Overcoming the data barrier: transfer learning for 90-day mortality prediction in general surgery - a retrospective...

Area of Science:

• Surgical oncology
• Medical artificial intelligence
• Machine learning in healthcare

Background:

• Preoperative risk stratification is crucial for optimizing surgical outcomes and patient decision-making in general surgery.
• Data scarcity poses a significant challenge for developing high-dimensional artificial intelligence (AI) models for surgical applications.
• Transfer learning (TL) offers a solution by enabling knowledge transfer from pre-trained models to new, data-limited surgical domains.

Purpose of the Study:

• To evaluate the efficacy of transfer learning (TL) in enhancing the performance of artificial intelligence (AI) models for predicting 90-day mortality in general surgery.
• To benchmark TL models against conventional machine learning (ML) models and established risk scores.

Main Methods:

• A multicenter study included 14,922 patients undergoing advanced general surgery.
• Large-scale source models were trained on 85 preoperative parameters for mortality prediction.
• Organ-specific fine-tuning was applied for esophageal, liver, pancreatic, and colorectal surgeries.
• TL models were compared to standard ML models and conventional risk scores using AUROC, AUPRC, and F1-score.

Main Results:

• Transfer learning (TL) significantly improved Area Under the Precision-Recall Curve (AUPRC) by 38% for esophageal, 14% for liver, and 8% for pancreatic surgery.
• Patient age and Charlson Comorbidity Index (CCI) were consistently high-weight features across TL models.
• All neural networks (NNs) developed using TL outperformed the ASA physical status and CCI in mortality prediction.

Conclusions:

• Machine learning models demonstrate superior performance compared to conventional methods for preoperative mortality prediction.
• Transfer learning effectively addresses data limitations in surgical AI, significantly boosting model performance.
• TL presents a promising strategy for overcoming data constraints in the development of AI for surgery.