| Abstract|| |
INTRODUCTION: Massive blood loss and transfusions may be associated with mortality. However, blood donations have been disrupted globally, particularly in the current pandemic era. At present, machine learning (ML) is one of the modern approaches used to predict clinical outcomes. The main objective of this study was to forecast massive packed red cell (PRC) transfusions using ML. In addition, a secondary objective focused on developing an ML-based web application to assist neurosurgeons with high-risk patients and to allocate limited resources effectively during the pandemic.
METHODS: This was a historical cohort study involving individuals who had undergone neurosurgical operations. Using a 70:30 splitting method, the total dataset was randomly separated into a training dataset and a testing dataset. The supervised ML models of various algorithms were trained, whereas the testing dataset was used to validate the ML models.
RESULTS: Among 3006 neurosurgical patients, massive PRC transfusion was observed in 4.26%. The preoperative factors significantly associated with massive transfusions were used to train the ML models. For the testing process, naive Bayes (NB) algorithms had the highest performance for massive transfusion prediction with area under the curve of 0.75 (95% confidence interval: 0.72–0.80). Therefore, the NB model was deployed as a web application, which is a simple-to-use tool via https://psuneurosx.shinyapps.io/massive_transfusion/.
CONCLUSIONS: Massive PRC transfusions in neurosurgical operations occurred in 4.26%. NB algorithms showed acceptable predictability for massive PRC transfusions via the simple web application, though it may be a challenge to implement in general practice as a computerized clinical decision support system.
Keywords: Machine learning, massive transfusions, neurosurgical operations, prediction, web application
|How to cite this URL:|
Taweesomboonyat C, Kaewborisutsakul A, Sungkaro K. Prediction of massive transfusions in neurosurgical operations using machine learning. Asian J Transfus Sci [Epub ahead of print] [cited 2022 Dec 4]. Available from: https://www.ajts.org/preprintarticle.asp?id=356885
| Introduction|| |
Massive blood loss can occur in various situations during neurosurgical operations, such as for traumatic brain injury, neuro-oncology surgery, and ruptured aneurysm clipping.,, Prior studies performed preoperative embolization in central nervous system tumors to reduce intraoperative blood loss, while Gitto et al. reported massive bleeding from a ruptured pseudoaneurysm of the middle cerebral artery. Moreover, previous studies have reported that intraoperative massive blood loss is associated with mortality.,
Blood donations have been disrupted and decreased globally in the pandemic era., From a survey of European countries, Chandler et al. reported that around half of blood donations diminished during the coronavirus (COVID-19) outbreak. Therefore, patients who have risk factors associated with massive blood loss leading to massive blood transfusions should be considered for the allocation of limited resources during the pandemic. Further, blood banks should be prepared to augment blood product supplies. From the literature review, prior studies proposed a predictive model for massive bleeding using several methods. Akaraborworn et al. proposed a massive transfusion scoring system from various clinical factors to predict the events in trauma patients. They found that the tool had an area under the receiver operating characteristics (ROC) (AUC) of 0.83 (95% confidence interval [CI]: 0.78–0.91). Chico-Fernández et al. compared the predictability of massive transfusions using the three-score system as follows: Assessment Blood Consumption (ABC), Emergency Transfusion Score (ETS), and Trauma-associated Severe Hemorrhage (TASH). Hence, AUCs of ABC, ETS, and TASH were 0.77 (0.72–0.83), 0.78 (0.72–0.84), and 0.89 (0.86–0.93), respectively.
