PICOT question

O R I G I N A L A R T I C L E The Journal of Nursing Research ▪ VOL. 29, NO. 1, FEBRUARY 2021

Predicting the Development of Surgery-Related Pressure Injury Using a

Machine Learning Algorithm Model Ji-Yu CAI1 • Man-Li ZHA1 • Yi-Ping SONG1 • Hong-Lin CHEN2*

1BSN, Graduate Student, School of Nursing, Nantong University, Nantong City, Jiangsu, People’s Republic of China • 2MS, Associate Professor, School of Nursing, Nantong University, Nantong City, Jiangsu, People’s Republic of China.

Copyright © 2020 The Authors. Published by Wolters Kluwer Health, Inc.

This is an open access article distributed under the Creative Commons Attribution License 4.0 (CCBY), which permits unrestricted use, distri- bution, and reproduction in any medium, provided the original work is properly cited.

ABSTRACT Background: Surgery-related pressure injury (SRPI) is a serious problem in patients who undergo cardiovascular surgery. Identi- fying patients at a high risk of SRPI is important for clinicians to recognize and prevent it expeditiously. Machine learning (ML) has been widely used in the field of healthcare and is well suited to predictive analysis.

Purpose: The aim of this study was to develop an ML-based predictive model for SRPI in patients undergoing cardiovascular surgery.

Methods: This secondary analysis of data was based on a single-center, prospective cohort analysis of 149 patients who underwent cardiovascular surgery. Data were collected from a 1,000-bed university-affiliated hospital. We developed the ML model using the XGBoost algorithm for SRPI prediction in pa- tients undergoing cardiovascular surgery based on major poten- tial risk factors. Model performance was tested using a receiver operating characteristic curve and the C-index.

Results: Of the sample of 149 patients, SRPI developed in 37, an incidence rate of 24.8%. The five most important predictors included duration of surgery, patient weight, duration of the cardio- pulmonary bypass procedure, patient age, and disease category. The ML model had an area under the receiver operating character- istic curve of 0.806, which indicates that the ML model has a mod- erate prediction value for SRPI.

Conclusions: Applying ML to clinical data may be a reliable approach to the assessment of the risk of SRPI in patients un- dergoing cardiovascular surgical procedures. Future studies may deploy the ML model in the clinic and focus on applying targeted interventions for SRPI and related diseases.

KEY WORDS: surgery-related pressure injury, machine learning, risk assessment, cardiovascular surgery.

Introduction Pressure injury is generally defined as localized damage to the skin and underlying soft tissue, usually because of the lo- cation of the skin and tissue over a bony prominence or of the use of a medical or other device (Xiong et al., 2019). During surgery, patients are affected by procedure-related factors

such as perioperative fasting, liquid fasting, postanesthesia compulsive position, and disinfectant-induced damp skin (Gao et al., 2018). Therefore, patients face an elevated risk of experiencing a surgery-related pressure injury (SRPI). A meta-analysis reported a general prevalence of SRPI of 18.96% (95% CI [15.3, 22.6]) among patients (Shafipour et al., 2016). An incidence rate of SRPI ranging from 0.3% to 57.4% and 18% among patients who underwent cardio- vascular surgery was identified in a systematic review (95% CI [14.0, 22.0]; Chen et al., 2012). Pressure injury is an im- portant safety indicator in healthcare systems. Pressure inju- ries not only adversely affect quality of life but also drain resources from healthcare systems worldwide (Girouard et al., 2008; Liao et al., 2018).

Commonly identified risk factors for SRPI include surgi- cal positioning, type of anesthesia, duration of surgery, extra- corporeal circulation, and pressure from internal retractors or from operating room staff (Campbell, 1989; Papantonio et al., 1994; J. Walsh, 1993). Evidence from clinical trials suggests that pressure injury is preventable in today's modern healthcare environment (Thomson & Brooks, 1999). Assessing the risk of pressure injury is recommended in clin- ical nursing care. Unfortunately, although some risk assess- ment tools for SRPI have been developed, they have limitations. Although the Braden Scale is a validated and widely used instrument for assessing pressure injury risk, this scale was developed for use in other care settings. The valid- ity and reliability of using the Braden Scale to assess pressure injury development have been established in a variety of pa- tient care settings. However, the results of a meta-analysis re- vealed that the Braden Scale had a low predictive value for SPRI development in patients who underwent surgery (He et al., 2012). Alternatively, the modified Norton scale has been used frequently in German hospitals. However, the sensitiv- ity and specificity of this scale are 41% and 88%, respectively

1

The Journal of Nursing Research Ji Yu CAI et al.

