Developing and optimizing a machine learning predictive model for post-thrombotic syndrome in a longitudinal cohort of patients with proximal deep venous thrombosis.

BACKGROUND

Post-thrombotic syndrome (PTS) is the most common chronic complication of deep venous thrombosis (DVT). Risk measurement and stratification of PTS are crucial for patients with DVT. This study aimed to develop predictive models of PTS using machine learning for patients with proximal DVT.

METHODS

Herein, hospital inpatients from a DVT registry electronic health record database were randomly divided into a derivation and a validation set, and four predictive models were constructed using logistic regression, simple decision tree, eXtreme Gradient Boosting (XGBoost), and random forest (RF) algorithms. The presence of PTS was defined according to the Villalta scale. The areas under the receiver operating characteristic curves, decision-curve analysis, and calibration curves were applied to evaluate the performance of these models. The Shapley Additive exPlanations analysis was performed to explain the predictive models.

RESULTS

Among the 300 patients, 126 developed a PTS at 6 months after DVT. The RF model exhibited the best performance among the four models, with an area under the receiver operating characteristic curves of 0.891. The RF model demonstrated that Villalta score at admission, age, body mass index, and pain on calf compression were significant predictors for PTS, with accurate prediction at the individual level. The Shapley Additive exPlanations analysis suggested a nonlinear correlation between age and PTS, with two peak ages of onset at 50 and 70 years.

CONCLUSIONS

The current predictive model identified significant predictors and accurately predicted PTS for patients with proximal DVT. Moreover, the model demonstrated a nonlinear correlation between age and PTS, which might be valuable in risk measurement and stratification of PTS in patients with proximal DVT.