Using Machine-Learning (ML) models to predict risk of Venous Thromboembolism (VTE) following spine surgery

Abstract

Study Design:

Retrospective Cohort Study

Objectives:

Venous Thromboembolism (VTE) is a potentially high-risk complication for patients undergoing spine surgery. Although guidelines for assessing VTE risk in this population have been established, development of new techniques that target different aspects of the medical history may prove to be of further utility. The goal of this study was to develop a predictive machine learning (ML) model to identify non-traditional risk factors for predicting VTE in spine surgery patients.

Summary of Background Data:

A cohort of 63 patients was identified who had undergone spine surgery at a single center from 2015–2021. 31 patients had a confirmed VTE while 32 had no VTE. 113 attributes were defined and collected via chart review. Attribute categories included demographics, medications, labs, past medical history, operative history, and VTE diagnosis.

Methods:

The Waikato Environment for Knowledge Analysis (WEKA) software was used in creating and evaluating the ML models. Six classifier models were tested with 10-fold cross validation and statistically evaluated using t-tests.

Results:

Comparing the predictive ML models to the control model (ZeroR), all predictive models were significantly better than the control model at predicting VTE risk, based on the 113 attributes (p < 0.001). The RandomForest model had the highest accuracy of 88.89% with a positive predictive value (PPV) of 93.75%. The SimpleLogistic algorithm had an accuracy of 84.13% and defined risk attributes to include calcium and phosphate lab values, history of cardiac comorbidity, history of previous VTE, anesthesia time, SSRI use, antibiotic use, and antihistamine use. The J48 model had an accuracy of 80.95% and it defined hemoglobin lab values, anesthesia time, beta blocker use, dopamine agonist (DA) use, history of cancer, and Medicare use as potential VTE risk factors.

Conclusions:

Further development of these tools may provide high diagnostic value and may guide chemoprophylaxis treatment in this setting of high-risk patients.

Introduction

Venous Thromboembolisms (VTE) are post-surgical complications that carry a significant burden of morbidity and mortality¹. Specifically, in spine surgery, Deep Vein Thromboses (DVT) and Pulmonary Embolisms (PE) incidences are increased due to various factors such as increased surgery durations and decreased postoperative mobility^1,2,4. While various risk factors have been explored and their impact on VTE risk studied, there is still difficulty in applying these risk factors to predict VTE occurrence post spine surgery.

Various studies have retrospectively identified a set of factors that contribute to VTE risk, and some have even quantified the risk using their own scoring guidelines or models. These tools are helpful in gauging which patients with risk factors will benefit most from prophylactic anticoagulation. Chemoprophylaxis cannot be administered to all patients due to its rare but serious complications, which can include epidural hematomas, infection, and retropharyngeal hematomas. These complications can lead to paralysis, or even death. Hence, it is imperative to establish a predictive model that can effectively determine VTE risk in patients who have undergone spine surgery to accurately administer both mechanical and chemoprophylactic in an appropriate manner for this population^5–8.

Many studies have identified risk factors through their own risk assessment guidelines based off regression analyses. A retrospective study from Piper et al found 13 VTE preoperative and operative risk factors through regression and created a scoring system where patients scoring 7/13 or greater had a 100-fold increased risk of developing VTE post spine surgery compared to patients who scored 0³. There are also standardized guidelines from the North American Spine Society (NASS) and American College of Chest Physicians (ACCP) that recommend interventions such as mechanical prophylaxis for all patients, but chemoprophylaxis on a case-by-case basis, leaving institutions to determine specific risk factors within their own patient populations.

Standardized scores and guidelines may not include all potential risk factors that could be used to identify patients with a high risk of VTE complications postoperatively. Understanding all aspects of their medical chart may prove to be of further utility. Thus, we plan to also apply novel use of predictive machine learning (ML) models to evaluate pre-operative medical charts of patients who underwent spine surgery and predict which patients are at highest risk for developing VTE postoperatively. With the absence of literature evaluating the effectiveness of such artificial intelligence models, this pilot project hopes to illustrate how the use of such tools can provide diagnostic value in this setting.

