Abstract:

Objective

To develop a dose prediction model for ultrasound-guided thrombin treatment of pseudoaneurysms using the Light Gradient Boosting Machine （LightGBM） algorithm.

Methods

A retrospective analysis was conducted on 84 patients diagnosed with femoral artery pseudoaneurysms via ultrasound and treated with ultrasound-guided thrombin injection at the First Affiliated Hospital of Nanjing Medical University between January 2018 and December 2024.Patients were categorized into three groups based on thrombin dosage：low-dose （＜500 IU， 30 cases）， mediumdose （≥500 IU and ＜1000 IU， 36 cases）， and high-dose （≥1000 IU， 18 cases）.The cohort was randomly divided into a training set （67 cases） and a validation set （17 cases） at an 8：2 ratio.Feature variables were screened using ordinal logistic regression analysis to construct a logistic-based thrombin dose prediction model.Additionally， LightGBM contribution-based feature selection was applied to build a LightGBM-based dose prediction model.Model performance was evaluated using overall accuracy，micro-average area under the curve （AUC）， recall， F1-score， and receiver operating characteristic（ROC） curve analysis.

Results

In the training set， the logistic regression model demonstrated an overall accuracy of 0.677 and a micro-average AUC of 0.744 （95% confidence interval ［CI］：0.674-0.815）； in the validation set， the corresponding values were 0.686 and 0.758 （95%CI：0.624-0.891）.The LightGBM-based model exhibited superior performance， with a training set overall accuracy of 0.930，micro-average AUC of 0.975 （95%CI：0.955-0.995）， micro-average recall of 92.5%， and micro-average F1-score of 0.899.In the validation set， it achieved an overall accuracy of 0.804， micro-average AUC of 0.872 （95%CI：0.766-0.978）， micro-average recall of 76.5%， and micro-average F1-score of 0.722.

Conclusion

The LightGBM-based thrombin dose prediction model effectively forecasts thrombin dosage requirements， offering a valuable reference for achieving precision and individualized treatment in ultrasound-guided thrombin injection therapy.

Key words: Ultrasonography, interventional, Pseudoaneurysm, Thrombin, LightGBM, Predictive model, Machine learning