A risk prediction model for venous thromboembolism in hospitalized patients with thoracic trauma: a machine learning, national multicenter retrospective study

Table 1 Prediction performance of the machine learning models in the training and testing sets

Model	LR	RF	SVM	MLP	GBM
Train_auc	0.798	0.997	1	0.757	0.843
Test_auc	0.815	0.879	0.739	0.728	0.838
auc_ci	(0.788,0.842)	(0.856,0.902)	(0.708,0.77)	(0.697,0.759)	(0.812,0.864)
Specificity	0.734	0.687	0.589	0.684	0.899
Specificity_ci	(0.702,0.766)	(0.653,0.721)	(0.553,0.624)	(0.65,0.718)	(0.877,0.92)
Sensitivity	0.8	0.956	0.822	0.756	0.644
Sensitivity_ci	(0.689,0.911)	(0.889,1.0)	(0.711,0.933)	(0.622,0.867)	(0.511,0.778)
F1	0.261	0.27	0.193	0.219	0.392
Youden Index	0.534	0.642	0.411	0.439	0.544
MCC	0.273	0.314	0.193	0.216	0.374
Kappa	0.182	0.19	0.101	0.133	0.339
npv	0.984	0.996	0.982	0.979	0.976
npv_ci	(0.973,0.993)	(0.99,1.0)	(0.968,0.993)	(0.965,0.99)	(0.964,0.987)
ppv	0.156	0.158	0.109	0.128	0.282
Ppv_ci	(0.113,0.203)	(0.117,0.201)	(0.077,0.145)	(0.09,0.169)	(0.194,0.369)
plr	3.011	3.049	1.998	2.39	6.392
nlr	0.272	0.065	0.302	0.357	0.395
mAP	0.597	0.717	0.636	0.565	0.64

RF (0·879, 95% CI 0·856–0·902) had the best curve performance. In terms of the negative predictive rate and mAP, the RF (npv = 0·996 95% CI 0·99–1·0, mAP = 0·717) also achieves the best performance. The ROC curve (Fig. 2. A), precision‒recall curve (Fig. 2. B), calibration curve (Fig. 2. C), and DCA decision curve (Fig. 2. D) for all the candidate models in the test set are shown in Fig. 2. On the basis of these results, RF was selected as the final model for subsequent external validation

ISSN: 1749-7922