Abstract
Successful monitoring of deep vein thrombosis (DVT) remains a challenging problem after gynecological laparoscopy. Thus, this study aimed to create and validate predictive models for DVT with the help of machine learning (ML) algorithms. A total of 489 patients from the Cancer Biology Research Center, Tongji Hospital were included in the study between January 2017 and February 2023, and 35 clinical indicators from electronic health records (EHRs) were collected within 24h of patient admission. Risk factors were identified using the least absolute shrinkage and selection operator (LASSO) regression. Then, the three commonly used DVT prediction models are random forest model (RFM), generalized linear regression model (GLRM), and artificial neural network model (ANNM). In addition, the predictive performance of various prediction models (i.e. the robustness and accuracy of predictions) is evaluated through the receiver operating characteristic curve (ROC) and decision curve analysis (DCA), respectively. We found postoperative DVT in 41 (8.38%) patients. Based on the ML algorithm, a total of 13 types of clinical data were preliminarily screened as candidate variables for DVT prediction models. Among these, age, body mass index (BMI), operation time, intraoperative pneumoperitoneum pressure (IPP), diabetes, complication and D-Dimer independent risk factors for postoperative DVT and can be used as variables in ML prediction models. The RFM algorithm can achieve the optimal DVT prediction performance, with AUC values of 0.851 (95% CI: 0.793–0.909) and 0.862 (95% CI: 0.804–0.920) in the training and validation sets, respectively. The AUC values of the other two prediction models (ANNM and GLRM) range from 0.697 (95% CI: 0.639–0.755) and 0.813 (95% CI: 0.651–0.767). In summary, we explored the potential risk of DVT after gynecological laparoscopy, which helps clinicians identify high-risk patients before gynecological laparoscopy and make nursing interventions. However, external validation will be needed in the future.
Keywords: deep vein thrombosis, gynecological laparoscopy, machine learning algorithms, prediction, risk factor
1. Introduction
In recent years, with the continuous improvement of surgical technology and surgical instruments and equipment, compared with the traditional laparotomy, gynecological laparoscopy has incomparable advantages such as less physical damage to patients, less stress reaction, less blood loss of patients, and faster recovery after surgery.[1,2] However, it should be noted that deep vein thrombosis (DVT), as one of the common serious complications of gynecological laparoscopy, causing delays in recovery short-term and possible impacts on quality of life long-term.[3] According to previous research reports, the risk of DVT after laparoscopic surgery is not lower than that of traditional gynecological surgery.[4,5] The most serious complication of DVT is pulmonary embolism, which is closely related to higher mortality and poor prognosis. In addition, the formation of DVT can have irreversible effects on the patient’s limb motor function, thereby reducing their daily activity ability and quality of life.[6–8] Therefore, it is of great significance to understand the relevant risk factors of DVT after gynecological laparoscopy and guide the early prevention and intervention of clinical deep vein examination to reduce the incidence of DVT after gynecological laparoscopy.
Nowadays, with the rapid development of artificial intelligence, machine learning(ML), data mining theory and the improvement of computer computing power, under the guidance of the national “precision medicine” strategy, medical data mining has developed rapidly and become a research hotspot.[9,10] Undoubtedly, ML-based techniques are being increasingly adapted for use in the medical field because of their high accuracy.[11] Early prediction and prevention of DVT can effectively reduce patient pain while improving outcomes. Although research on the risk factors for DVT is prevalent, there is a stark lack of clinical predictive models for DVT occurrence specifically in patients after gynecological laparoscopy.
Given this situation, to investigate the applicability of ML-based techniques to predict DVT for patients after gynecological laparoscopy, this study aims to extract the electronic health records (EHRs) of patients received gynecological laparoscopy using different ML-based algorithms, to develop a predictive model to assist in better clinical decision-making and optimal allocation of medical resources.
2. Materials and methods
2.1. Study population
We retrospectively selected patients who performed laparoscopy in the department of gynecology of Tongji Hospital from January 2017 to February 2023. Inclusive criteria: patients who received gynecological laparoscopy for the first time; Patients who underwent preoperative venous color ultrasound examination for DVT; Patients who have not used contraceptives, other hormone medications, or medications that may affect their coagulation function before surgery; Patients who are informed and agree to participate in this study. Exclusion criteria: Patients undergoing traditional gynecological open surgery; Patients taking anticoagulant drugs before surgery; Patients with abnormal coagulation function, such as hemophilia, leukemia, etc; Patients who are unwilling to participate in this study. As this study belongs to a retrospective cohort study, according to the policy issued by the ethics review agency, patients can be exempted from informed consent forms. The ethics committee (No. TJ-IRB2023) encrypts the personal information of all patients included in this study to prevent leakage. The patients included in the study and the process of predicting the model were shown in Figure 1.