Currently, disruptive technologies have been applied in daily life such as consumer segmentation and suggestion systems in various businesses, as well as internet of things by engineers and in agriculture., The background of these technologies is closely related to one type of artificial intelligence that became machine learning (ML). ML is one of the modern approaches used to predict clinical outcomes that has been applied in several fields of medicine via various algorithms. ML algorithms are programs that can learn from histological data and improve the predictability. Supervised ML and the common algorithms of supervised ML have been used to forecast clinical outcomes as follows: naive Bayes (NB) support vector machine (SVM), k-nearest neighbors (KNN), decision tree (DT), random forest (RF), and artificial neural network (ANN)., These algorithms have been utilized in neurosurgical operations. Tunthanathip et al. used NB, SVM, KNN, DT, RF, and ANN for predicting surgical site infections in cranial operations and found that NB had the highest predictability with AUC of 0.76, while RF was an algorithm that had an acceptable predictive performance of 2-year survival in patients with glioblastoma (AUC 0.81). In addition, the supervised ML has been studied for transfusion prediction from literature reviews. Feng et al. used ML to predict blood transfusions in patients who had undergone surgery. They reported that the light gradient boosting machine algorithm had the best predictability with AUC of 0.90, while Liu et al. used ML for forecasting blood transfusions in patients who underwent mitral valve surgery. They reported that the CatBoost algorithm had the highest AUC of 0.88. However, there is a lack of ML-based studies for predicting massive transfusions from the literature review. Therefore, the present study aimed to predict massive packed red cell (PRC) transfusions using various ML algorithms. In addition, the secondary objective was to develop and deploy a web application from the highest predictability ML model for use in general practice.
| Methods|| |
Study design and study population
The historical cohort study was carried out by searching for patients who had been admitted and undergone neurosurgical operations at a tertiary hospital between January 2014 and January 2019. Patients who did not have data for cross-matching and transfusion or patients who were dead on arrival were excluded. Hence, medical records for 3006 patients were reviewed, and clinical characteristics were collected as follows: age, sex, comorbidity, neurosurgical disease, American Society of Anesthesiologists (ASA) classification, operation, estimated blood loss, preoperative hematological laboratories, body mass index, and number of transfused PRC units.
The primary endpoint of the present study was the massive PRC transfusion event for each patient as the binary classifier. In detail, the operational definition of massive PRC transfusion was a patient who received more than 4 units of PRC within 1 h or more than 10 units of PRC within 24 h.,
We calculated the sample size using ROC with the AUC formula. Based on an AUC of 0.89 from the study of Chico-Fernández et al., a minimum of 232 patients from the testing data would be needed to evaluate the predictability of the ML models with a given marginal error of 0.05.
The present study was approved by the Human Research Ethics Committee (REC 65-052-10-1). Informed consent was not required from patients due to the nature of the retrospective study design. However, patients' identification numbers were encoded before analysis.
The baseline characteristics were analyzed using proportion and percentage, whereas the continuous variables were described using mean and standard deviation (SD). Hence, the Chi-square test was performed to evaluate the difference in the proportions between the massive transfusion group and nonmassive transfusion group, while the independent t-test was used for comparing means between the two groups. For effect size estimation, binary logistic regression was performed to identify the factors associated with massive transfusions. <0.05 was considered statistically significant. Afterward, the significant variables were included to train by ML in the next process.
The common algorithms of supervised ML were achieved for forecasting massive PRC transfusions as follows: NB, SVM, KNN, DT, and ANN. NB is a supervised algorithm based on the Bayes theorem that works out the probability of events, while SVM is a linear model for classification of a problem by a line or hyperplane which separates data into biclassifiers or multiclassifiers.,, DT is a tree-like model of decisions and their possible consequences, including the event outcomes. RF is an ensemble learning method that operates by constructing a multitude of DTs at the training period and the output of the RF is the event outcome selected by most DTs.,, ANN is a neuron-like model inspired by the central nervous system. This algorithm is a computational model comprising several interconnected groups of nodes that receive inputs and deliver outputs based on their predefined activation functions.,,
From the prior step of statistical analysis, various significant clinical characteristics were inputs into the supervised algorithms for predicting the binary outputs (transfusion versus nontransfusion) via the train-test processes. In detail, the training dataset (n = 2,104) and testing dataset (n = 902) were created using a 70:30 splitting procedure. Training of the ML model was performed with 5-fold cross-validation, and the training model for each algorithm was developed. Therefore, the predictability of these ML models was evaluated with the testing dataset. The performance of each algorithm was calculated for sensitivity, specificity, positive predictive value, negative predictive value (NPV), and accuracy from the confusion matrix. Further, the ROC with AUC was estimated for ML classification, with an AUC of ≥0.7 indicating acceptable predictability.,
ML was trained using the python Collaboratory program (Python software foundation) with the “scikit-learn” package. For the secondary outcome, the best training model was developed and deployed as a web application using R version 3.6.2 with the “shiny” package.