Methods

Data Collection

A patient cohort was established with 31 patients who had undergone spine surgery at from 2016–2021 with confirmed incidence of VTE within a 90-day postoperative period, and 32 patients who had undergone spine surgery at from 2015–2019 with no incidence of VTE (n = 63). The 32 control patients were randomly selected using random number generator. 113 attributes were collected through manual chart review and data pulled via the Tripartite Request Assessment Committee (TRAC) at New York Presbyterian. There were four attribute categories: demographics, preoperative history, operative history and postoperative history. The following attributes were collected for demographics: sex, height, weight, BMI, insurance, ethnicity, and age. For preoperative history, labs (CBC, BMP, Coags, LFTs), past medical history (cardiac comorbidity, hypertension, hyperlipidemia, obstructive sleep apnea, asthma, COPD, anxiety, depression, osteoporosis, diabetes, hypothyroidism, coagulopathy, history of venous thromboembolism, history of cancer, gastroesophageal reflux disease, celiac disease, irritable bowel syndrome, autoimmune condition, benign prostatic hypertrophy, arthritis, vertigo, neurological comorbidity, glaucoma, and history of STIs), social history (smoking, alcohol, and other drug use), and medications (NSAID, acetaminophen, lipid lowering therapy, Xa inhibitor, H2 blocker, ACE inhibitor, PPI, antiemetic/prokinetic, antiplatelet, anticonvulsant, beta blocker, opioid, opioid antagonist, alpha blocker, SSRI, SNRI, 5-HT modulator, benzodiazepine, hormone therapy, corticosteroid, antileukotriene, vitamin supplement, immunosuppressant, antibiotic, antifungal, tricyclic antidepressant, antihistamine, PDE inhibitor, diuretic, CCB, dopamine agonist, anticholinergic, bronchodilator, alpha agonist, antituberculosis therapy, prostaglandin, antipsychotic, diabetes therapy, laxatives, sedative/hypnotics, muscle relaxant, DMARDs, antimalarial therapy) were collected. Intraoperative data collected included anesthesia time, EBL, and if there were any intraoperative complications. Postoperatively, LOS, non VTE postoperative complication, and incidence of VTE were collected.

All categorical variables (including insurance, ethnicity, past medical history, medications, perioperative complications, and VTE incidence) were converted to binary, where 0 indicated a negative result and 1 was a positive result (except in the case of sex, where 0 = male and 1 = female).

Classifying and Training Machine Learning Models

The Waikato Environment for Knowledge Analysis (WEKA) software was used in creating and evaluating the ML models. All binary variables were converted to a readable format using the unsupervised NumerictoBinary function. 5% of the data was missing across all patient attributes, and any missing data points were imputed to provide full data sets for the model. Six classifier models were tested with 10-fold cross validation: ZeroR, OneR, Logistic, SimpleLogistic, J48 (C4.5) and RandomForest with VTE as the predicting output on 53 patients with 10 patients used as test cases. Models were statistically analyzed against each other using t-tests in WEKA.

Results

Machine Learning

Out of the VTE patient cohort, 23 patients had a PE (74%), and 8 patients had a DVT (26%) (Table 1). VTE diagnosis was established on average 5.13 +/− 2.14 days post spine surgery using Doppler or Computed tomography angiography (CTA).

Table 1.

Patient Cohort Demographics

	Control	95% CI	VTE	95% CI	p-value
N	32	31
Age	55.38	[53.6, 57.1]	62.29	[60.7, 63.9]	0.0525
% Female	17 (53%)		19 (61%)		0.6131
BMI	29.42	[28.2, 30.6]	28.78	[27.9, 29.7]	0.7870
PE Incidence			23 (74%)
VTE Incidence			8 (26%)
VTE Diagnosis			5.13	[4.38, 5.88]

The ZeroR model served as the control and had a 50.79% accuracy. The other rule model, OneR had an accuracy of 85.71%. Out of the regression models, the SimpleLogistic model had an 84.12% accuracy rate, and the Logistic model had an 80.95%. The last two models created and validated were decision tree models. The J48 model had an accuracy of 80.95%, and the RandomForest model had the best accuracy at 88.89%. Compared to the ZeroR model, all models predicted VTE risk significantly better (Figure 1). The SimpleLogistic Model used the following weighted attributes to calculate VTE risk: low calcium and high phosphate lab values, past medical history of cardiac comorbidity or previous VTE, total anesthesia time during spine surgery, and use of SSRI, antibiotic, and/or antihistamine medications. The J48 Tree Model used the following weighted attributes to calculate VTE risk: hemoglobin lab values, past medical history of cancer, total anesthesia time during spine surgery, use of beta-blockers or dopamine agonists, and Medicare use. In this model, total anesthesia time had the highest weight, and use of dopamine agonists and use of Medicare had the lowest weight.

