Machine Learning Model to Predict Postmastectomy Breast Reconstruction Complications

Postmastectomy breast reconstruction (PMBR) improves patients' quality of life, but patients often lack reliable, individualized information about complication risk. Machine learning (ML) can analyze complex clinical data to generate personalized risk estimates, facilitating shared decision-making.

To develop and validate ML models trained on both structured data and manually abstracted variables from unstructured clinical notes to predict major complications after PMBR.

This prognostic study used retrospective data from female patients aged 18 years or older who underwent unilateral or bilateral therapeutic mastectomy with immediate or delayed implant-based or autologous reconstruction at 2 academic centers in the US from 2012 to 2022. Demographic, treatment, and surgical variables were extracted from electronic health records with a 1-year postoperative follow-up period. Extreme gradient boosting (XGBoost) and random forest models were trained on 80% of the cohort (329 individuals) and tested on 20% of the cohort (82 individuals). Patients with bilateral prophylactic mastectomy, distant metastases, mixed autologous and implant-based reconstruction, or less than 12 months of follow-up were excluded. Of more than 4000 eligible patients, a random sample of 411 underwent manual health record review for variable abstraction. Data were analyzed from September through November 2024.

PMBR.

Outcomes of interest were major complications, defined as unplanned reoperations or rehospitalizations within 1 year of reconstruction. Model performance was evaluated using the area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC).

The sample included 411 female patients (667 breasts) receiving implant-based (290 individuals [70.6%]) or autologous (121 individuals [29.4%]) PMBR, with a median (IQR) age of 51.3 (44.0-58.3) years. The overall major complication rate was 25.8% (106 individuals). The XGBoost model outperformed the random forest model, achieving an AUROC of 0.83 (95% CI, 0.72-0.94) and an AUPRC of 0.62 (95% CI, 0.55-0.69; baseline: 0.26) on the test set, compared with 0.74 (95% CI, 0.66-0.82) and 0.56 (95% CI, 0.50-0.62), respectively, for the random forest model. Top predictors of major complications included smoking, adjuvant radiotherapy, body mass index, age, and diabetes. Model performance remained consistent across reconstructive modalities.

In this prognostic study of PMBR outcomes, an internally validated ML model trained on both structured and unstructured clinical data was used to predict 1-year major complications. Such models support personalized risk assessment, inform decision-making, and provide a foundation for future externally validated and prospectively tested decision-support tools.