Machine learning with interpretability predict surgical site infection after posterior cervical surgery

Ideal tools should not only investigate risk factors, but also provide explicit auxiliary answer for whether a patient will develop surgical site infection (SSI) or not. Machine learning (ML) models have ability to carry out complicated predictive medical tasks. We intend to develop ML models to predict SSI after posterior cervical surgery and interpret the outcome. We retrospectively analyzed 235 patients who had undergone posterior cervical surgery between June 2013 to April 2019 at Zhongda Hospital Aliated to Southeast University. We established Articial neural networks (ANN), XGBClassier (xgboost), KNeighborsClassier (KNN), Decision tree classier (decision tree), Random forest classier (random forest) and support vector classier (SVC). Receiver operating characteristic (ROC) curve, area under the curve (AUC) score, accuracy score, recall score, F1 score and precision score were calculated to measure models’ performance. Shapley values were calculated using SHapley Additive exPlanations (SHAP) to determine relative feature importance of xgboost model.


Conclusion:
We had successful established ML models in individualized predicting SSI after posterior cervical surgery.
Xgboost, ANN, decision tree and random forest achieved excellent performance which could provide auxiliary information for clinical decision makers. The interpretable model focuses on contribution of important features to the predictive result. It can improve the acceptance of clinicians on ML and promote ML's application in the actual clinical work.

Background:
Posterior cervical surgery is the major surgical approach for the treatment of cervical diseases. Surgical site infection (SSI) is one of most common severe postoperative complications following spinal surgery.
[1] The rates of SSI after spinal surgery had been reported ranging from 0.7-12.0% [2]. The SSI can induce devastating complications including neurological injury, paralysis, pseudarthrosis, sepsis and death, with prolonging the length of hospital stay and increasing costs [1,3,4]. Identifying risk factors associated with SSI is a signi cant focus in past researches.
Previous studies had identi ed many risk factors of SSI after spinal surgery, including age, diabetes, body mass index (BMI), malnutrition, American Society of Anesthesiologists score (ASA score), smoking status, revision, posterior approach, operative time, osteotomy, and fusion length. [2,5,6] Although many studies had investigated factors associating with SSI, it mainly based on clinicians' analytical judgment, that recognizing whether a patient will develop SSI or not and deciding subsequent treatment. Ideal tools are needed to provide explicit auxiliary answer to doctors.
Though logistic regression analysis had been developed to predict SSI after instrumented thoracolumbar spine surgery and showed reasonable discriminative ability and calibration. [7] But logistic regression analysis is static and often relies on prede ned relationships (eg, linear), which makes it less robust for large data sets with complex relationships between patient characteristics and outcomes.
[8] In recent years, arti cial intelligence (AI) has made great progress in medical management. As a major branch of AI, machine learning (ML) has ability to learn from complicated data and nonlinear relationships[8], and achieves better performance from big data. To our best knowledge, no studies have been conducted to use ML models predicting SSI after posterior cervical surgery. Therefore, the purpose of this study is to develop ML models for predicting SSI and interpret the outcome.

Methods:
1. Patient population and data processing We retrospectively analyzed 235 patients who had undergone posterior cervical surgery for cervical spondylosis, ossi cation of the posterior longitudinal ligament (OPLL), cervical stenosis, cervical acute trauma, cervical tumors and atlantoaxial disorder between June 2013 to April 2019 at Zhongda Hospital A liated to Southeast University. The Diagnostic criteria of SSI were as follows [9]: (1)Purulent drainage, with or without laboratory con rmation, from the super cial incision or deep incision; (2) Organisms isolated from an aseptically obtained culture of uid or tissue from the incision; (3) The incision spontaneously dehisces or is deliberately opened by a surgeon when the patient has at least one of following signs or symptoms: fever (>38ºC), localized pain, or tenderness, localized swelling, redness, or heat, unless site is culture-negative; (4) An abscess or other evidence of infection is found on direct examination, during reoperation, or by histopathologic or radiologic examination; The following variables were analyzed as input: sex, age, BMI, diabetes, hypertension, preoperative nutritional status (including serum albumin, serum total cholesterol, total blood lymphocyte count), the prognostic nutritional index (PNI) (10 × serum albumin [g/dL] + 0.005 × total lymphocyte count [/μL]), [5] operation time, intraoperative blood loss, incision length, postoperative drainage and drainage time. BMI≥25 is overweight. Parameters used to train models were shown in Table 1. To overcome the low sample size for SSI cohorts, the SMOTE approach, con rmed in previous research [10], was utilized to generate SSI sample for reducing class imbalance. Then the dataset was randomly divided into training set and test set at 7:3. Gradient descent method was used to reduce loss of the ANN, and dropout was added to inhibit over tting. The training was performed for 1300 epochs and the learning rate was 0.0005. Receiver operating characteristic (ROC) curve, area under the curve (AUC) score, accuracy score, recall score, F1 score and precision score were calculated to measure models' performance. Shapley values were calculated using SHapley Additive exPlanations (SHAP) to determine relative feature importance, for improving interpretability of our xgboost model. [11] Results: There were 58 females and 177 males who underwent posterior cervical surgery, in which 17 patients were diagnosed as SSI and the incidence of SSI was 7.23%. Total patients had a mean age of 60.10 (SD: ± 11.29) years, BMI of 24.36 (SD: ± 3.20) kg/m 2 , which shown in Table 2. Performance of models on accuracy score, recall score, F1 score and precision score were shown in Table 3. Xgboost, ANN, decision tree and random forest achieved excellent prediction result in testing set.

