Prediction of the Risk of Preoperative Deep Vein Thrombosis in Non-fractured Patients Awaiting Total Joint Arthroplasty Using a Support Vector Machine

Haosheng Wang, Tingting Fan, Yanhua Chen, Wenle Li, Fujiang Zhao, Qiuming Liu, Mingyu YangResearch 2021 · DOI

TL;DR

Researchers have developed a Support Vector Machine (SVM) machine learning model to predict the risk of "silent" blood clots (Deep Vein Thrombosis) in patients before they undergo hip or knee replacements. By analyzing five key clinical markers—including fibrinogen levels and blood pressure history—the model achieved an impressive 98% accuracy, potentially offering a vital screening tool to prevent fatal surgical complications.

What Problem Does This Paper Address?

Total Joint Arthroplasty (TJA), such as hip and knee replacements, is a highly effective treatment for advanced joint disease. However, Venous Thromboembolism (VTE)—which includes Deep Vein Thrombosis (DVT) and its potentially fatal counterpart, Pulmonary Embolism (PE)—remains a significant concern for surgeons.

While much research focuses on preventing clots after surgery, this paper tackles the "silent" threat of preoperative DVT. In patients who haven't suffered a recent fracture, DVT is often asymptomatic and can go undetected without expensive, universal ultrasound screening. Missing a pre-existing clot before surgery significantly increases the risk of the clot dislodging during the procedure. Currently, there is a lack of reliable, data-driven tools to identify which non-fractured patients are at high risk before they reach the operating table.

Methods

The authors conducted a retrospective study of 369 elective TJA patients. To identify the most relevant predictors, they first used traditional statistical methods (univariate and multivariate logistic regression) to screen clinical, imaging, and laboratory data.

The core of the study is the application of a Support Vector Machine (SVM). SVM is a supervised learning algorithm that works by mapping data into a high-dimensional space to find the optimal "hyperplane" (a boundary) that separates two groups—in this case, patients with DVT and those without.

Key technical details include: * Kernel Function: They used a Radial Basis Function (RBF) kernel to handle non-linear relationships in the data. * Optimization: Parameters (C and γ) were tuned using a grid search. * Validation: The model was validated using 10-fold cross-validation to ensure the results weren't just a fluke of the specific data split.

Datasets

The study utilized a clinical dataset from a single institution (Taizhou Central Hospital) covering patients treated between March 2015 and August 2020.

  • Total Patients: 369 (21 with confirmed DVT, 348 without).
  • Ground Truth: All patients underwent preoperative color Doppler ultrasound of the lower extremities to confirm the presence or absence of thrombi.
  • Features: The dataset included age, BMI, medical history (hypertension, diabetes, etc.), and a suite of laboratory markers (D-dimer, fibrinogen, hemoglobin, etc.).

Key Insights and Ideas

  • The "Big Five" Predictors: The study identified five independent risk factors that most strongly predict preoperative DVT: preoperative fibrinogen, age, D-dimer levels, history of hypertension, and preoperative hematocrit (Hct).
  • Fibrinogen as a Leading Indicator: Fibrinogen emerged as the most significant predictor. The authors note that it isn't just a clotting factor but also an inflammatory marker, suggesting that the body's inflammatory state is a major driver of preoperative clot formation.
  • The Hypertension Connection: Patients with hypertension were significantly more likely to have a pre-existing clot, likely due to chronic vascular inflammation and endothelial dysfunction associated with high blood pressure.
  • Machine Learning vs. Statistics: While traditional regression identified the factors, the SVM model provided a more robust way to combine these variables into a predictive tool, handling the complexity of medical data better than simple linear models.

Key Results

The SVM model demonstrated high performance across several metrics:

  • Accuracy: 0.98 (98%)
  • Precision: 0.92
  • Recall: 0.93
  • Area Under the Curve (AUC): The model achieved a maximum AUC of 0.94 and an average of 0.77 in 10-fold cross-validation.

In practical terms, these numbers suggest the model is highly effective at identifying the small subset of patients (5.7% in this study) who actually have a clot, without generating an overwhelming number of false alarms.

Limitations and Future Work

The authors acknowledge several hurdles before this can be used in daily clinical practice:

  • Sample Size: The number of DVT-positive cases (21) was relatively small, which can lead to "overfitting" where the model learns the specific quirks of those 21 people rather than general biological truths.
  • Single-Center Bias: Because the data came from one hospital, the model might not perform as well on different populations (e.g., patients in different countries or with different ethnic backgrounds).
  • Retrospective Nature: The study looked backward at existing data. A "prospective" study—where the model makes predictions in real-time before the ultrasound is performed—is needed to prove its clinical utility.
  • External Validation: The model needs to be tested on entirely new datasets from other institutions to ensure its generalizability.

Summary updated

Abstract

<title>Abstract</title> <bold>Background: </bold>We developed a potential useful alternative prediction model based on the support vector machine (SAM) algorithm to predict the risk of preoperative deep vein thrombosis (DVT) in non-fractured patients awaiting total joint arthroplasty (TJA). <bold>Methods: </bold>From March 2015 to August 2020, a retrospective review of the preoperative ultrasound examination findings of lower extremity venous vessels was performed on non-fractured patients of 369 elective TJA. Based on the ultrasound examination findings of preoperative lower extremely venous vessels, these patients were divided into two groups: the DVT group and the Non-DVT group. We collected the clinical, imaging, and laboratory findings from an electronic medical record system. These variables were imported into univariate, multivariate and logistic regression analysis to identify the risk factor for preoperative DVT. According to published literature and clinical experience, a series of variables were selected to construct a prediction model based on the SVM machine learning algorithm. <bold>Results: </bold>Among the 369 patients, preoperative DVT was observed in 21 patients (5.7%). The Multivariate regression analysis showed the following 5 independent factors associated with preoperative DVT: preoperative fibrinogen odds ratio [OR] = 7.306), age (OR = 1.133), history of hypertension (OR = 3.848), preoperative hematocrit (OR = 0.315), and D-dimer (OR = 2.032). The SVM model achieved a maximum and average area under the receiver operating characteristic curve (AUC) of 0.94 and 0.77 in the 10-fold cross-validation. Meanwhile, the accuracy, precision, and recall of the model were 0.98, 0.92, and 0.93, respectively. Additionally, the confusion matrix showed the classification results of the discriminant analysis.<bold>Conclusions: </bold>SVM machine modeling is a promising method for the prediction of the risk of DVT in non-fractured patients awaiting TJA. However, future external validation is needed.
BibTeX
@inproceedings{wang2021prediction,
  author = {Haosheng Wang and Tingting Fan and Yanhua Chen and Wenle Li and Fujiang Zhao and Qiuming Liu and Mingyu Yang},
  title = {{Prediction of the Risk of Preoperative Deep Vein Thrombosis in Non-fractured Patients Awaiting Total Joint Arthroplasty Using a Support Vector Machine}},
  booktitle = {Research},
  year = {2021},
  doi = {10.21203/rs.3.rs-844532/v1},
}
Figure