(Feuchtinger et al., 2007). The Waterlow score is another widely used instrument for predicting pressure injury development. The studies indicated that this scale, although suitable for predicting postoperative morbidity and mortality in surgical patients, did not predict the likelihood of SRPI formation (Thorn et al., 2013). These traditional risk assessment scales are based on clin- ical experience dating from the 1970s and 1980s and lack suffi- cient data and evidence-based support.

Machine learning (ML), a recently developed field of smart technology, is increasingly being applied in the con- struction of predictive models. Incorporating ML improves the performance of predictive models and does not require explicit programming or manual guidance (Ethem, 2004; Marsland, 2009). Thus, ML is now widely recognized as an effective resource in medicine and healthcare. In a previ- ous study, an ML-based predictive model constructed for uri- nary tract infections in the emergency department was superior to existing predictive models (Taylor et al., 2018). Another ML algorithm, XGBoost, was used to construct a risk prediction model for incident essential hypertension and achieved satisfactory predictive accuracy (Ye et al., 2018). Ap- plying ML to longitudinal clinical data has provided a scal- able tool for expanding screening for risk of nonfatal suicide attempts in adolescents (C. G. Walsh et al., 2018). Furthermore, ML techniques effectively and efficiently use large amounts of clinical data and are well suited for use in predicting SRPI.

The aims of this study were to integrate ML information with clinical information to predict SRPI risk among patients and to validate the validity of the developed ML predictive model as a reference for future studies.

Methods

Sample This study adopted a secondary analysis approach using data from a prospective study of patients who had consecutive car- diac surgeries that was conducted to predict the incidence of SRPI using ML (Lu et al., 2017). The sample consisted of data on patients who underwent cardiac surgery and aortic surgery at a teaching hospital between January 2015 and December 2015. The inclusion criteria included all patients, regardless of age, with a pressure injury at the time of admission and be- fore surgery. This study was approved by the ethics committee of the School of Nursing, Nantong University (Approval No. 2018056) in China.

Data Collection Data for each patient were obtained from the original health records. A wide range of relevant predictors noted in prior studies, including demographic characteristics, SRPI infor- mation, and corticosteroid information, among others, were considered. Demographic characteristics included age, gen- der, weight, and disease category; SRPI information in- cluded number of ulcers, ulcer severity as determined using the National Pressure Ulcer Advisory Panel classification

2

(Choi et al., 2016), anatomical location, and outcome; corti- costeroid information included administration, type of drug used, drug dosage, and frequency of drug administration; and risk factors included use of vasoactive drugs, experiencing hypotensive periods, hemoglobin level, albumin level, and use of corticosteroids (Feuchtinger et al., 2005).

Model Construction Models were constructed using all of the variables collected, and descriptive statistics were used to compare the baseline characteristics and outcomes. Univariate chi-square tests and t tests were used to compare categorical variables and continuous variables, respectively. P values (< .10) were deemed significant. The prospective data were used to con- struct the prediction model. XGBoost, an algorithm in ML, was used to generate predictive estimates on the basis of fea- tures retained in the univariate analysis (Plagnol et al., 2012). XGBoost is a novel, sparsity-aware algorithm used in condi- tions of sparse data and weighted quantile sketch for approx- imate tree learning. It is designed to improve the speed and performance of gradient-boosted decision trees. The final predictive estimate is calculated by summing the scores in the corresponding leaves of each tree (Ye et al., 2018). XGBoost adds an estimator to provide a better approximation. At each iteration, a new prediction model is built, and each model learns to correct the previous stage model. XGBoost does not require linear features or linear interactions between features and thus is a significantly better classifier than other al- gorithms. The model usually refers to the mathematical struc- ture of how to make prediction-dependent variables given the independent variables. This algorithm has been recognized as having good accuracy, flexibility, and speed.

Model Evaluation The performance of the model was evaluated based on the mixture matrix. The primary indicator of this prediction model was the area under the curve (AUC) of the receiver op- erating characteristic. The secondary indicators were sensitiv- ity and accuracy with 95% CI. Most classification models have an AUC between 0.5 and 1, a random classifier has an AUC of 0.5, and a perfect classifier has an AUC of 1 (Qiao et al., 2018). The definition of sensitivity is the proportion of positive results out of the number of samples that were posi- tive and the proportion of negative results out of the number of samples that were negative (Taylor et al., 2018).