Figure 1.
The flow chart of patient selection and data process.
2.2. Diagnostic criteria for postoperative DVT
The diagnostic criteria for DVT are as follows: Clinical manifestations: postoperative pain, numbness, swelling, and limited movement of the affected side of the lower limb; Physical signs: The circumference of the affected lower limb is greater than that of the healthy side, and the homans sign is positive; Auxiliary examination: Color Doppler ultrasound examination of lower limb veins shows a large amount of echo in the venous cavity.
2.3. Acquisition of EHR-based clinical parameters
We mainly rely on the HIS medical record system of Tongji Hospital to collect relevant data, which are as follows: Preoperative data of patients, such as age, body mass index (BMI), education level, smoking history, hypertension history, diabetes history, D-dimer; Intraoperative data (surgical time, ASA grading, intraoperative blood transfusion); and Postoperative data (complications, such as bleeding, organ injury, subcutaneous emphysema, excessive blood loss, urinary system infection, etc.), postoperative bedridden days, etc.
2.4. Construction and performance evaluation of DVT prediction model
All patients were randomly assigned to a 70% training set and a 30% validation set. As for the predictive variables, we used the least absolute shrinkage and selection operator (LASSO) and Pearson correlation coefficient for variable selection. The final prediction model for DVT was based on three commonly used algorithms, namely: Random Forest Model (RFM), Artificial Neural Network Model (ANNM), and Generalized Linear Regression Model (GLRM).[12–14] In addition, the clinical predictive performance of various prediction models was evaluated using multi-dimensional prediction tools, including decision curve analysis (DCA), clinical impact curve (CIC), and area under the receiver operating characteristic (AUROC) curve.[15]
2.5. Statistical analysis
In this study, Wilcoxon rank sum test or t test was used for continuous variables, while chi square test was used to compare categorical variables. The correlation between two continuous variables was evaluated using Pearson correlation coefficient. The charts in this study were obtained using R Studio software (available from the following website: https://www.r-project.org/). A two tailed P value of < .05 was defined as having significant statistical significance.
3. Results
3.1. Analysis of general baseline data characteristics of patients
In this study, a total of 41 patients experienced postoperative DVT and were classified as the “DVT” group, while 448 patients who did not experience DVT were considered as the “Non-DVT” group. The incidence rate of postoperative DVT in the training and validation sets was 6.43% (22 out of 342) and 12.93% (19 out of 147), respectively. The types of diseases mainly included 124 cases of accessory surgery (including the treatment of ectopic pregnancy, salpingostomy, pelvic inflammation, Ovarian cyst and infertility examination, etc.), 286 cases of total hysterectomy, 143 cases of uterine fibroid, and 218 cases of laparoscopic gynecological malignant tumor surgery. The baseline data were summarized in Table 1.
Table 1.
The characteristics of included patients.
| Variables | Training cohort | P value | Testing cohort | P value | ||||
|---|---|---|---|---|---|---|---|---|
| Overall (N = 342) | non-DVT (N = 320) | DVT (N = 22) | Overall (N = 147) | non-DVT (N = 128) | DVT (N = 19) | |||
