Development and Internal Validation of Supervised Machine Learning Algorithms for Predicting the Risk of Surgical Site Infection Following Minimally Invasive Transforaminal Lumbar Interbody Fusion Under Quadrant Channel

Background: Machine Learning (ML) is rapidly growing in capability and is increasingly applied to model outcomes and complications in medicine. Surgical site infections (SSI) are a common postoperative complication in spinal surgery. This study aimed to develop and validate supervised ML algorithms for predicting the risk of SSI following minimally invasive transforaminal lumbar interbody fusion (MIS-TLIF) under the Quadrant channel. Methods: This single-central retrospective study included a total of 705 cases between May 2012 and October 2019. Data of patients who underwent MIS-TLIF under the Quadrant channel was extracted by the electronic medical record system. The patient’s clinical characteristics, surgery-related parameters, and routine laboratory tests were collected. Univariate and multivariate logistic regression analyses were used to screen and identify independent risk factors for SSI. Then, the independent risk factors were imported into six ML algorithms, including k-Nearest Neighbor (KNN), Decision Tree (DT), Support Vector Machine (SVM), Random Forest (RF), Multi-Layer Perceptron (MLP), and Naïve Bayes (NB), to develop a prediction model for predicting the risk of SSI following MIS-TLIF under Quadrant channel. During the training process, 10-fold cross-validation was used for validation. Indices like the area under the receiver operating characteristic (AUC), sensitivity, speci�city, and accuracy (ACC) were reported to test the performance of ML models. Results: Among the 705 patients, SSI occurred in 33 patients (4.68%). The univariate and multivariate logistic regression analyses showed that preoperative glycated hemoglobin A1c (HbA1c), estimated blood loss (EBL), preoperative albumin, body mass index (BMI), and age were all independent predictors of SSI. In predicting SSI, six ML models posted an average AUC of 0.60-0.80 and an ACC of 0.80-0.95, with the NB model standing out, registering an average AUC and an ACC of 0.78 and 0.90. Then, the feature importance of the NB model was reported. Abbreviations: BMI, body mass index; COPD, chronic obstructive pulmoriary disease; CHF, chronic heart failure; ASA, American Society of Anesthesiologists; EBL, estimated blood loss; HbA1c, glycated hemoglobin A1c; Alb, albumin; RBC, red WBC, prothrombin APTT: thromboplastin time; Fib, �brinogen.


Background
Minimally invasive transforaminal lumbar interbody fusion (MIS-TLIF) is a classic minimally invasive operation for the treatment of lumbar degenerative diseases such as lumbar disc herniation, lumbar spinal stenosis, and lumbar spondylolisthesis. The operation uses channels to complete decompression and bone graft fusion and xation. Compared with the traditional open TLIF, it better retains the paravertebral muscle structure, less intraoperative bleeding and faster postoperative recovery [1][2][3]. It has become the mainstream operation scheme for minimally invasive fusion surgery at present.
Although surgical site infections (SSI) following MIS-TLIF are lower than those following open TLIF, previous literature has reported that the incidence of SSI after MIS-TLIF is not uncommon as this procedure is widely used in the clinic [4,5]. Surprisingly, over 20% of the patients who were re-admitted within 30 days after operation were caused by SSI [6,7]. Meanwhile, the rise in the number of spinal surgeries results in an expanding number of SSI, which poses a critical challenge to patients and clinicians. At present, there are studies on the analysis of preoperative and intraoperative risk factors for SSI after total spinal open surgery, but there is little literature investigating the risk factors for SSI following MIS-TLIF under the Quadrant channel. Previously, most studies only described these risk factors as relative risks (RR) or odds ratios (OR) [7,8], which are not su cient to evaluate the risk of SSI following MIS-TLIF under the Quadrant channel. Therefore, a comprehensive prediction model combining risk factors is needed to assess SSI risk for patients. Identi cation of high-risk surgical populations might help target interventions to patients at high risk, reduce the risk of hospitalization, and improve the clinical outcome.
Machine learning (ML) is a form of arti cial intelligence (AI) in which algorithms automatically learn and improve by identifying patterns and complex relationships, with the ultimate goal of making decisions using minimal human intervention [9][10][11][12][13]. Through the good performance exhibited by ML algorithms in medical big data, we have the potential to obtain more superior prediction tools than traditional statistical modeling modalities under certain conditions for better prediction of the risk of SSI. It is frustrating that there are currently very few studies training ML algorithms to predict SSI risk. As such, this study aimed to develop and validate ML-based models using preoperative and intraoperative variables to predict the risk of SSI following MIS-TLIF under the Quadrant channel.

