Research Background
Trauma is a key public health issue and the primary basis of death and disability across the globe. Over 180,000 persons die each year or approximately 1 death each 3 minutes from traumatic wounds (Centers for Disease Control and Prevention, 2014) and it is by now the principal cause of death within the below age set (Bouamra and Lesko, 2014). Several demographic and therapeutic aspects have been associated with death after experiencing traumatic wounds like age and post-injury wellbeing circumstances. Nevertheless, these aspects are extremely interconnected and recognizing pertinent elements or at-threat-of-death populace groups is not easy. The expansion of patient result Predicting techniques has been a significant part of trauma exploration right from the early 1970s (Palocsay et al., 1996). This analysis offers the growth of graphical model that can employ data concerning trauma patients’ circumstances so as to precisely foretell their medical result.
This study begins by backdrop details regarding the research. Subsequently, it defines the trauma forecast literature review that comprises prior applied variables and findings. After that, it contrasts accessible prediction approaches that comprise statics and simulated intelligent techniques. Afterward, it analyses the universal hypothesis of Bayesian network and the basic idea of inexact likelihood, to then demonstrate the devise and methodology of the study. Lastly, we assess some additional factors like the moral approval and the devise for future investigation.
Problem Statement
Exploring pre-existing medical circumstances in evaluating injury results is an ongoing apprehension within the trauma health area. As a result, it is significant to construct a model that can: encompass the interacting elements that could affect different results; represent intricate interaction effectively within a network configuration; and deduce provisional dependencies available between the aspect and the results.
Research Aim & Objectives
The projected study will aspire to carry on rigorous investigation on the associated dimensions of foretelling Trauma results that influence and enhance the wellbeing care factors of medical institution. As a result, the subsequent main goals shall set the procedures for the investigator:
- To explain numerous factors of demographic and medicinal aspects that could influence trauma findings.
- To supply modernized, precise predictions on changing patient state using a controllable amount of variables.
- To demonstrate the relationships between the aspects and to characterize the associations in the presence of ambiguity of insufficient information.
- To build up an appropriate Bayesian Network archetype for caregivers and health care bodies.
- To supply recommendations for improving the projected prediction technique.
Research Benefits/Significance
There are numerous significant parts where this research forms an innovative contribution to study within this area such as future plans of direct care and relocating sick persons to help the growth of trauma services. Moreover, the model can be employed to help in other diverse medical choice-making roles. The outcomes of the projected research will emphasize the mainly influencing aspects that amplify the danger of death. This will propose offering fast initial cure to persons who might have additional risk results from the damage.
Our findings must be of notice not just to the health care society, but to the simulated intelligent group also.
Proposed Chapters
The research will be consisted of the subsequent five sections:
Chapter 1 – Introduction
Chapter 2 – Literature Review
Chapter 3 – Study Methodology
Chapter 4 – Data Review & Discussion
Chapter 5 – Conclusion & Suggestions
2. Literature Review
Trauma Prediction
Variables established to be affecting trauma result have been investigated in numerous studies. Some assessed age the capacity to foretell result from trauma (death forecast is possibly the mainly elementary utilize of injury-cruelty scoring, trailed by other result actions); contrast of the therapeutic techniques; a pre- and inter-infirmary triage device; an instrument for quality-enhancement and avoidance programme (Chawda et al, 2004). As a result, when planning a result appraisal on a group of sick persons one should generate conclusions founded on all of these different contributing aspects.
Nevertheless, the accessible clustering principle/algorithms do suit widely into four basic classes: The Trauma literature deemed factors (criteria) founded on one of these aspects or a combination of a number of them.
Critical symptoms comprising systolic blood pressure, pulse rate, rate of respiratory, body heat, and level of awareness are employed to dispense an early caution score to evaluate sickness severity. During trauma care, other devices for physiology scoring like the Glasgow Coma Score, and the abridged wound score have been built to evaluate trauma harshness before comprehensive diagnoses can be carried out for trauma sick persons.
1) Age, gender, anatomy and mechanism of injury
Numerous prior retrospective group researches have discovered age is a widespread aspect that could foretell trauma result (Imen et al. 2015) (Dent et al., 1995) ;( Walcott et al., 2014) above 65 years of age (Joseph et al., 2013) (Liebman et al., 2010). Gender have projected in other earlier investigations (Liebman et al., 2010) (Imen et al. 2015), and source of wound (Imen et al. 2015).