| Age (median [IQR]) | 42.50 [36.00, 48.00] | 42.00 [36.00, 47.00] | 54.50 [47.25, 59.75] | <.001 | 44.00 [38.00, 50.00] | 42.00 [38.00, 48.00] | 58.00 [50.00, 61.00] | <.001 |
| BMI (median [IQR]) | 23.15 [21.10, 24.70] | 22.80 [21.00, 24.42] | 28.55 [26.30, 29.80] | <.001 | 23.60 [21.45, 25.55] | 23.20 [21.20, 25.00] | 30.80 [28.35, 31.30] | <.001 |
| Degree (%) | ||||||||
| Illiteracy | 90 (26.3) | 86 (26.9) | 4 (18.2) | .484 | 30 (20.4) | 29 (22.7) | 1 (5.3) | .355 |
| Primary school | 71 (20.8) | 66 (20.6) | 5 (22.7) | 37 (25.2) | 32 (25.0) | 5 (26.3) | ||
| Junior high school | 82 (24.0) | 74 (23.1) | 8 (36.4) | 37 (25.2) | 31 (24.2) | 6 (31.6) | ||
| High school and above | 99 (28.9) | 94 (29.4) | 5 (22.7) | 43 (29.3) | 36 (28.1) | 7 (36.8) | ||
| Smoking (%) | ||||||||
| Yes | 117 (34.2) | 110 (34.4) | 7 (31.8) | .99 | 43 (29.3) | 38 (29.7) | 5 (26.3) | .975 |
| No | 225 (65.8) | 210 (65.6) | 15 (68.2) | 104 (70.7) | 90 (70.3) | 14 (73.7) | ||
| Diabetes (%) | ||||||||
| Yes | 100 (29.2) | 80 (25.0) | 20 (90.9) | <.001 | 51 (34.7) | 34 (26.6) | 17 (89.5) | <.001 |
| No | 242 (70.8) | 240 (75.0) | 2 (9.1) | 96 (65.3) | 94 (73.4) | 2 (10.5) | ||
| Hypertension (%) | ||||||||
| Yes | 96 (28.1) | 93 (29.1) | 3 (13.6) | .189 | 37 (25.2) | 32 (25.0) | 5 (26.3) | 1 |
| No | 246 (71.9) | 227 (70.9) | 19 (86.4) | 110 (74.8) | 96 (75.0) | 14 (73.7) | ||
| Disease type (%) | ||||||||
| Benign | 124 (36.3) | 113 (35.3) | 11 (50.0) | .247 | 45 (30.6) | 39 (30.5) | 6 (31.6) | 1 |
| Malignant | 218 (63.7) | 207 (64.7) | 11 (50.0) | 102 (69.4) | 89 (69.5) | 13 (68.4) | ||
| Nurse type (%) | ||||||||
| I | 240 (70.2) | 223 (69.7) | 17 (77.3) | .609 | 104 (70.7) | 89 (69.5) | 15 (78.9) | .568 |
| II | 102 (29.8) | 97 (30.3) | 5 (22.7) | 43 (29.3) | 39 (30.5) | 4 (21.1) | ||
| ASA (%) | ||||||||
| I | 182 (53.2) | 170 (53.1) | 12 (54.5) | 1 | 72 (49.0) | 63 (49.2) | 9 (47.4) | 1 |
| II | 160 (46.8) | 150 (46.9) | 10 (45.5) | 75 (51.0) | 65 (50.8) | 10 (52.6) | ||
| Operation time (%), h | ||||||||
| <1 | 236 (69.0) | 232 (72.5) | 4 (18.2) | <.001 | 93 (63.3) | 90 (70.3) | 3 (15.8) | <.001 |
| ≥1 | 106 (31.0) | 88 (27.5) | 18 (81.8) | 54 (36.7) | 38 (29.7) | 16 (84.2) | ||
| IBT (%) | ||||||||
| Yes | 97 (28.4) | 91 (28.4) | 6 (27.3) | 1 | 46 (31.3) | 43 (33.6) | 3 (15.8) | .195 |
| No | 245 (71.6) | 229 (71.6) | 16 (72.7) | 101 (68.7) | 85 (66.4) | 16 (84.2) | ||
| IPP (%), mm Hg | ||||||||
| ≤15 | 240 (70.2) | 234 (73.1) | 6 (27.3) | <.001 | 102 (69.4) | 97 (75.8) | 5 (26.3) | <.001 |
| >15 | 102 (29.8) | 86 (26.9) | 16 (72.7) | 45 (30.6) | 31 (24.2) | 14 (73.7) | ||
| Complication (%) | ||||||||
| Yes | 110 (32.2) | 92 (28.7) | 18 (81.8) | <.001 | 36 (24.5) | 21 (16.4) | 15 (78.9) | <.001 |
| No | 232 (67.8) | 228 (71.2) | 4 (18.2) | 111 (75.5) | 107 (83.6) | 4 (21.1) | ||
| PB (%), days | ||||||||
| ≤5 | 236 (69.0) | 231 (72.2) | 5 (22.7) | <.001 | 104 (70.7) | 97 (75.8) | 7 (36.8) | .001 |
| >5 | 106 (31.0) | 89 (27.8) | 17 (77.3) | 43 (29.3) | 31 (24.2) | 12 (63.2) | ||
| Blood type (%) | ||||||||
| A | 174 (50.9) | 168 (52.5) | 6 (27.3) | <.001 | 66 (44.9) | 60 (46.9) | 6 (31.6) | <.001 |
| B | 159 (46.5) | 152 (47.5) | 7 (31.8) | 76 (51.7) | 68 (53.1) | 8 (42.1) | ||
| AB | 3 (0.9) | 0 (0.0) | 3 (13.6) | 4 (2.7) | 0 (0.0) | 4 (21.1) | ||
| O | 6 (1.8) | 0 (0.0) | 6 (27.3) | 1 (0.7) | 0 (0.0) | 1 (5.3) | ||
| D-Dimer (%), mg/L | ||||||||