| Results|| |
Clinical and radiological characteristics
Demographic data are shown in [Table 1]. The mean age of the cohort was 46.42 years ([SD] 20.67); the mean age of the massive transfusion group was significantly higher than the nonmassive transfusion group (P < 0.009). The major neurosurgical conditions were brain tumor, traumatic brain injury, and cerebral aneurysm at 45.7%, 15.4, and 13.6, respectively. Craniotomy (36.3%), decompressive craniectomy (12.1%), and burr hole (8.7%) were common neurosurgical operations. Moreover, 47.5% of all cases were emergency operations. As a result, massive PRC transfusion occurred in 4.26% of the present cohort.
[Table 2] presents preoperative hematologic laboratories. For comparing the mean between two groups, patients with transfusions had more anemia, lower platelet counts, and more significant coagulopathy than the other group. Moreover, the mean intraoperative blood loss in the massive transfusion group was significantly higher than in the nonmassive transfusion group (P < 0.001).
|Table 2: Comparison of preoperative hematologic laboratories by massive transfusion (n=3006)|
Click here to view
Factors associated with massive transfusions
Increased age caused a significantly higher risk for massive PRC transfusion (odds ratio [OR]: 1.012, 95% CI: 1.002–1.021). Operations for patients suffering from traumatic brain injuries had a higher risk for massive PRC transfusion than surgery for brain tumors (OR: 1.9, 95% CI: 1.24–2.87). Moreover, decompressive craniectomy had a significant risk for massive PRC transfusion when compared with craniotomy procedures (OR: 1.6, 95% CI: 1.09–2.41). Other clinical characteristics that significantly increased the risk of intraoperative massive PRC transfusion included chronic warfarin usage and increased ASA classification. For preoperative laboratories, the risk factors associated with blood transfusion were prolonged partial prothrombin time ratio and international normalized ratio, whereas a low level of hemoglobin, hematocrit, and platelet count were protective factors for massive transfusion, as shown in [Figure 1].
|Figure 1: The odds ratio of variables associated with massive transfusion. ASA = American Society of Anesthesiologists, DC = Decompressive craniectomy, TBI = Traumatic brain injury, INR = International normalized ratio|
Click here to view
After the random splitting procedure, the training dataset was used to build the predictive model from the preoperative factors significantly associated with massive transfusions in the former step. Various algorithms of supervised ML classification were trained, and the ML model for each algorithm was turned and optimized for the best parameters using the caret package with 5-fold cross-validation. As a result, NB had the highest AUC with an acceptable performance of predictability and AUC of 0.75 (95% CI: 0.72–0.80), as shown in [Figure 2]. The NB model had high specificity, NPV, and accuracy, as shown in [Table 3]. Therefore, the NB model was used to develop and deploy by web service as a simple-to-use web application via https://psuneurosx.shinyapps.io/massive_transfusion/or QR code scanning, as shown in [Figure 3].
|Figure 2: Receiver operating characteristics curves of each algorithm. NB = Naïve Bayes, SVM = Support vector machine, knn = k-Nearest neighbors, DT = Decision tree, RF = Random forest, ANN = Artificial neural network|
Click here to view
|Figure 3: Screenshot of a web application using the naïve Bayes algorithm|
Click here to view
| Discussion|| |
Massive blood bleeding and transfusions are unexpected intraoperative events that can lead to mortality and disability., We found that the incidence of massive transfusion in neurosurgical operations was 4.26%. Because of the lack of articles mentioning this result in neurosurgical patients, prior studies reported incidence of massive blood transfusion in trauma patients ranging from 3% to 8% in their cohorts,,, which was similar to the incidence compared with our result.