Patients
The study was approved by the institutional review board (IRB) of Taizhou Central Hospital and was performed following national and international guidelines. The ethics committee waived the requirement for informed consent due to the retrospective nature of the study. A total of 705 patients, referred from May 2012 to October 2019 to the Department of Orthopedics, Taizhou Central Hospital (Taizhou, China) for MIS-TLIF owing to lumbar spine diseases, were enrolled in this research. The selection criteria were as follows: (1) age ≥ 18 years; (2) diagnosis of lumbar degenerative disease, including lumbar disc herniation, lumbar spinal stenosis, lumbar spondylolisthesis, and lumbar instability; (3) undergoing single-or multi-segment MIS-TLIF under Quadrant channel. The exclusion criteria were as follows: (1) Patients were preoperatively complicated by active infection in the spine or other parts of the body, cauda equina syndrome, spinal deformity, and tumors. (2) Emergency surgery.
All operations were done by 4 experienced spinal surgeons at our institution. Patients were followed for a minimum of 1 year after the index operation to monitor all complications and incidences of revision surgery. All complications were recorded by the available electronic medical system. The primary outcome of interest was the occurrence of SSI. The diagnosis of SSI was based on the CDC criteria (Centre for Disease Control and Prevention) [14]. The primary diagnostic criteria for SSI were as follows.
(1) the surgical wound showed redness, swelling, heat, pain, wave sense or combined with abscess, and drooling(2) Bacterial culture of wound exudate was positive (3) Patients with re debridement surgery, positive intraoperative lavage uid or tissue bacterial culture (4) Laboratory tests for blood routine, Creactive protein, procalcitonin, as well as MRI examination or histopathological examination were performed to con rm SSI.

Data collection
The personal and medical information of patients, including age, body mass index (BMI), smoking, alcohol, chronic heart failure (CHF), diabetes, COPD, stroke, rheumatic disease, and diagnosis. The preoperative laboratory examinations, including blood routine examination, glycated hemoglobin A1c (HbA1c), liver function test, serum potassium, and serum calcium, were collected from the electronic medical record (EMR) system. Meanwhile, we recorded the surgery-related parameters, including operation time, intraoperative blood loss, surgical level, and the number of fusion segments in this investigation.

Statistical analysis
The statistical analyses were performed using SPSS for Windows software (ver. 26.0; SPSS Inc., Chicago, IL, USA). For continuous, normally distributed variables, means were calculated and compared using the independent student t test; otherwise, the Mann-Whitney U test was used for group comparisons. The chi-square test was used to analyze qualitative variables. Values of P < 0.05 were considered signi cant.

Training and evaluation of ML models
Forward step-wise multivariable logistic regression analysis, including covariates identi ed in univariate logistic analysis, was used to identify independent factors associated with SSI following MIS-TLIF under Quadrant channel. Meanwhile, the odds ratio (OR) and 95% con dence intervals (CIs) were reported.
In its most basic sense, machine learning uses programmed algorithms that learn and optimize their operations by analyzing input data to make predictions within an acceptable range [15][16][17][18]. In this study, 10-fold cross-validation was adopted, which means patients were randomly divided into a training set and a validation set at a ratio of 9:1 in each round. After 10 rounds, where all the patients are used for the test, the evaluation process ends. We developed ve types of ML algorithms to model our data: k-Nearest Neighbor (KNN), Decision tree (DT), Support vector machine (SVM), Random Forest (RF), Multi-Layer Perceptron (MLP), Naïve Bayes (NB). When training the prediction model, we used 10-fold crossvalidation (CV) on the training data to avoid over tting and nd the optimal hyperparameters. Then, Python software was used to further train the ML algorithms. To assess the performance of the ML algorithms, the area under the receiver operating characteristic (AUC) value, and sensitivity, speci city, the whole dataset accuracy (ACC), and confusion matrix of ML algorithms were reported. Of these, the ACC was determined by the following formula: In these formulas, TP: true positive; TN: true negative; FP: false positive; and FN: false negative. When we evaluated the performance of different ML algorithms, the closer the AUC was to 1, the better the classi cation performance of the model. In general, AUC > 0.7 indicated good performance of the model. We next determined the AUC values by comparing different ML algorithms, with the best performing algorithm as the nal predictive model. Data analysis was performed using programs written in Python programming. The Python packages, including Scikit-learn 0.24.2 [19] and Python 3.8, are adopted for ML algorithms. A two-tailed P < 0.05 was deemed statistically signi cant. The owchart of the study is presented in Fig. 1.
There were no signi cant differences in remaining parameters between the SSI and Non-SSI groups. The detailed results of the parameters are shown in Table 1.