Results

Patient Characteristics This study included data from 149 patients, with ages rang- ing from 4 to 77 years. Seventy-nine patients were male, and 70 were female. In terms of disease category, 35 of the patients had congenital heart disease, 85 had valvular heart disease, 20 had coronary heart disease, and nine had macrovascular

Figure 1 Receiver Operating Characteristic Curve for Surgical-Related

Machine-Learning-Based Risk Prediction Model VOL. 29, NO. 1, FEBRUARY 2021

disease. Selected patient characteristics stratified by cardio- vascular surgical patients are shown in Table 1.

Pressure Injuries by Machine Learning Model Versus Logistic Regression

Model Performance

The ML tool XGBoost was selected to construct the SRPI prediction model. The outcome evaluation index was the oc- currence of SRPI. A score of 1 would be assigned to patients who developed SRPI, and a score of 0 would be assigned oth- erwise. We evaluated the SRPI prediction model in the form of confusion matrices, with sensitivity, specificity, and the Youden Index. In the prediction model, three patients were correctly predicted as positive for SRPI, and 34 were erro- neously predicted as negative. One hundred twelve patients with no SRPI were correctly predicted as negative. The values of sensitivity and specificity were, respectively, 8.11% and 100%. The Youden Index was derived using the thresh- old at which the sum of sensitivity and specificity achieves the highest value. The value of Youden Index was calculated as 0.081.

Table 1

Baseline Patient Characteristics (N = 149)

Characteristic Without SRPI

(n = 112) With SRPI (n = 37)

pn % n %

Gender .815 Male 60 53.6 19 51.4 Female 52 46.4 18 48.6

Disease category .074 Congenital heart disease 32 28.6 3 8.1 Valvular heart disease 61 54.5 24 64.9 Coronary artery disease 13 11.6 7 18.9 Thoracic aortic aneurysms 6 5.4 3 8.1

Vasoactive agents intraoperatively

.768

Yes 30 26.8 9 24.3 No 82 73.2 28 75.7

Vasoactive agents postoperatively

.737

Yes 76 67.9 24 64.9 No 36 32.1 13 35.1

Corticosteroids perioperative .018 Yes 5 4.5 6 16.2 No 107 95.5 31 83.8

M SD M SD p

Age (years) 48.2 18.3 54.7 15.0 .053

Weight (kilograms) 59.8 15.3 59.1 15.4 .805

Surgery duration (minutes) 221.7 85.8 263.6 93.0 .013

CPB duration (minutes) 48.9 23.1 48.9 23.1 .996

Note. SRPI = surgery-related pressure injury; CPB = cardiopulmonary bypass.

Positive predictive value and negative predictive value were used to describe the performance of the SRPI prediction model. The positive predictive value was 100%, indicating that 100% of patients who developed SRPI were predicted to do so. The negative predictive value was 76.71%, indicat- ing that 76.71% of patients with no SRPI were predicted to do so.

By adopting the ML tool XGBoost, the developed pres- sure injury prediction model performed at an AUC value of 0.806. Figure 1 shows the receiver operating characteristic curve of the ML prediction model.

Predictor Importance

Nine predictors entered the ML model. These predictors were patient age, gender, disease category, weight, duration of surgery, duration of cardiopulmonary bypass procedure, perioperative corticosteroid administration, use of intraop- erative vasoactive agents, and use of postoperative vasoac- tive agents.

Finally, the importance metrics were aggregated to sum- marize the five predictors that were important in this ML model. These included, in rank order of decreasing impor- tance, (a) duration of surgery (in minutes), (b) weight (in ki- lograms), (c) duration of cardiopulmonary bypass procedure (in minutes), (d) age (in years), and (e) disease category (e.g., congenital heart disease). The importance of these five risk factors in predicting SPRI was 0.426, 0.193, 0.131, 0.126, and 0.124, respectively. Findings indicate that duration of surgery was the most important risk factor for SRPI. The proportional importance of each input variable is shown in Figure 2. Risk factors of insignificant importance, including gender, perioperative corticosteroid administration, use of

3

Figure 2 Relative Importance of Studied Variables to Surgical-Related Pressure Injuries

The Journal of Nursing Research Ji Yu CAI et al.

intraoperative vasoactive agents, and use of postoperative vasoactive agents, are not shown.

