A Machine Learning Model to Predict Complications Following Head and Neck Microvascular Surgery.

Abstract

Objectives: Microvascular free tissue transfer is the primary reconstructive method in patients with complex defects of the head and neck. Defect site complications vary; incidence ranges from 15% to 70%. The use of machine learning via mutual information analysis has recently been described in the context of head and neck cancer, but this does not distinguish between causal relationships. The current study explores the first causal machine learning model to predict surgical outcomes in patients undergoing head and neck microvascular reconstruction.

Methods: A retrospective chart review was conducted at a single institution from 2022 to 2025. A causal machine learning analysis using Peter-Clark algorithmic discovery, external knowledge inputs, and large language models was performed to determine predictors of complications.

Results: The dataset included 155 subjects with 116 features. These 116 features were generated from the extracted clinical variables after preprocessing, categorical encoding, numerical transformation, and derivation of model-ready analytic features. Flap failure occurred in 3.2% (n = 5) of patients, and defect site complications occurred in 31.6% (n = 49). Mutual information analysis identified osteoradionecrosis/chondroradionecrosis, preoperative albumin, and peripheral vascular disease as the most consistently informative features. Lower preoperative hemoglobin and albumin were identified as potential causal predictors in the structural model.

Conclusions: Causal machine learning analysis is a powerful tool that assists in the prediction of outcomes for patients undergoing head and neck free flap reconstruction and can identify potentially modifiable preoperative factors to aid in risk assessment.

Level of evidence: 3.