| <0.5 | 110 (32.2) | 93 (29.1) | 17 (77.3) | <.001 | 51 (34.7) | 35 (27.3) | 16 (84.2) | <.001 |
| ≥0.5 | 232 (67.8) | 227 (70.9) | 5 (22.7) | 96 (65.3) | 93 (72.7) | 3 (15.8) | ||
| PT (median [IQR]), s | 10.60 [9.80, 11.40] | 10.60 [9.78, 11.40] | 10.30 [9.90, 11.25] | .435 | 10.80 [9.70, 11.45] | 10.80 [9.78, 11.40] | 10.70 [9.50, 11.70] | .982 |
| PTA (median [IQR]), % | 99.00 [88.00, 112.00] | 99.00 [87.75, 112.00] | 98.50 [93.00, 112.50] | .758 | 100.00 [88.00, 115.50] | 99.50 [88.00, 114.25] | 106.00 [91.50, 118.00] | .393 |
| INR (median [IQR]) | 0.96 [0.85, 1.04] | 0.96 [0.85, 1.03] | 0.96 [0.83, 1.08] | .639 | 0.97 [0.90, 1.06] | 0.98 [0.90, 1.07] | 0.94 [0.84, 1.02] | .076 |
| Fibrinogen (median [IQR]), g/L | 2.72 [2.32, 3.25] | 2.69 [2.30, 3.22] | 3.14 [2.70, 3.30] | .046 | 2.83 [2.36, 3.28] | 2.83 [2.40, 3.28] | 3.16 [1.98, 3.45] | .806 |
| APTT (median [IQR]), s | 29.00 [26.00, 34.00] | 29.00 [26.00, 33.00] | 30.50 [25.50, 34.00] | .433 | 29.00 [25.00, 33.00] | 29.00 [25.00, 33.00] | 29.00 [26.00, 31.50] | .647 |
| TT (median [IQR]), s | 12.00 [10.00, 14.00] | 12.00 [10.00, 14.00] | 11.00 [10.00, 14.00] | .775 | 13.00 [11.00, 14.00] | 13.00 [11.00, 14.00] | 13.00 [9.00, 14.00] | .637 |
| Hb (median [IQR]), g/L | 101.00 [87.00, 114.75] | 100.50 [87.00, 115.00] | 102.00 [89.25, 113.25] | .963 | 101.00 [85.50, 114.00] | 102.00 [86.75, 113.25] | 101.00 [80.00, 121.00] | .926 |
APTT = activated partial thromboplastin time, ASA = American Society of Anesthesiologists, BMI = body mass index, Hb = hemoglobin, IBT = intraoperative blood transfusion, INR = international normalized ratio, IPP = intraoperative pneumoperitoneum pressure, IQR = inter-quartile range, PB = postoperative bedtime, PT = prothrombin time, PTA = prothrombin activity, TT = thrombin time.
3.2. Screening of predictive variables
To minimize the bias caused by non-normal distributions as much as possible, we set a reduction coefficient (i.e. penalty coefficient) to retain the weight of the predicted variables during the gradual screening process. Specifically, as shown in Figure 2, during the cross validation of candidate parameters, we input the optimal lambda(λ) of the LASSO regression model into the prediction model, which includes 13 prediction variables, with λ = 22. This was ultimately used to construct the DVT prediction model.
Figure 2.
Predictor variable selection based on the LASSO regression method. (a) Optimal parameter (lambda) selection in the LASSO model. (b) LASSO coefficient profiles of the candidate features.
3.3. Establishment and validation of GLRM
We conducted multivariable logistic regression analysis on the included candidate variables, and finally included age, BMI, operation time, IPP, diabetes, complications and D-dimer as the prediction variables based on Akaike information quasi test, as shown in Table S1, Supplemental Digital Content, http://links.lww.com/MD/L159 In the constructed nomogram prediction visualization graph (Fig. 3), we evaluated the risk of DVT according to the total score obtained from the assignment of each column of prediction variables. Based on bootstrap, the robustness of the GLRM prediction model can reach C-index = 0.697, indicating that the prediction model has ideal robustness and accuracy.
Figure 3.
Nomogram prediction model for predicting DVT. (A)Nomogram predicts risk of pulmonary infection; (B) The calibration curves for the nomogram. BMI = body mass index, IPP = intraoperative pneumoperitoneum pressure.