Discussion In this prospective observational study of surgical patients, we constructed a risk prediction model of SRPI development. The prediction model achieved an AUC of 0.806, indicating that this model has moderate predictive accuracy for SRPI in patients undergoing cardiovascular surgery.

In this study, an ML algorithm called XGBoost was adopted for feature selection and model construction. XGBoost is a nonparametric algorithm that does not assume a functional relationship between outcome and features, as is required for linear regression models (Jensen et al., 2012). As a result, this supervised ML method is a robust approach for han- dling correlated features. The resulting ML prediction model captured potentially powerful predictors of the development of SRPI. In this study, the highest risk of SRPI was borne by patients with longer surgery and cardiopulmonary bypass procedure durations, lower body weight, older age, and con- genital heart disease.

Predictive models for SRPI have been developed previ- ously. In a prior study conducted by the authors of this study, a nomogram score was constructed on the basis of logistic re- gression to predict the development of SRPI (Lu et al., 2017). The constructed logistic regression model contained four in- dependent risk factors: duration of surgery, weight, disease category, and perioperative corticosteroid administration. The three risk factors duration of surgery, weight, and dis- ease category were consistent with the ML model developed in this study. Moreover, age and duration of cardiopulmo- nary bypass procedure were also considered as important factors in the ML model. The logistic regression model was

4

shown to have moderate power for predicting SRPI, with an AUC of 0.725, an outcome that was similar to this study. However, the ML model has more accurate discrimination power than the nomogram score. ML has been proposed by several authors as an approach to wound-tissue recognition. Kosmopoulos and Tzevelekou used ML for pressure injury di- agnosis and presented some exploratory results (Kosmopoulos & Tzevelekou, 2007). Kaewprag et al. used the Bayesian net- work algorithm in ML to explore risk factors for pressure in- juries in patients in intensive care units. That study indicated that the sensitivity of the ML predictive model was nearly three times higher than the logistic regression model, with no decline in overall accuracy (Kaewprag et al., 2017). Therefore, similar to logistic regression, ML may also be used as a technique along with data mining to improve assess- ment of risk of the development of SRPI.

The five most important variables based on the mean de- crease in SRPI accuracy were, in descending order, duration of surgery, body weight, duration of cardiopulmonary by- pass procedure, age, and disease category. A previous study (Chen et al., 2018) built an artificial neural network model to investigate the independent risk factors for SRPI in pa- tients undergoing cardiovascular surgery. The factors identi- fied included disease category, perioperative corticosteroid administration, age, and duration of surgery, and the impor- tance of these factors to predicting SRPI was 0.268, 0.136, 0.237, and 0.360, respectively. In this study, two new risk factors, namely, weight and duration of cardiopulmonary bypass procedure, were identified. Duration of surgery is rec- ognized as a high-risk factor for the development of SRPI. Ex- tended duration of procedures leads to increased duration of hypoperfusion, ischemia of local compressed tissues, and de- creasing temperature of the compressed position skin, which increase the risk of SRPI (Chen et al., 2017; O'Connell, 2106; Shen et al., 2015). According to a previous study, every 1-hour extension in surgery duration increases the risk of SRPI by 96% (Gao et al., 2018). In a meta-analysis of the as- sociation between duration of surgery and SRPI risk, the point estimates for surgery duration at 300 and 600 minutes, re- spectively, increased the SRPI risk by 3.653 and 13.344 times that of the risk at 60 minutes (Chen et al., 2017).

Weight was identified in this study as an important vari- able affecting SRPI risk. Although previous studies have shown that poor nutritional status is a common risk factor for pressure injury, only one previous study reported finding a correlation between higher risk of pressure injury and lower body mass in surgical patients (Gao et al., 2018). Most previous studies have cited body mass index as a potential risk factor and showed low body mass index as a significant predictor of SRPI (Alderden et al., 2018; Aloweni et al., 2019). In this study, we did not include body mass index, be- cause it is not clear whether it is an independent risk factor for SRPI. Future studies may further confirm the predictive effect of body mass on SRPI.