Discussion:
Plenty of studies had investigated risk factors of SSI after spinal surgery.[2, 5, 6] However, when a patient with several factors needs spine surgery, the exact SSI incidence possibility is ambiguous and the treatment decision still depended on clinicians' experience. While intelligent algorithms based on previous complicated patient data will provided a wealth of information. A paper had applied a Deep Neural Network model using 35 unique variables to predict SSIs after posterior spinal fusions. [12] They mainly pay attention on demographics, combined disease and operation type. [12] We hold that more speci c variables are necessary to improve ML predictive models. To our best knowledge, no study had yet explored ML models to predict SSI after posterior cervical surgery.
Our models shown reliably predive performance based on patient data from this institution. Xgboost, random forest, ANN and decision tree got predictive accuracy more than 90% in test set. The ability to predict SSI may assist surgeons in identifying potential risk patients, better selecting candidates for posterior cervical surgery, managing patient expectations and timely intervene to improve outcomes.
These lends to precision-based spine care and more bespoke management to optimize patient outcomes and improve daily function, with decreasing healthcare-related costs. [11] We established sex different models to nd the most robust one, which contain one deep learning and two model ensembles. Deep learning has superior tting ability than traditional models, suitable to big data. Model ensembles, like xgboost and random forest, combines plenty weak classi ers to increase accuracy on ML tasks. However, in this study xgboost and random forest got slightly better predictive ability than ANN, the deep learning algorithm, which may explain by the low order of magnitude data.
Interestingly, the decision tree also achieved satisfactory result after grid searching for best parameters. This may due to our ne-tuning work and well feature selection.
We input more individual features like, nutritional status and operation data, differing from national level database taking into account plenty demographic and general management variables. We thought these may better providing personalized preoperative status and provide more individual information. However, these variables were all manual inputted and laborious.
A main challenge of ML needed to overcome is Black box, that limits application of ML in medicine. It is that most AI technologies operate based on opaque logic and hardly understandable to users. Interpretation of xgboost based on SHAP is a solution of Black box. Beyond get an unexplainable prediction, we could examine contribution of each valuable on the individual prediction. Two examples are shown in Fig 3 and Fig 4. Drainage volume and drainage duration were identi ed as predictors in the present study. Prolonged drainage duration has frequently been cited as independent risk factors for SSI. [4,13,14] Liu et al used multivariable analysis to con rm that prolonged drainage duration was a risk factor for SSI after lumbar spinal surgery. 4 Moreover, Rao et al suggested to remove the drain as early as possible for reducing infection rate following spinal fusion. 17 Drains may induce local tissue in ammation and become direct access for bacteria by ascending the drain tube, thus increasing the risk of infection. 18 Also, a study demonstrated two-fold reduction of SSI with implementation of prolonged prophylactic systemic antibiotics regimen for the duration of drain. [15] Previous researches had suggested a positive relationship between BMI and SSI. Pesenti et al con rmed that obesity is a signi cant risk factor for SSI after spinal arthrodesis and BMI > 30 kg m −2 as a signi cant risk factor in most of studies.
[6] Another meta-analysis found 21 % increase in risk of spinal SSI for every 5-unit increase in BMI, after adjusting for diabetes and other confounders.
[16] It has been reported that increasing subcutaneous adipose tissue raises the likelihood of fat necrosis and thus increases infection risk. [17] A multivariate analysis found that perioperative blood loss≥500 mL was a risk factor for SSI in spine surgery.
[18] Zhou et al indicated that the SSI incidence of blood loss 500 mL was twice those blood loss<500 mL. [19] Meanwhile, they found surgical time≥3 hours have higher incidence of spinal SSI, which may be explained by that prolonging operative duration increases the chance of contamination in surgical wounds. [19] However, some studies found no relation between the surgical time and SSI. [20] Besides "operative time", total anesthetic time is also an independent predictor of SSI, when patient remains in the operative room environment. [21] Our model identi es cholesterin, PNI and albumin as risk factors, while the total cholesterol had no signature between SSI and non-SSI in previous researchs. [22,23] Lower preoperative PNI and lower serum albumin were found be risk factor for SSI after spine surgery. [4,5,24]  Although performance of ML models on SSI after posterior cervical surgery are robust, this study has noteworthy limitations. First, the sample of this study is small, and further larger sample inputting may improve prediction results. Also, the data retrieved from a single institution, and further external validation is necessary to elevate models' expansibility.

Conclusions:
We

Declarations
Ethics approval and consent to participate: The study had been approved by Zhongda Hospital A liated to Southeast University ethics committee and the reference number was 2021ZDSYLL222-P01. This study had met the requirements for exemption from informed consent, and approved to exemption from informed consent by Zhongda Hospital A liated to Southeast University ethics committee.
Consent for publication: Not Applicable.
Availability of data and materials: The datasets used and analysed during the current study are available from the corresponding author on reasonable request.
Competing interests: The authors have no relevant nancial or non-nancial interests to disclose.
Funding: No funding was received for conducting this study.
Authors' contributions: GuanRui Ren and ZhiYang Xie contributed to the study conception and design.  Patient Example 2. Low probability of non-SSI. PNI (prognostic nutritional index)