3.4. Constructing and developing DVT prediction model using machine learning algorithms
In this study, two sets of machine learning algorithms (i.e. ANNM and RFM) were used to develop and validate DVT models. As shown in Table S2, Supplemental Digital Content, http://links.lww.com/MD/L160 in the RFM prediction model, the weight values of age, BMI, operation time, IPP, diabetes, complications and D-dimer become candidate variables for predicting DVT based on RFM (Fig. 4A). Interestingly, in ANNM, age, BMI, operation time, IPP, diabetes, complications and D-dimer were also candidate variables for predicting DVT, and the weight and distribution of each variable given by two different groups of algorithms were inconsistent, so the advantages and disadvantages of the algorithm were also reflected in the subsequent prediction effectiveness evaluation (Fig. 4B, Table S3, Supplemental Digital Content, http://links.lww.com/MD/L161).
Figure 4.
Construction of DVT prediction model via (A) RFM and (B) ANNM. Notes. The application prediction model formula of RFM is as follows: C = argmax (∑ (Ci)), where Ci represents the type of In prediction for the i-th tree, C is the final classification result, and I is the number of trees. The formula of ANNM is as follows: θ = θ − η × ∇ (θ). J(θ) Among them η It’s the learning rate, so (θ). J(θ) Represents the gradient change of the loss function (i.e. J(θ)).
3.5. Performance of DVT prediction models
The ROC curve in this study showed that the RFM model had higher predictive power than the ANNM model in both training and validation sets, with area under the curve (AUCs) of 0.851 and 0.862, and 95% confidence intervals (CI) of 0.793–0.909 and 0.804–0.920, respectively. The AUCs of the ANNM model were 0.795 and 0.813, and 95% CI were 0.737–0.853 and 0.755–0.871, respectively. The predictive performance of clinical indicators before DVT surgery was shown in Table 2, which also indicated that the predictive performance of the DVT prediction model developed based on ML algorithm was significantly better than GLRM. In addition, the DCA curve in Figure 5 indicated that RFM was clearly more robust than the other two prediction models in predicting DVT.
Table 2.
Comparison of predictive efficacy of DVT prediction models via ROC curves.
| Model | Training set | Internal validation set | ||||
|---|---|---|---|---|---|---|
| AUC Mean | AUC 95% CI | Variables* | AUC Mean | AUC 95% CI | Variables* | |
| RFM | 0.851 | 0.793–0.909 | 7 | 0.862 | 0.804–0.920 | 7 |
| ANNM | 0.795 | 0.737–0.853 | 8 | 0.813 | 0.755–0.871 | 8 |
| GLRM | 0.697 | 0.639–0.755 | 7 | 0.709 | 0.651–0.767 | 7 |
Variables included in the model.
Figure 5.
Prediction performance of DVT risk based on different supervised algorithm. (A) DCA for 3 prediction models in the training set; (B) DCA for 3 prediction models in the testing set.
3.6. Evaluation of prediction accuracy of ML-based DVT prediction model
To further evaluate the optimal predictive performance of the RFM model, we used clinical impact curves to evaluate the classification accuracy of the prediction model. As shown in Figure S1, Supplemental Digital Content, http://links.lww.com/MD/L158 RFM can effectively distinguish between DVT patients and non-DVT patients, which also suggested that RFM was a reliable tool for preoperative evaluation of DVT in patients and may become an effective screening and prediction tool for determining DVT, which can be applied to DVT risk assessment.
4. Discussion
So far, laparoscopy has been widely used in gynecological benign and malignant diseases due to its advantages of minimal injury, less postoperative wound pain and fast recovery. However, due to the artificial pneumoperitoneum during the operation, the abdominal pressure can be greater than the lower limb venous blood flow return pressure, which will change the venous hemodynamics, expand the lower limb venous vessels, slow down the blood flow, and increase the intravascular pressure, affecting the blood flow return.[3,16] Clinically, the DVT refers to the abnormal coagulation of blood in the deep vein cavity, blocking the venous lumen, leading to venous reflux obstacles.[5] The main veins of the whole body can be affected, especially in the lower limbs.[4] Severe DVT can promote the occurrence of pulmonary embolism and then endanger life. Previous studies have reported that the incidence of DVT after gynecological laparoscopic surgery ranges from 7.6% to 11.55%.[8,17] Previous studies have shown that the postoperative blood viscosity of patients undergoing gynecological laparoscopy has significantly increased, and the incidence of DVT is about 8%.[18,19] Therefore, it is necessary to take preventive and intervention measures as soon as possible to reduce the incidence of DVT.