Predictive ML algorithms performance
Comparisons of the performance of prediction among the six ML algorithms models in the whole set are detailed in Fig. 2. The result of the 10-fold CV of six ML algorithms in the whole set was demonstrated in Fig. 3. It turned out that the NB model demonstrated the highest performance in predicting SSI, whose average AUC was 0.86, sensitivity 0.667, speci city 0.89, and accuracy 0.90 in whole sets. Accordingly, we chose the NB model as the nal prediction model.

Relative importance of variables
According to the indicators, the NB model is determined as the nal prediction model. The relative importance of variables in the NB model was shown in Fig. 4. The importance of high-ranking variables in the NB model was arranged as follows in descending order: preoperative HbA1c, EBL, operation time, preoperative Alb, BMI, and age.

Discussion
In this work, we developed and validated multiple ML algorithms to predict the risk of SSI following MIS-TLIF under the Quadrant channel. We found that the NB model provided improved predictive ability compared to other models. This study produced interesting ndings concerning the added value of machine-learning methods, contrary to some of the wider literature on clinical prediction models. This ML-based model may potentially assist clinical decision-making when assessing and advising patients after MIS-TLIF under the Quadrant channel and when deciding on ongoing management.
MIS-TLIF is well known to have many advantages compared to open TLIF in terms of reducing muscle injury, smaller incisions, and scarring, especially for elderly patients who can not bear the trauma of traditional open TLIF surgery [20][21][22]. In contrast, SSI after MIS-TLIF can cause additional prolonged antibiotic use, prolonged length of stay, increased risk of fusion failure, and even the need for re-invasive surgery or removal of the implant [23,24], which not only increases the medical burden on patients and society but also causes increased mortality in patients.
With the continuous development of minimally invasive spine surgery, MIS-TLIF has gradually attracted the attention of spine surgeons. It has been reported that the incidence of SSI after lumbar interbody fusion is 2.4-5% [25][26][27][28]. In this investigation, the incidence of SSI was 4.78, which was consistent with the previous report. Although risk factors associated with SSI after MIS-TLIF have been previously reported, little attention has been paid to the incidence of SSI following MIS-TLIF under the Quadrant channel. More importantly, we still lack a concise and reliable tool to identify high-risk groups and assist clinicians in their decision-making. Therefore, in this study, several ML-based models were constructed using AI techniques to predict the risk of SSI.
Many studies have shown that diabetes mellitus is an important risk factor for SSI after spine surgery [23,[29][30][31]. Usually, glucose metabolism is closely related to the structure and pathophysiology of blood vessels. Disorders of glucose metabolism are inextricably linked to vascular structure and dysfunction, as has been demonstrated in many studies. And vascular pathological alterations cause local microcirculatory disorders leading to tissue ischemia and hypoxia, which provide a favorable environment for the growth of anaerobic bacteria and pose a potential risk for delayed postoperative wound healing and surgical site infection. Also, systemic immune de ciency caused by diabetes is the main reason for the occurrence of SSI. Therefore, surgeons usually strictly control patients' perioperative blood glucose and control preoperative fasting glucose below 6.9 mmol/L before considering surgical treatment [32]. Interestingly, glycosylated hemoglobin has also been identi ed as a risk factor for SSI.
Chronic hyperglycemic state leads to cell-mediated immune and macrophage dysfunction, reduced overall immune function of the body, and diminished ability to respond to bacterial viruses. And the relative importance in our prediction model was ranked rst. Therefore, we need to be more cautious about the perioperative glycemic management of patients.
In the present study, we found that high BMI is also a risk factor for SSI. In a large retrospective study by Onyekwelu et al. [33], BMI > 30 kg/m2 was found to be a high-risk factor for SSI after lumbar spinal fusion. The problems caused by obesity are present in almost all orthopedic surgeries, not only in spine surgery. Obesity can lead to endocrine disruption and insulin resistance, causing impaired glucose tolerance and even diabetes. On the other hand, obese patients may have a variety of underlying diseases that cause a decrease in the patient's immune capacity. Equally worrisome is the increased vascularity in the adipose tissue of obese patients and the even greater incidence of postoperative fat liquefaction, which also increases the risk of SSI.
