Application of machine learning models in predicting the risk of thromboembolic events in patients with nonvariceal gastrointestinal bleeding

Table 1 Characteristics of patients in the thromboembolism group and non-thromboembolism group, mean ± SD

Variable	Thromboembolism group (n = 294)	Non-thromboembolism group (n = 854)	P value
Age (years)	73.10 ± 11.55	64.85 ± 16.97	< 0.001
Gender (male) (%)	64.07	69.56	0.088
Heart rate (beats/minute)	83.21 ± 15.18	86.92 ± 18.87	0.002
Hemoglobin (g/L)	80.15 ± 27.47	90.01 ± 30.61	< 0.01
Albumin (g/L)	3306 ± 6.62	34.91 ± 9.87	0.003
Intensive care unit admission (%)	22.37	10.77	< 0.001
History of anticoagulant drug use (%)	58.64	25.41	< 0.001
D-dimer level (μg/L)	4177.46 ± 8129.90	2114.73 ± 6339.46	< 0.001
Length of hospital stay (days)	12.90 ± 9.52	8.57 ± 6.88	< 0.001
Use of hemostatic drugs (%)	10.17	11.82	0.04
Thromboembolism history (%)	35.25	10.42	< 0.001
Alanine aminotransferase (U/L)	23.19 ± 50.04	23.55 ± 64.62	0.931
Creatinine (μmoI/L)	121.59 ± 123.56	115.05 ± 144.02	0.486
International normalized ratio	5.57 ± 6.62	1.17 ± 0.64	0.044
Prothrombin time (seconds)	14.03 ± 9.73	13.56 ± 6.97	0.376
History of nonsteroidal anti-inflammatory drug use (%)	15.93	17.68	0.493
Shock (%)	10.17	10.77	0.772
Red cell distribution width (%)	15.56 ± 2.99	15.09 ± 5.04	0.13
Education (%)1	9.52	15.69	0.09

¹Proportion of individuals with education level of high school or above.

Table 2 Characteristics of machine learning models in the internal validation sets, mean (95%CI)

Model	Accuracy	Precision	Sensitivity	Specificity	F1	Area under the receiver operating characteristic curve	P value
L1 regularized logistic regression	0.736 (0.697-0.771)	0.43 (0.362-0.501)	0.716 (0.628-0.79)	0.741 (0.698-0.781)	0.537	0.793 (0.750-0.837)	< 0.01
Support vector machines	0.701 (0.661-0.738)	0.401 (0.340-0.465)	0.802 (0.72-0.864)	0.673 (0.627-0.716)	0.534	0.804 (0.757-0.851)	< 0.01
Categorical boosting	0.754 (0.716-0.789)	0.455 (0.386-0.526)	0.75 (0.664-0.82)	0.755 (0.712-0.794)	0.567	0.818 (0.777-0.859)	< 0.01
Random forest	0.678 (0.638-0.716)	0.38 (0.321-0.443)	0.793 (0.711-0.857)	0.647 (0.6-0.691)	0.514	0.798 (0.755-0.842)	< 0.01
Extreme gradient boosting	0.71 (0.67-0.746)	0.402 (0.338-0.47)	0.724 (0.637-0.797)	0.706 (0.661-0.747)	0.517	0.772 (0.723-0.821)	< 0.01
D-dimer	0.621 (0.579-0.661)	0.309 (0.253-0.371)	0.621 (0.530-0.704)	0.621 (0.574-0.666)	0.413	0.618 (0.552-0.683)	-

P value reflected the comparisons between each machine learning model and D-dimer.

Table 3 Characteristics of machine learning models in the external validation sets, mean (95%CI)

Model	Accuracy	Precision	Sensitivity	Specificity	F1	Area under the receiver operating characteristic curve	P value
L1 regularized logistic regression	0.78 (0.728-0.825)	0.395 (0.296-0.504)	0.711 (0.566-0.823)	0.793 (0.737-0.84)	0.508	0.805 (0.735-0.875)	< 0.01
Support vector machines	0.77 (0.717-0.815)	0.389 (0.295-0.492)	0.778 (0.637-0.875)	0.768 (0.71-0.817)	0.519	0.806 (0.727-0.884)	< 0.01
Categorical boosting	0.826 (0.778-0.866)	0.466 (0.343-0.592)	0.6 (0.455-0.73)	0.869 (0.82-0.906)	0.524	0.815 (0.746-0.885)	< 0.01
Random forest	0.738 (0.683-0.785)	0.344 (0.255-0.445)	0.711 (0.566-0.823)	0.743 (0.683-0.794)	0.464	0.804 (0.736-0.872)	< 0.01
Extreme gradient boosting	0.727 (0.672-0.776)	0.318 (0.23-0.421)	0.622 (0.476-0.749)	0.747 (0.688-0.798)	0.421	0.746 (0.661-0.831)	< 0.01
D-dimer	0.734 (0.68-0.782)	0.258 (0.166-0.379)	0.356 (0.232-0.502)	0.806 (0.751-0.851)	0.299	0.51 (0.403-0.617)	-

P value reflected the comparisons between each machine learning model and D-dimer.