture bleeding color. We compared the color difference of the flap photos with pixel analysis. The outcome variable is the condition of the flap at the moment of flap observing, which were labeled as a normal flap, a flap with arterial insufficiency, or a flap with venous insufficiency.
Through the supervised machine learning approach, the data were divided into training and testing sets. The predictive ability of different machine learning algorithms (K-Nearest Neighbors, Decision Tree, Random Forest, and AdaBoost) were evaluated. To overcome data imbalances, random over sampling example (ROSE) method was used. Prediction accuracy and predictive power (AUC) were measured.

Result
A total of 805 data were collected. The flap's surface temperature and the color difference between the flap and surround skin, turgor of flap skin paddle, and puncture bleeding color were highly correlated to the flap condition. The prediction accuracy score ranged from 96% to 98%, and the AUROC ranged from 0.97 to 0.99 by using the above algorithms. The model based on K-Nearest Neighbors performed the most effective prediction.

Conclusion
Our results showed ML models could make precise predictions whether the flap has vascular compromised, which is greatly helpful for postoperative free flap monitoring and decision making. The clinical application of these models will reduce the burden and encourage microsurgeon to perform microvascular reconstructive surgery more actively.

Download