Intriguingly, age was also identi ed as an independent risk factor in this study. However, we are also aware that this is a non-modi able risk factor. Previous studies reported that the risk of SSI after spinal surgery was three times higher in patients aged ≥ 60 years than in the younger group [1,27]. Similar results were observed in the present study. Although it is still controversial whether advanced age is a risk factor for SSI, the current mainstream of scholars believe that elderly patients have reduced metabolism, degenerated immune function, and may be accompanied by numerous underlying diseases, which signi cantly reduce the reserve capacity of the organism and make it more sensitive to surgical trauma strikes. The ability of local tissues to resist infection as well as self-repair is weakened, increasing the risk of SSI following spine surgery.
Although univariate logistics regression analysis indicated a statistical difference in operation time and EBL between the two groups (P < 0.05), in multivariate logistic regression analysis, we found that EBL and operation time were identi ed as risk factors for SSI. It has been reported in the literature that prolonged operation time is a signi cant risk factor for SSI that can not be ignored. Prolonged operation time leads to increased tissue stretching time, aggravating local ischemia and tissue in ammatory factor release, while prolonged surgical wound exposure leads to an increased possibility of direct blood entry of pathogenic bacteria from the wound. Whereas the relationship between EBL and SSI has been debatable.
According to previous literature, intraoperative blood transfusion > 1000 ml was reported as a risk factor for postoperative SSI in spine surgery and cardiac surgery [7,27]. Pullter et al. [33] investigated 3175 patients undergoing spine surgery and found that the incidence of postoperative SSI was 2.5% in patients with intraoperative bleeding < 1000 ml, and 3.5% in patients with intraoperative bleeding > 1000 ml. The incidence of SSI was 5.6% (p = 0.001) and was identi ed as an independent risk factor for SSI.
In the present study, we found that low serum albumin was an important risk factor for SSI. Serum albumin level is closely related to the nutritional status, liver and kidney function of patients. Hypoproteinemia will lead to immune suppression, local tissue edema, impaired microcirculation, and increased tissue uid exudation, which will affect normal wound healing and create conditions for bacterial colonization. Previous literature has reported preoperative hypocalcemia as a risk factor for postoperative spinal infection [30], and another interesting nding is that electrolyte variables such as blood calcium and potassium in this investigation were not statistically signi cantly different between the SSI and Non-SSI groups. Patients with toxic shock syndrome are usually found to suffer from hypocalcemia, which may be due to an increase in calcitonin secretion caused by the infection, resulting There is, however, little research in the available literature on applying ML algorithms to SSI following MIS-TLIF. As far as we know, this is the rst study to develop a prediction model using ML algorithms for risk evaluation of SSI with easy-to-use clinical and laboratory data. Fortunately, our prediction model evidences good predictive performance of the model, which distinguishes itself from linear models adopted by previous research.
In this study, several limitations of this study require attention. First, the nature of a retrospective study might have introduced a selection bias. Second, the ML algorithm model we established, to some extent, was con ned to one single center, which might restrict its generalizability pending further validation in real-world scenarios. Third, further independent external validation is required to con rm these ndings.

Conclusions
We developed and validated ML algorithms for individualized prediction of SSI following MIS-TLIF under the Quadrant channel by utilizing readily available preoperative variables. The ML-based prediction model can identify whether patients are at high risk of SSI and may assist the clinician in decision making. In the future, we hope to enlarge the sample size in the following study and obtain more reliable