The impact of cardiopulmonary bypass procedure duration was not adequately explored in a previous systematic review

Machine-Learning-Based Risk Prediction Model VOL. 29, NO. 1, FEBRUARY 2021

(Rao et al., 2016). In this study, duration of cardiopulmonary bypass procedure was deemed as one of the most significant of the examined variables. Our finding that patients who un- dergo cardiopulmonary bypass are more likely to develop SRPI agrees with other previous studies (Alderden et al., 2018; Gao et al., 2018). Moreover, SRPI experienced after cardiopulmo- nary bypass may be related to poor peripheral circulation perfusion caused by intraoperative hypothermia, utilization of a warming blanket after cardiopulmonary bypass, and/or lack of subcutaneous tissue protection at heel (Gao et al., 2018). Further studies should pay attention to the effective prevention of heel-related SRPI.

The variables that were found to have no significant predic- tive effect are also informative for future research. Corticosteroid is acommon variable associatedwith SRPI in cardiovascular sur- gical patients, as the use of corticosteroid decreases the levels of growth factor, which is deemed as an important factor for pres- sure injury development (Feuchtinger et al., 2005; Wicke et al., 2000; Yang et al., 2013). However, in this study, corticosteroid administration was not found to be an important risk factor. Future researchers may take into account these factors and multidisciplinary wound care.

Limitations This study had limitations. First, the data were from a single healthcare institution within a confined geographic region. Thus, the generalizability of our findings may be limited. Sec- ond, not collecting data prospectively may affect the perfor- mance of the ML prediction model developed in this study. Third, the severity of all SRPI instances in this study were Stage 1.

Conclusion An ML model for predicting SRPI risk in cardiovascular sur- gical patients was constructed in this study. Integrating pre- dictive analysis into clinical care holds the potential to better identify high-risk patients and provide appropriate predictive intervention. Future studies may build on these findings to develop a potentially more robust and sensitive ML model for predicting SRPI risk.

Acknowledgments Thisstudy was supported by the Postgraduate Research & Prac- tice Innovation Program of Jiangsu Province (KYCX18_2430). We thank all of the people who participated in this study.

Author Contributions Study conception and design: JYC, HLC Data collection: JYC, MLZ, YPS Data analysis and interpretation: JYC, MLZ, YPS Drafting of the article: JYC Critical revision of the article: HLC

Accepted for publication: January 10, 2020 *Address correspondence to: Hong-Lin CHEN, No. 19, QiXiu Road, Nantong City 226001, Jiangsu Province, People’s Republic of China. Tel: +86-513-85051856; E-mail: honglinyjs@126.com The authors declare no conflicts of interest.

Cite this article as: Cai, J.-Y., Zha, M.-L., Song, Y.-P., & Chen, H.-L. (2021). Predicting the development of surgery-related pressure injury using a machine learn- ing algorithm model. The Journal of Nursing Research, 29(1), Article e135. https://doi.org/:10.1097/jnr.0000000000000411

References Alderden, J., Pepper, G. A., Wilson, A., Whitney, J. D., Richardson,

S., Butcher, R., Jo, Y., & Cummins, M. R. (2018). Predicting

pressure injury in critical care patients: A machine-learning

model. American Journal of Critical Care, 27(6), 461–468.

https://doi.org/10.4037/ajcc2018525

Aloweni, F., Ang, S., Fook-Chong, S., Agus, N., Yong, P., Goh, M.,

Yong, P., Goh, M., Tucker-Kellogg, L., & Soh, R. C. (2019). A

prediction tool for hospital-acquired pressure ulcers among

surgical patients: Surgical pressure ulcer risk score. Interna-

tional Wound Journal, 16(1), 164–175. https://doi.org/10.1111/

iwj.13007

Campbell, K. (1989). Pressure point measures in the operating

room. Journal of Enterostomal Therapy, 16(3), 119–124.

Chen, H. L., Chen, X. Y., & Wu, J. (2012). The incidence of pressure

ulcers in surgical patients of the last 5 years: A systematic re-

view. Wounds, 24(9), 234–241.

Chen, H. L., Shen, W. Q., Liu, P., & Liu, K. (2017). Length of surgery

and pressure ulcers risk in cardiovascular surgical patients: A

dose–response meta-analysis. International Wound Journal,

14(5), 864–869. https://doi.org/10.1111/iwj.12722

Chen, H. L., Yu, S. J., Xu, Y., Yu, S. Q., Zhang, J. Q., Zhao, J. Y., Liu,

P., & Zhu, B. (2018). Artificial neural network: A method for predic-

tion of surgery-related pressure injury in cardiovascular surgical

patients. Journal of Wound, Ostomy, and Continence Nursing,

45(1), 26–30. https://doi.org/10.1097/won.0000000000000388

Choi, E. P. H., Chin, W. Y., Wan, E. Y. F., & Lam, C. L. K. (2016).