Correspondingly, NB had the highest AUC with an acceptable performance of predictability in the present study that should be further developed and deployed as a real-world predictive tool for physicians and neurosurgeons to use in general practice. From prior studies, several ML algorithms have been studied to predict clinical outcomes. Chang et al. used SVM to predict transfusions in orthopedic operations with an AUC of 0.703–0.707 for intraoperative transfusion, while the NB algorithm was highlighted and had the highest AUC in the present study. This is potentially explained by the NB classifier not requiring as many data points to be trained and being able to deal with high-dimensional features. Moreover, this algorithm is not sensitive to irrelevant features.,, Tunthanathip et al. used various ML algorithms to predict surgical site infection in neurosurgical operations. They found that the NB was highlighted with an AUC of 0.76. To the best of the authors' knowledge, little evidence exists concerning the prediction of massive transfusion using ML. The present study was the first paper to reveal the predictability of ML for forecasting massive transfusions.
The COVID-19 pandemic has highly disrupted blood donations in several countries around the world,,, thus requiring the optimization of limited-resource blood products., We deployed the NB-based web application as a simple tool that could be integrated in general practice. The NB has high specificity as well as few false positive results; the application may be used as a confirmatory tool to enhance physicians' awareness in real-world situations. The web application is one of the implications by ML for decision-making in medical practice. Vodrahalli et al. performed telehealth using ML-based image assessment in dermatology, while Wijesinghe et al. used a deep learning-based telemedicine system for the early diagnosis of diabetic retinopathy and foot ulcers in diabetic patients. Therefore, modern health technologies have been driven by artificial intelligence, ML, and deep learning technologies. In addition, ML has been commonly used in several fields of medicine as follows: triage, diagnosis, treatment, and prognostication.
The limitations in the study should be considered in terms of the multicollinearity that may be recognized. However, we aimed to use all significant preoperative features for training the ML model because high dimensions supported the learning processes., External validation should be performed to confirm the predictability of ML in the future., In addition, various clinical prediction tools have been proposed for predicting massive transfusions such as ABC, ETS, and TASH. Comparison predictability among these various prediction tools should be performed for preparation and resource allocation in the pandemic era. In addition, multicenter trials should be performed in the future due to the increasing prevalence of massive transfusions and to confirm the tool's predictability. Because the web application in the present study has the ML model in the web server, it has become a simple and user-friendly tool to distribute for external validation in other hospitals. In addition, the ML web application is a challenge to become the computerized clinical decision support system (CDSS) for physicians to balance blood preparation and utilization for patients facing high-risk massive blood transfusions., Similarly, Souza et al. conducted a systematic review of 58 previous studies and reported that computerized CDSSs effectively improved the process of care for multimodality, such as screening and treatment, thus enhancing patient outcomes, safety, cost of care, and patient/provider satisfaction.
| Conclusions|| |
Massive PRC transfusions occur in 4.26% of neurosurgical operations. NB algorithms have acceptable predictability for massive PRC transfusions via the simple tool. ML has played an important role in modern health technologies as the CDSS to support physicians with decision-making as well as balancing blood preparation and utilization.
Ethics approval and consent to participate
The study was approved by the institutional committee on research ethics (REC. 65-052-10-1).
Financial support and sponsorship
Conflicts of interest
There are no conflicts of interest.
| References|| |
Kobayashi K, Ozkan E, Tam A, Ensor J, Wallace MJ, Gupta S. Preoperative embolization of spinal tumors: Variables affecting intraoperative blood loss after embolization. Acta Radiol 2012;53:935-42.
Ishihara H, Ishihara S, Niimi J, Neki H, Kakehi Y, Uemiya N, et al
. The safety and efficacy of preoperative embolization of meningioma with N-butyl cyanoacrylate. Interv Neuroradiol 2015;21:624-30.
Gitto L, Richardson TE, Serinelli S, Diana F, Peschillo S, Domenicucci M. Massive intracranial bleeding due to the rupture of a rare spontaneous pseudoaneurysm of the middle cerebral artery in a pediatric patient: case report with clinical, radiological, and pathologic findings. Forensic Sci Med Pathol 2019;15:474-80.