The results of this study found that age, intraoperative pneumoperitoneum pressure > 15mmHg, history of diabetes, D-dimer > 0.5 mg/L, operation time ≥ 1 hour, and complications are all independent influencing factors for the occurrence of DVT after gynecological laparoscopy. Therefore, it is necessary to take preventive measures against these factors to reduce the occurrence of DVT. Previous studies have shown that the probability of DVT occurring for the first time increases exponential growth with age, among which age > 60 is an independent risk factor for DVT occurring after gynecological surgery.[17,18] We speculate that the main reason may be that as age increases, the amount of activity decreases correspondingly, and the venous blood flow velocity slows down. Especially as age increases, the elasticity of the vascular wall in elderly patients decreases, and endothelial cells are easily damaged.[20] When endothelial cells release more coagulant promoting substances, the risk of DVT also increases accordingly. In addition, intraoperative pneumoperitoneum pressure > 15 mm Hg is also an independent risk factor that cannot be ignored. The pneumoperitoneum pressure of gynecological laparoscopy is usually 12 to 15 mm Hg, while the normal pressure of human inferior vena cava is 2 to 5 mm Hg, which means that the greater the abdominal cavity pressure is, the greater the pressure of inferior vena cava will be, leading to slower blood flow.[21,22] At the same time, the blood is hypercoagulable, which is easy to cause endothelial damage and induce platelet activation, adhesion and release, thus increasing the risk of DVT occurrence.
In this study, we also found that the history of diabetes is an independent risk factor analysis of postoperative DVT. There is no doubt that the elderly women often have a history of diabetes, and diabetes is very easy to cause vascular endothelial damage, the activation and release of inflammatory factors, the generation of a large number of oxygen free radicals, leading to the activation of platelets and the coagulation system. A meta-analysis showed that the risk of DVT in patients with diabetes could increase by 1.5 to 2 times, which was consistent with the conclusion of this study, and further confirmed that preoperative hyperglycemia was a high-risk factor for thrombosis after gynecological surgery.[23] In addition, previous studies have shown that the rise of D-dimer is of great significance in identifying lower limb arterial thrombosis.[24,25] The critical value of D-dimer ≥ 0.5 mg/L has a good predictive value for the occurrence of DVT in elderly inpatients, with a sensitivity of 100% and a specificity of 41.8%.[26] This suggests that the rise of D-dimer level can also be considered as an effective indicator of coagulation system abnormalities, which has been confirmed in this study, because D-dimer occupies a large weight in the three types of prediction models. Of course, we also explored the potential correlation between surgical time and the risk of DVT occurrence. Consistent with previous research results, we found that surgical time ≥ 1 hour is an independent risk of postoperative DVT occurrence. Therefore, it is speculated that the long surgical time may be related to the patient’s poor self condition and excessive intraoperative blood loss. At the same time, the anesthesia time is longer, and the muscles lose their retraction function due to narcotic, resulting in venous blood retention in the lower limbs.
In addition to the recognized objective risk factors mentioned above, we also found that preoperative complications in patients are independent risk factors that cannot be ignored. In recent years, with the improvement of laparoscopy instruments and equipment, as well as the improvement of doctors’ surgical techniques, the overall incidence of postoperative complications continues to decrease.[27] However, at this stage, the indications of laparoscopy continue to expand, the complexity and difficulty of surgery continue to increase, and the prevention and management of postoperative complications are still not optimistic.[17,28] This study shows that the incidence of DVT in patients with postoperative comorbidities is higher than that in patients without postoperative comorbidities, which may be related to the stress response triggered by postoperative comorbidities and changes in hemodynamics caused by excessive postoperative blood and fluid loss. In this study, due to sample size limitations, the correlation between specific comorbidities and DVT was not analyzed, which will be the focus of the next study on the occurrence of DVT after laparoscopic surgery.
In this study, we built the DVT prediction model, an ensemble model derived from three supervised learning algorithms (RFM, ANNM, and GLRM), that enabled accurate prediction of potential risk of DVT for patients who received laparoscopy in advance using clinical information in EHRs on admission, which may provide clinically important information for risk stratification and treatment planning. Importantly, the RFM displayed an AUC ranging from 0.851 to 0.862 in the training and validation cohorts. The prognostic implications of DVT prediction model might facilitate more responsive health systems that are conducive to patients with high-risk DVT via early identification, and ensuing instant intervention as well as intensive care and monitoring, thus, hopefully assisting to optimize nursing intervention strategies after laparoscopy. Consistent with previous research reports, strengths of RFM-based prediction model also include its stability and practicability upon patients with or without several missing features. Interestingly, from the candidate variables in the RFM iteration, it can be seen that the 13 features for DVT prediction were readily accessible and frequently monitored in routine clinical practice.[29,30] For example, Age and sex were basic information. Age, BMI, and diabetes were easily observed symptoms, while operation time and IPP were physical signs available at hand. Presence of complication and D-Dimer could be ascertained by referring to previous EHRs and patients or their family doctors.