Classified sick persons founded on anatomy of damage are imperative. For instance, Predicting result of traumatic head wound. TBI is varied conditions of basis, pathology, harshness, and prognosis that fronts diagnostic difficulties. Nevertheless, TBI described by a Glasgow unconsciousness scale (GCS) mark of 8 or fewer in the initial post-traumatic period (Imenet al., 2015). Furthermore, describing the anatomy of wound assists with means of patient’s analysis. For instance, the connection linking the medical effects of severe optical and orbital trauma with the results on automated tomography (CT) (Chaudharyet al., 2015). Prior studies have concentrated on predicting the findings of one composition of injury like chest wall damage (Pressleyet al, 2012), severe whiplash suffering (Gehrt et. al. 2015) and foretelling visual result (Ainbinder et al. 2003)
2) Vital Signs
Conventional critical signs are viewed like a significant section of trauma evaluation; in spite of their poor prognostic worth in this respect (Bruijnset al., 2013). Crucial indications have been extensively applied in trauma forecast models. For instance, Imen et al. 2015 have employed respiratory rate, heart rate, systolic and diastolic pressure of the blood as a forecast aspects and Palocsay et al. (1996) employed Systolic pressure of blood (S) and rate of Respiratory R). Pupil reaction by (Dent et al., 1995);(Walcott et al., 2014).
3) Post-injury conditions
Numerous injuries differ between patients as measured by various aspects like age, pre-existing illness and most likely hereditary predisposition (Chawda et al., 2004). Having a post damage state would influence the injury result like pregnancy. The treatment of a pregnant trauma sick persons warrants deliberation of numerous issues particular to pregnancy like contact to emission and other likely dermatogens, the necessity to evaluate circumstances that are exclusive to pregnancy and are associated with trauma such as preterm labour and placental abruption ) (Jain et al., 2015).
4) Scoring systems
The scoring mechanism can help during the earlier expansion of management strategies that might precisely foretell in the first 24 hours of care the patients that are mainly probable to have poor results and need thorough attention to permit early involvement to enhance effects.
In actual fact, quantifying damage needs 3 things: foremost a lexicon that separates the constant background of human wound into a set of distinct damages; next, an assess of severity for every damage; and third, a model that sums up the shared severity of the entire injuries that an independent patient has suffered as a sole numeric worth (Osler et al., 2008). More than 30 years ago, the entire 3 roles were performed concurrently when Baker et al implemented the abridged injury scale (AIS) (Committee on Medical Aspects of Automative Safety, 1971) inventory of damages and its attendant professional-allocated severities, and described the injury sternness score (ISS) as the total of the squares of the cruelties of the lone worst damage in every of the 3 mainly injured body parts (Baker et al., 1974).
ISS score rapidly became the average gauge of trauma and was almost immediately integrated as the quantity of trauma in extra inclusive models of death after trauma. For instance, (Imen et al. 2015) (Dent et al., 1995) ;( Walcott et al., 2014). Nevertheless, the ISS has significant restrictions. Since it relies on the expert-allocated severities of the AIS listing, the ISS accedes to the lack of accuracy of the trauma AIS and was soon integrated as the quantity of trauma (Osler et al., 2008). Furthermore, the Injury Severity Score has confines linked to disregarding numerous damages within the same body part (Bouamra and Lesko, 2014).
The Abridged Injury Scale generated by: Alliance for the of expansion of Automotive Medicine (AAAM) established in 1971 to assist vehicle collision explorers expanded in 1990 to be extra pertinent to medical review and study. It estimated within 24 hours following admittance (Imen et al. 2015).
A great deal of other research deemed other kinds of variables like psychological aspects as forecast elements that influence patient’s health result. For instance, Gehrt et. al. (2015) established that severe whiplash wound Patients having negative sickness perceptions were further possible to suffer neck ache and affected functioning capacity at 12 months contrasted to patients having optimistic sickness perceptions. Other aspects after sick person’s admission include forms of surgical procedure, harshness of surgery (Liebman et al., 2010).
Trauma Outcomes
Little is recognized regarding the last functional result of patients having numerous wounds, how speedy they recuperate, and the quantity that have outstanding disabilities (Chawdaet al.,2004). There are various measures that have been used for determining patients effects like intensive care unit (ICU) admittance, intubation and automatic aeration, duration of stay at hospital (Crystal M. et al, 2012) (Malone et al. 2001) (Krishnan et al. 2014). Result group disparities might as well turn out to be larger with further days of study. For instance, within the Hypovolemia set, when examining 31 days rather than 16 days of sequential statistics, the degree of implication of the dissimilarity between mean Principle element standards for sodium augmented (Crump et al., 2011). Other principle functioning capacity was evaluated using a self-report calendar in the previous month of the follow-up time. The sick persons were requested to list days having sick leave and decreased functioning hours because of the disaster (Gehrt et. al. 2015)
Progressively, researches have assessed the psychological influence on persons who survive a disturbing physical wound , negative association between Glasgow Coma Scale and despair, with the claim that the psychological result lasts above the physical one (Rainey et al, 2014) Survival and bereavement (Imen et al., 2015).