Figure 1.

Accuracy of VTE risk predictive models. All models did significantly better than the ZeroR (50.79%) when validated with the test set after 10-fold cross validation training (* indicating p<0.05). The most notable models were the RandomForest with the highest accuracy (88.89%) and the SimpleLogistic model (84.13%).

Discussion

There is limited literature exploring the use of artificial intelligence and ML models in predicting VTE incidence and identifying key risk factors via an algorithm. Thus, the purpose of this study was to develop a proof-of-concept ML model that could parse through various pre-operative and intraoperative risk factors collected from the EMR, and identify which risk factors were most likely to contribute to VTE incidence and utilize these factors to predict future risk in this patient population.

A recent study looked at various ML algorithms and their utility in predicting VTE risk specifically following posterior lumbar fusion. They found five clinical variables to be most significant: age >65, obesity grade II or above, CAD, functional status, and prolonged operative time⁹. Another study evaluated VTE risk factors in cervical spine surgery, citing male sex, fluid/electrolyte disorders, and pulmonary circulation disorders to be most significant in this patient sample⁴.

There are other studies that looked at logistic regressions to identify VTE related risk factors. A meta-analysis of 21 studies found that a higher rate of VTE complications post spine surgery was closely associated with increased age, longer surgery duration, greater blood loss, and patients with a history of hypertension, preoperative walking disability, or diabetes¹⁰.

We also found similar VTE risk attributes. For example, our SimpleLogistic model found a history of cardiac comorbidities and increased anesthesia time, which is a proxy for operative time as associated with increased VTE risk. Similarly, the J48 model found increased anesthesia time as well as beta blocker use as VTE risk factors. In this case, the beta blocker use can be attributed to underlying comorbidities such as CAD, CHF, and hypertension, all which are typically treated with beta blockers. A history of cancer or active malignancy has also been shown to be an established risk factor for VTE incidence¹¹.

The studies exploring hemoglobin levels on VTE incidence have been variable. One study from Braekkan et al in 2009 noted how greater hematocrit and associated hematological variables such as hemoglobin led to at 2.4-fold increase in unprovoked VTE, more specifically in men¹². However, a more recent study from Chi et al in 2018 showed that hospitalized patients with anemia were at a twofold greater risk of symptomatic DVT or non-fatal PE despite pharmacologic VTE prophylaxis use during their hospital stay, regardless of sex¹³. Of note, these studies did not evaluate VTE risk in patients undergoing surgery, and thus the findings may not be fully translatable to our study and patient population.

There is no evidence in the literature exploring the association between VTE incidence and other factors such as serum calcium and phosphate levels, dopamine agonist use, chronic antibiotic use, SSRI use, and antihistamine use. Further studies are warranted to investigate the effect of these factors on VTE incidence.

The limited data set and sample size significantly limited the power of the study. With a cohort of 63 patients, it is challenging to generalize both the findings and the AI model to larger populations. This study was meant to provide a pilot proof of concept that ML, when used in conjunction with various types of attributes, can provide unknown associations between clinical procedures and outcomes. A larger retrospective review or prospective study applying the proposed models would be necessary to evaluate the model’s validity and efficacy in determining VTE risk and chemoprophylaxis eligibility to reduce incidence.

Conclusion

The utility of pre-operative ML modeling for predicting VTE risk in patients undergoing spine surgery has been demonstrated in a proof of concept pilot model to be robust and accurate. Further investigation and development of these tools may provide high diagnostic value and may guide risk stratification and chemoprophylaxis regimens in this setting of high-risk patients.