Development of a Novel Interpretable Transformer-Based Deep Learning Model for Predicting Postoperative Hypokalemia...

Area of Science:

• Endocrinology
• Neurosurgery
• Artificial Intelligence in Medicine

Background:

• Postoperative hypokalemia is a common complication after pituitary adenoma surgery.
• Early identification of hypokalemia is crucial for improving patient outcomes.
• Predictive models can aid in timely clinical interventions.

Purpose of the Study:

• To develop an interpretable predictive model for early postoperative hypokalemia in pituitary adenoma patients.
• To identify significant preoperative predictors of hypokalemia.
• To compare a Transformer-based model with traditional machine learning models.

Main Methods:

• Retrospective cohort study of 280 pituitary adenoma patients.
• Development of a Transformer-based model (CliTab-Transformer).
• Comparison with XGBoost and Multilayer Perceptron (MLP) models.
• Evaluation using accuracy, F1 score, sensitivity, and AUC with cross-validation.
• Interpretability analysis using a novel Transformer-Explainability method.

Main Results:

• CliTab-Transformer demonstrated superior performance over XGBoost and MLP.
• Key predictive factors identified include gender, hypertension, serum potassium, and disease duration.
• The model achieved higher accuracy (0.836) and F1 score (0.845) compared to benchmarks.

Conclusions:

• The Transformer-based model offers improved prediction of postoperative hypokalemia in pituitary adenoma patients.
• Attention mechanisms effectively capture complex feature interactions in clinical data.
• The model provides interpretable and clinically relevant insights for early intervention.