The necessity for a prediction technique
It would be an incredible benefit for clinicians to be capable to foretell with an elevated level of precision the result of a trauma patient. Nevertheless, several of the earlier work had some benefits like including big losses of information and the trial size is comparatively little together with one hospital (Imen et al. 2015) that makes the execution of the outcome have many challenges.
Prognostic forms that have several clinical gains when foretelling the result of independent trauma patients are pertinent. The variables that are foretold to have an unconstructive consequence on outcome within a model can as well direct clinicians in their revival effort of suffering victims (Bouamra and Lesko, 2014)
Prediction Methods
The classic modeling techniques used for trauma forecast models thus far can be widely categorized into: statistical techniques and artificial intellect or information mining based approaches:
Statistical Methods
Earlier work by Tanoueet al., (2013) employed one-way study of variance to assess the connection between radiological changeable and unfortunate neurodevelopment results. The Bartlett experiment was employed to check for variance homogeneity.
A different research by Gehrtet al., (2015) used Chi-squared examination to assess the connection between death and the various aspects (p value of below 0.05 were deemed noteworthy. risk aspects were assessed by multivariate study using numerous logistic stepwise degeneration processes.
The Logistic Regression form is one of the mainly statics approaches that have been employed during trauma forecast (Malone et al., 2001). (Krishnan et al., 2014) (Liebman et al., 2010). It determines the connection between the clear-cut reliant variable and one or extra self-sufficient variables through approximating likelihoods. LR is an utmost-likelihood technique that has been applied in numerous trauma result investigations (Kong et al., 2016).
Logistic regression contains flaws in coping with too many extremely correlated forecasters at a time (Hossain, M., &Muroma, C., 2012). Abdel-Atyet al. (2008) created a noteworthy enhancement over this method by using random forest that determines the variable significance through forming a compilation of categorization trees produced by unsystematic sampling of statistics and random variable assortment. Even though the technique these days is broadly accepted for its steadiness, unbiased outcome and capability to manage big variable space with little test size, it yet can be vulnerable to biasness once any or a set of variables have comparatively bigger amount of classes (Strobl et al.,2007).
As a result of the data restrictions and the disadvantages of standard statistics methods, a Bayesian technique offers the solution from a diverse viewpoint. Bayesian method regards the testing allocation unrelated to the statistical implications since it takes into account events that have not happened yet. Bayesian implication is done by means of Bayes theorem wherein posterior allocation is described by the possibility function that has sample data times the previous distribution of factor of the concern. While Bayes implication follows the official rules of likelihood hypothesis, Bayes estimators are steady, asymptotically effective, and acceptable under mild circumstances.
When formed using statistical methods, the majority of them become dropped as section of modeling procedure. Therefore, it is significant to use techniques that can put up with connected variables and make finest employ of each accessible piece of detail to enhance the forecast success (Hossain, M., & Muroma, C., 2012). Furthermore, using above two diverse result measures provides the chance to explore the function of other forecasters in every measure separately (Gehrt et. al. 2015). Graphical forms are frequently suggested in this situation because they are capable to model interconnected data, and to characterize intricate relationships successfully within a network structure.
Artificial Intelligent Techniques
Neural system founded modeling approaches (such as probabilistic neural system) can hold correlated reliant variables. Nevertheless, they anticipate adequate previous knowledge concerning the trouble domain shown through the interconnection among the forecasters (Hossain, M., &Muroma, C., 2012).
The neural system model we chose for the trauma result prediction trouble is the extensively employed multi-layer, supply forward system with controlled learning. Nonetheless, a major concern in any neural system application is when preparation should be ceased. The effort of Rumelhart and others (Rumelhart et al., 1994; Hintonand Touretzky, 1991; Hinton, 1992) recommends that employing a “monitoring group” in forming this choice frequently enhances analytical performance of the ensuing neural system (Palocsay et al., 1996). Neural system education algorithms aim to reduce the amount of the squared mistakes that the net produces on the preparation set. One of the mainly complex issues encountered in using these algorithms is measuring whether a system qualified on a set of trials will be capable to simplify to novel, never-before-seen situations (Palocsay et al., 1996).