This study still has the following limitations. First, as this study is limited to a single center, future studies should expand the sample size and conduct prospective cohort studies across multiple centers. Second, due to the influence of retrospective cohort studies, the patients we included may have selection bias, so it is still necessary to control for potential data bias that may lead to result bias; Third, the clinical parameters included in this study are limited, and in the future, it is still necessary to incorporate patient imaging data and molecular biology parameters into the prediction model in order to optimize the predictive performance of DVT.
5. Conclusion
In summary, our combination of advanced ML algorithms and clinical features can timely and accurately classify high-risk DVT patients, especially the DVT prediction model based on RFM, which can help clinical healthcare professionals identify DVT in a timely manner and provide appropriate clinical care to promote postoperative recovery and improve patient prognosis.
Author contributions
Conceptualization: Dongxue Wang.
Investigation: Min Hou.
Methodology: Min Hou.
Resources: Dongxue Wang.
Software: Xiao Chen, Dongxue Wang.
Supervision: Xiao Chen.
Visualization: Min Hou.
Writing – original draft: Xiao Chen, Min Hou.
Writing – review & editing: Xiao Chen, Min Hou, Dongxue Wang.
Supplementary Material
Abbreviations:
- ANNM
- artificial neural network model
- AUC
- area under the curve
- BMI
- body mass index
- DCA
- decision curve analysis
- DVT
- deep vein thrombosis
- EHRs
- electronic health records
- GLRM
- generalized linear regression model
- LASSO
- least absolute shrinkage and selection operator
- ML
- machine learning
- RFM
- random forest model
- ROC
- receiver operating characteristic curve
The authors have no funding and conflicts of interest to disclose.
The datasets generated during and/or analyzed during the current study are not publicly available, but are available from the corresponding author on reasonable request.
Supplemental Digital Content is available for this article.
How to cite this article: Chen X, Hou M, Wang D. Machine learning-based model for prediction of deep vein thrombosis after gynecological laparoscopy: A retrospective cohort study. Medicine 2024;103:1(e36717).
Contributor Information
Xiao Chen, Email: chenxiaotj_2023@163.com.
Min Hou, Email: 859283963@qq.com.
References
- [1].Shore EM, Lefebvre GG, Grantcharov TP. Gynecology resident laparoscopy training: present and future. Am J Obstet Gynecol. 2015;212:298–301, 298.e1. [DOI] [PubMed] [Google Scholar]
- [2].Lee SH, Oh SR, Cho YJ, et al. Comparison of vaginal hysterectomy and laparoscopic hysterectomy: a systematic review and meta-analysis. BMC Women’s Health. 2019;19:83. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [3].Tian Q, Li M. Risk factors of deep vein thrombosis of lower extremity in patients undergone gynecological laparoscopic surgery: what should we care. BMC Women’s Health. 2021;21:130. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [4].Gutzeit O, Lauterbach R, Loberman Z, et al. Laparoscopic sacrocolpopexy complication: ilio-femoral deep vein thrombosis. Eur J Obstet Gynecol Reprod Biol. 2020;247:270–1. [DOI] [PubMed] [Google Scholar]
- [5].Yu R, Nansubuga F, Yang J, et al. Efficiency and safety evaluation of prophylaxes for venous thrombosis after gynecological surgery. Medicine (Baltimore). 2020;99:e20928. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [6].Carr CE, Chambers L, Jernigan AM, et al. Short- and long-term outcomes for single-port risk-reducing salpingo-oophorectomy with and without hysterectomy for women at risk for gynecologic cancer. Int J Gynecol Cancer. 2021;31:215–21. [DOI] [PubMed] [Google Scholar]
- [7].Ackroyd SA, Hernandez E, Roberts ME, et al. Postoperative complications of epidural analgesia at hysterectomy for gynecologic malignancies: an analysis of the National Surgical Quality Improvement Program. Int J Gynecol Cancer. 2020;30:1203–9. [DOI] [PubMed] [Google Scholar]
- [8].Santoso JT, Evans L, Lambrecht L, et al. Deep venous thrombosis in gynecological oncology: incidence and clinical symptoms study. Eur J Obstet Gynecol Reprod Biol. 2009;144:173–6. [DOI] [PubMed] [Google Scholar]
- [9].Handelman GS, Kok HK, Chandra RV, et al. eDoctor: machine learning and the future of medicine. J Intern Med. 2018;284:603–19. [DOI] [PubMed] [Google Scholar]