Evaluation of the internal and external responsiveness of the

Pressure Ulcer Scale for Healing (PUSH) tool for assessing

acute and chronic wounds. Journal of Advanced Nursing, 72(5),

1134–1143. https://doi.org/10.1111/jan.12898

Ethem, A. (2004). Introduction to machine learning. MIT Press.

Feuchtinger, J., Halfens, R., & Dassen, T. (2007). Pressure ulcer

risk assessment immediately after cardiac surgery—Does it

make a difference? A comparison of three pressure ulcer risk

assessment instruments within a cardiac surgery population.

Nursing in Critical Care, 12(1), 42–49. https://doi.org/10.1111/j.

1478-5153.2006.00198.x

Feuchtinger, J., Halfens, R. J., & Dassen, T. (2005). Pressure ulcer

risk factors in cardiac surgery: A review of the research litera-

ture. Heart & Lung, 34(6), 375–385. https://doi.org/10.1016/j.

hrtlng.2005.04.004

Gao, L., Yang, L., Li, X., Chen, J., Du, J., Bai, X., & Yang, X. (2018).

The use of a logistic regression model to develop a risk assess-

ment of intraoperatively acquired pressure ulcer. Journal of

Clinical Nursing, 27(15–16), 2984–2992. https://doi.org/10.

1111/jocn.14491

5

The Journal of Nursing Research Ji Yu CAI et al.

Girouard, K., Harrison, M. B., & VanDenkerkof, E. (2008). The

symptom of pain with pressure ulcers: A review of the litera-

ture. Ostomy Wound Management, 54(5), 30–40.

He, W., Liu, P., & Chen, H. L. (2012). The Braden Scale cannot be

used alone for assessing pressure ulcer risk in surgical patients:

A meta-analysis. Ostomy Wound Management, 58(2), 34–40.

Jensen, P. B., Jensen, L. J., & Brunak, S. (2012). Mining electronic

health records: Towards better research applications and clin-

ical care. Nature Reviews Genetics, 13(6), 395–405. https://doi.

org/10.1038/nrg3208

Kaewprag, P., Newton, C., Vermillion, B., Hyun, S., Huang, K., &

Machiraju, R. (2017). Predictive models for pressure ulcers from

intensive care unit electronic health records using Bayesian net-

works. BMC Medical Informatics & Decision Making, 17(2, Suppl.),

Article No. 65. https://doi.org/10.1186/s12911-017-0471-z

Kosmopoulos, D. I., & Tzevelekou, F. L. (2007). Automated pres-

sure ulcer lesion diagnosis for telemedicine systems. Engi-

neering in Medicine & Biology Magazine IEEE, 26(5), 18–22.

https://doi.org/10.1109/EMB.2007.901786

Liao, Y., Gao, G., & Mo, L. (2018). Predictive accuracy of the

Braden Q Scale in risk assessment for paediatric pressure ul-

cer: A meta-analysis. International Journal of Nursing Sci-

ences, 5(4), 419–426.

Lu, C. X., Chen, H. L., Shen, W. Q., & Feng, L. P. (2017). A new no-

mogram score for predicting surgery-related pressure ulcers

in cardiovascular surgical patients. International Wound Journal,

14(1), 226–232. https://doi.org/10.1111/iwj.12593

Marsland, S. (2009). Machine learning: An algorithmic perspective

(1st ed.). Chapman & Hall/CRC.

O'Connell, M. P. (2006). Positioning impact on the surgical patient.

Nursing Clinics of North America, 41(2), 173–192. https://doi.

org/10.1016/j.cnur.2006.01.010

Papantonio, C. T., Wallop, J. M., & Kolodner, K. B. (1994). Sacral

ulcers following cardiac surgery: Incidence and risks. Advances

in Wound Care, 7(2), 24–36.

Plagnol, V., Curtis, J., Epstein, M., Mok, K. Y., Stebbings, E.,

Grigoriadou, S., Wood, N. W., Hambleton, S., Burns, S. O.,

Thrasher, A. J., Krummararatne, D., Doffinger, R., & Nejentsev, S.