Barbosa RR, Rowell SE, Sambasivan CN, Diggs BS, Spinella PC, Schreiber MA, et al
. A predictive model for mortality in massively transfused trauma patients. J Trauma 2011;71:S370-4.
Akaraborworn O, Chaiwat O, Chatmongkolchart S, Kitsiripant C, Chittawatanarat K, Morakul S, et al
. Prediction of massive transfusion in trauma patients in the surgical intensive care units (THAI-SICU study). Chin J Traumatol 2019;22:219-22.
Stanworth SJ, New HV, Apelseth TO, Brunskill S, Cardigan R, Doree C, et al
. Effects of the COVID-19 pandemic on supply and use of blood for transfusion. Lancet Haematol 2020;7:e756-64.
Loua A, Kasilo OM, Nikiema JB, Sougou AS, Kniazkov S, Annan EA. Impact of the COVID-19 pandemic on blood supply and demand in the WHO African Region. Vox Sang 2021;116:774-84.
Chandler T, Neumann-Böhme S, Sabat I, Barros PP, Brouwer W, van Exel J, et al
. Blood donation in times of crisis: Early insight into the impact of COVID-19 on blood donors and their motivation to donate across European countries. Vox Sang 2021;116:1031-41.
Akaraborworn O, Siribumrungwong B, Sangthong B, Thongkhao K. Massive blood transfusion for trauma score to predict massive blood transfusion in trauma. Crit Care Res Pract 2021;2021:3165390.
Chico-Fernández M, García-Fuentes C, Alonso-Fernández MA, Toral-Vázquez D, Bermejo-Aznarez S, Alted-López E. Massive transfusion predictive scores in trauma. Experience of a transfusion registry. Med Intensiva 2011;35:546-51.
Howard J. Artificial intelligence: Implications for the future of work. Am J Ind Med 2019;62:917-26.
Zhang MZ, Wang LM, Xiong SM. Using machine learning methods to provision virtual sensors in sensor-cloud. Sensors (Basel) 2020;20:E1836.
Tunthanathip T, Sae-Heng S, Oearsakul T, Sakarunchai I, Kaewborisutsakul A, Taweesomboonyat C. Machine learning applications for the prediction of surgical site infection in neurological operations. Neurosurg Focus 2019;47:E7.
Tunthanathip T, Oearsakul T. Machine learning approaches for prognostication of newly diagnosed glioblastoma. Int J Nutr Pharmacol Neurol Dis 2021;11:57-63. [Full text]
Feng Y, Xu Z, Sun X, Wang D, Yu Y. Machine learning for predicting preoperative red blood cell demand. Transfus Med 2021;31:262-70.
Liu S, Zhou R, Xia XQ, Ren H, Wang LY, Sang RR, et al
. Machine learning models to predict red blood cell transfusion in patients undergoing mitral valve surgery. Ann Transl Med 2021;9:530.
Hajian-Tilaki K. Sample size estimation in diagnostic test studies of biomedical informatics. J Biomed Inform 2014;48:193-204.
Tunthanathip T. Translational Medicine in Neurosurgery. Bangkok: Sahamit Pattana Printing; 2022.
Patil V, Shetmahajan M. Massive transfusion and massive transfusion protocol. Indian J Anaesth 2014;58:590-5.
] [Full text]
Chidester SJ, Williams N, Wang W, Groner JI. A pediatric massive transfusion protocol. J Trauma Acute Care Surg 2012;73:1273-7.
Zewdie K, Genetu A, Mekonnen Y, Worku T, Sahlu A, Gulilalt D. Efficiency of blood utilization in elective surgical patients. BMC Health Serv Res 2019;19:804.
Taweesomboonyat C, Kaewborisutsakul A, Tunthanathip T, Saeheng S, Oearsakul T. Necessity of in-hospital neurological observation for mild traumatic brain injury patients with negative computed tomography brain scans. JHSMR 2020;28:267-74.