- [10].MacEachern SJ, Forkert ND. Machine learning for precision medicine. Genome. 2021;64:416–25. [DOI] [PubMed] [Google Scholar]
- [11].Deo RC. Machine learning in medicine. Circulation. 2015;132:1920–30. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [12].Wang J, Liu S. Visual information computing and processing model based on artificial neural network. Comput Intell Neurosci. 2022;2022:4713311. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [13].Rigatti SJ. Random forest. J Insur Med. 2017;47:31–9. [DOI] [PubMed] [Google Scholar]
- [14].Meurer WJ, Tolles J. Logistic regression diagnostics: understanding how well a model predicts outcomes. JAMA. 2017;317:1068–9. [DOI] [PubMed] [Google Scholar]
- [15].Vickers AJ, Holland F. Decision curve analysis to evaluate the clinical benefit of prediction models. Spine J. 2021;21:1643–8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [16].Vedantham S. Using it wisely: catheter-directed thrombolysis for deep vein thrombosis. Lancet Haematol. 2020;7:e6–7. [DOI] [PubMed] [Google Scholar]
- [17].Schmeler KM, Wilson GL, Cain K, et al. Venous thromboembolism (VTE) rates following the implementation of extended duration prophylaxis for patients undergoing surgery for gynecologic malignancies. Gynecol Oncol. 2013;128:204–8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [18].Yang T, Tian S, Wang Y, et al. Evaluation of risk factors for venous thromboembolism in patients who underwent gynecological surgery and validation of a fast-rating assessment table. Med Sci Monit. 2019;25:8814–9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [19].Martino MA, Williamson E, Rajaram L, et al. Defining practice patterns in gynecologic oncology to prevent pulmonary embolism and deep venous thrombosis. Gynecol Oncol. 2007;106:439–45. [DOI] [PubMed] [Google Scholar]
- [20].Halawani HM, Ripley-Hager CF, Naglak MC, et al. Venous thromboembolism after laparoscopic or robotic biliopancreatic diversion with duodenal switch Ninety-days outcome of a 10 years’ experience. Surg Obes Relat Dis. 2017;13:1984–9. [DOI] [PubMed] [Google Scholar]
- [21].Liu X, Wang X, Meng X, et al. Effects of patient position on lower extremity venous pressure during different types of hysterectomy. J Obstet Gynaecol Res. 2015;41:114–9. [DOI] [PubMed] [Google Scholar]
- [22].Liu C, Han Z, Zhang N, et al. Laparoscopic sleeve gastrectomy affects coagulation system of obese patients. Obes Surg. 2020;30:3989–96. [DOI] [PubMed] [Google Scholar]
- [23].Ageno W, Becattini C, Brighton T, et al. Cardiovascular risk factors and venous thromboembolism: a meta-analysis. Circulation. 2008;117:93–102. [DOI] [PubMed] [Google Scholar]
- [24].Kawaguchi R, Furukawa N, Kobayashi H. Cut-off value of D-dimer for prediction of deep venous thrombosis before treatment in ovarian cancer. J Gynecol Oncol. 2012;23:98–102. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [25].Habu Y, Mitsuhashi A, Hanawa S, et al. High prevalence of pulmonary embolism prior to cancer therapies in patients with ovarian and endometrial cancers detected by contrast-enhanced CT using D-dimer as an index. J Surg Oncol. 2021;124:106–14. [DOI] [PubMed] [Google Scholar]
- [26].Burgazli KM, Bilgin M, Kavukçu E, et al. Diagnosis and treatment of deep-vein thrombosis and approach to venous thromboembolism in obstetrics and gynecology. J Turk Ger Gynecol Assoc. 2011;12:168–75. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [27].Holzheimer RG. Laparoscopic procedures as a risk factor of deep venous thrombosis, superficial ascending thrombophlebitis and pulmonary embolism--case report and review of the literature. Eur J Med Res. 2004;9:417–22. [PubMed] [Google Scholar]
- [28].Ramirez PT, Nick AM, Frumovitz M, et al. Venous thromboembolic events in minimally invasive gynecologic surgery. J Minim Invasive Gynecol. 2013;20:766–9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [29].Paul A, Mukherjee DP, Das P, et al. Improved random forest for classification. IEEE Trans Image Process. 2018;27:4012–24. [DOI] [PubMed] [Google Scholar]
- [30].Menze BH, Kelm BM, Masuch R, et al. A comparison of random forest and its Gini importance with standard chemometric methods for the feature selection and classification of spectral data. BMC Bioinf. 2009;10:213. [DOI] [PMC free article] [PubMed] [Google Scholar]