(2012). A robust model for read count data in exome sequenc-

ing experiments and implications for copy number variant call-

ing. Bioinformatics, 28(21), 2747–2754. https://doi.org/10.1093/

bioinformatics/bts526

Qiao, Z., Sun, N., Li, X., Xia, E., Zhao, S., & Qin, Y. (2018). Using

machine learning approaches for emergency room visit predic-

tion based on electronic health record data. Studies in Health

Technology & Informatics, 247, 111–115.

Rao, A. D., Preston, A. M., Strauss, R., Stamm, R., & Zalman, D. C.

(2016). Risk factors associated with pressure ulcer formation in

critically ill cardiac surgery patients: A systematic review. Journal

6

of Wound Ostomy & Continence Nursing, 43(3), 242–247. https://

doi.org/10.1097/WON.0000000000000224

Shafipour, V., Ramezanpour, E., Gorji, M. A., & Moosazadeh, M.

(2016). Prevalence of postoperative pressure ulcer: A system-

atic review and meta-analysis. Electronic Physician, 8(11),

3170–3176. https://doi.org/10.19082/3170

Shen, W. Q., Chen, H. L., Xu, Y. H., Zhang, Q., & Wu, J. (2015). The

relationship between length of surgery and the incidence of

pressure ulcers in cardiovascular surgical patients: A retrospec-

tive study. Advances in Skin & Wound Care, 28(10), 444–450.

https://doi.org/10.1097/01.ASW.0000466365.90534.b0

Taylor, R. A., Moore, C. L., Cheung, K. H., & Brandt, C. (2018).

Predicting urinary tract infections in the emergency department

with machine learning. PLOS ONE, 13(3), Article e0194085.

https://doi.org/10.1371/journal.pone.0194085

Thomson, J. S., & Brooks, R. G. (1999). The economics of

preventing and treating pressure ulcers: A pilot study. Journal

of Wound Care, 8(6), 312–316. https://doi.org/10.12968/jowc.

1999.8.6.25879

Thorn, C. C., Smith, M., Aziz, O., & Holme, T. C. (2013). The

Waterlow score for risk assessment in surgical patients. Annals

of the Royal College of Surgeons of England, 95(1), 52–56.

https://doi.org/10.1308/003588413X13511609954770

Walsh, C. G., Ribeiro, J. D., & Franklin, J. C. (2018). Predicting suicide

attempts in adolescents with longitudinal clinical data and ma-

chine learning. The Journal of Child Psychology and Psychiatry,

59(12), 1261–1270. https://doi.org/10.1111/jcpp.12916

Walsh, J. (1993). Postop effects of OR positioning. RN, 56(2), 50–57.

Wicke, C., Halliday, B., Allen, D., Roche, N. S., Scheuenstuhl, H.,

Spencer, M. M., Roberts, A. B., & Hunt, T. K. (2000). Effects of ste-

roids and retinoids on wound healing. Archives of Surgery,

135(11), 1265–1270. https://doi.org/10.1001/archsurg.135.11.1265

Xiong, C., Gao, X., Ma, Q., Yang, Y., Wang, Z., Yu, W., & Yu, L.

(2019). Risk factors for intraoperative pressure injuries in pa-

tients undergoing digestive surgery: A retrospective study.

Journal of Clinical Nursing, 28(7–8), 1148–1155. https://doi.

org/10.1111/jocn.14712

Yang, J. J., Wang, X. L., Shi, B. W., & Huang, F. (2013). The angio-

genic peptide vascular endothelial growth factor-basic fibro-

blast growth factor signaling is up-regulated in a rat pressure

ulcer model. The Anatomical Record, 296(8), 1161–1168.

Ye, C., Fu, T., Hao, S., Zhang, Y., Wang, O., Jin, B., Xia, M., Liu, M.,

Zhou, X., Wu, Q., Guo, Y., Zhu, C., Li, Y. M., Culver, D. S., Alfreds,

S. T., Stearns, F., Sylvester, K. G., Widen, E., McElhinney, D., &

Ling, X. (2018). Prediction of incident hypertension within

the next year: Prospective study using statewide elec-

tronic health records and machine learning. Journal of

Medical Internet Research, 20(1), Article e22. https://doi.

org/10.2196/jmir.9268