Wade CE, del Junco DJ, Holcomb JB, Trauma Outcomes Group, Holcomb JB, Wade CE, et al
. Variations between level I trauma centers in 24-hour mortality in severely injured patients requiring a massive transfusion. J Trauma 2011;71:S389-93.
Holcomb JB, Zarzabal LA, Michalek JE, Kozar RA, Spinella PC, Perkins JG, et al
. Increased platelet: RBC ratios are associated with improved survival after massive transfusion. J Trauma 2011;71:S318-28.
Nunez TC, Voskresensky IV, Dossett LA, Shinall R, Dutton WD, Cotton BA. Early prediction of massive transfusion in trauma: Simple as ABC (assessment of blood consumption)? J Trauma 2009;66:346-52.
Chang CM, Hung JH, Hu YH, Lee PJ, Shen CH. Prediction of preoperative blood preparation for orthopedic surgery patients: A supervised learning approach. Appl Sci 2018;8:1559.
Ambika P. Chapter Thirteen – Machine learning and deep learning algorithms on the Industrial Internet of Things (IIoT). In: Raj P, Evangeline P. Advances in Computers. Amsterdam: Elsevier; 2020. p. 321-38.
Vodrahalli K, Daneshjou R, Novoa RA, Chiou A, Ko JM, Zou J. TrueImage: A machine learning algorithm to improve the quality of telehealth photos. Pac Symp Biocomput 2021;26:220-31.
Park MS, Jo H, Lee H, Jung SY, Hwang HJ. Machine learning-based COVID-19 patients triage algorithm using patient-generated health data from nationwide multicenter database. Infect Dis Ther 2022;11:787-805.
Gorji F, Shafiekhani S, Namdar P, Abdollahzade S, Rafiei S. Machine learning-based COVID-19 diagnosis by demographic characteristics and clinical data. Adv Respir Med 2022. [Epub ahead of print]. doi: 10.5603/ARM.a2022.0021. PMID: 35102543.
Tunthanathip T, Duangsuwan J, Wattanakitrungroj N, Tongman S, Phuenpathom N. Comparison of intracranial injury predictability between machine learning algorithms and the nomogram in pediatric traumatic brain injury. Neurosurg Focus 2021;51:E7.
Egert M, Steward JE, Sundaram CP. Machine learning and artificial intelligence in surgical fields. Indian J Surg Oncol 2020;11:573-7.
Goldrick S, Sandner V, Cheeks M, Turner R, Farid SS, McCreath G, et al
. Multivariate data analysis methodology to solve data challenges related to scale-up model validation and missing data on a micro-bioreactor system. Biotechnol J 2020;15:e1800684.
Tunthanathip T, Ratanalert S, Sae-Heng S, Oearsakul T, Sakarunchai I, Kaewborisutsakul A, et al
. Prognostic factors and clinical nomogram predicting survival in high-grade glioma. J Cancer Res Ther 2021;17:1052-8.
Kaewborisutsakul A, Tunthanathip T. Development and internal validation of a nomogram for predicting outcomes in children with traumatic subdural hematoma. Acute Crit Care 2022. [Epub ahead of print]. doi: 10.4266/acc.2021.01795. PMID: 35791657.
Moxey A, Robertson J, Newby D, Hains I, Williamson M, Pearson SA. Computerized clinical decision support for prescribing: Provision does not guarantee uptake. J Am Med Inform Assoc 2010;17:25-33.
Souza NM, Sebaldt RJ, Mackay JA, Prorok JC, Weise-Kelly L, Navarro T, et al
. Computerized clinical decision support systems for primary preventive care: A decision-maker-researcher partnership systematic review of effects on process of care and patient outcomes. Implement Sci 2011;6:87.
Department of Surgery, Division of Neurosurgery, Faculty of Medicine, Prince of Songkla University, Songkhla 90110
Source of Support: None, Conflict of Interest: None
[Figure 1], [Figure 2], [Figure 3]
[Table 1], [Table 2], [Table 3]