On the contrary to neural systems having “black box” structural design, the configuration of Bayesian networks gives the model builder immediate information regarding the temperament of the difficulty set and the comparative importance of variables to the result of interest. Additionally, through imputing present information into the form, the model builder gets a possibility of result. With a graphical demonstration of the network as well gives the model developer the possible rationale for the result and knowledge regarding extra details needed to confirm or disprove the predicted result (Crumpet al., 2011)
Not any of the logistic degeneration promotes vector machines (SVMs),and artificial neural nets (ANNs) models need solid information regarding the connection between precursor aspects and dependent results, and these techniques are totally data-motivated, which implies adequately big sample statistics are required to study forecast models (Kong et al., 2016).
Bayesian Network advantages
This research addresses the aforesaid shortcomings through suggesting a Bayesian belief network (BBN – also recognized as Bayesian net) based structure to build up Trauma wound forecast model.
There are numerous grounds for employing a Bayesian system in this circumstance. Foremost, the main benefits of Bayesian techniques are that they are capable to represent biased data such as professional view or beliefs and objective information.
During medical analysis, we must consider the details that the therapeutic hypothesis we have might not be absolute, we do not posses entire information concerning the sick, and the trials we run might be imprecise. A significant benefit of Bayesian techniques, different from frequentist techniques through which they are frequently compared, is that they can at all times yield an accurate response, even when no information at all is accessible. Lastly, latest Bayesian literature has concentrated on the possible implication of model ambiguity and the means it can be integrated into quantitative diagnoses (Ferson, 2005).
BBN is extremely efficient in circumstances where implications are not guaranteed logically but, instead, probabilistically (Hossain, M., &Muroma, C., 2012). BN is usually employed in numerous parts wherein inferences are needed, including representation scrutiny, equipment preservation, risk control, and medical analysis (Crumpet al., 2011)
Application of Bayesian networks offers latter analytical distributions for a specified group of variables to be diagnosed, thus offering a computationally efficient technique under circumstances of ambiguity for obtaining probabilities for a particular result or event (Crumpet al., 2011)
Bayesian Network with imprecise probability
Following its design, a Bayesian system can be queried by suitable implication algorithms so as to take out probabilistic data concerning the variables of notice (Corani, G.,Antonucci, A., &Zaffalon, M., 2012). Bayesian nets are exact models within the sense that precise numeric standards should be given as probabilities required for the model consideration. This prerequisite is at times too thin (Corani, G.,Antonucci, A., & Zaffalon, M., 2012). In reality, there are circumstances where a solitary probability allocation cannot appropriately explain the ambiguity concerning the condition of a variable (Abell ́an, J., Moral, S., 2005).
However, even when the spotlight is on learning possibilities from statistics, a creedal set might present a further dependable model of the ambiguity, particularly when dealing with small or unfinished data groups. Therefore, classifiers founded on Bayesian systems can be gainfully expanded to turn into creedal classifiers depending on creedal nets. Different from Bayesian classifiers, which always detect a single class as the one maximizing the posterior class likelihood, a creedal classifier operates with sets of allocations and might ultimately be incapable to differentiate a single group as that having uppermost likelihood (Corani, G.,Antonucci, A., &Zaffalon, M., 2012).
Imprecision
Small/unfinished information, expert’s (qualitative) data, undependable/deficient observations.
Creedal Sets
Walley’s behavioral hypothesis of inexact probabilities [75] offers an absolute probabilistic speculation, based disjointed lower provisions that simplifies to ambiguity deFinetti’s classical hypothesis [43].
- Research Methodology
The study will employ a graphical modeling method, particularly a Bayesian network (BN), to form the joint allocation of demographic and therapeutic aspects and trauma result.
What is Bayesian Network
Bayesian network is a technique of the artificial intellect (AI) likelihood and indecision society having multifarious application (e.g., reasoning underneath indecision, making forecasts of extremely uncertain occurrence, etc.). Instead of building a model concentrating on the crisis, BBN forms the system that can subsequently be applied to comprehend an occurrence or create predictions regarding events (Hossain, M., & Muroma, C., 2012). A Bayesian network defines a system of concern through specifying associations of provisional reliance between its elements. These relations are symbolized by a guided acyclic grid, and this, alongside a joint likelihood allocation for the nodes it has, forms a model that can be applied for making implications. The model is comprises of two sections.
The initial, the qualitative section represents the configuration of a BN as shown by a guided acyclic chart (digraph). The digraph’s joints symbolize the pertinent variables (aspects) from the sphere being formed that can be of diverse kinds (e.g. noticeable or latent, definite, arithmetical). A digraph’s curves symbolize probabilistic relations, i.e. they stand for the causal associations between factors. Next, the quantitative section links a node probability table (NPT) to every joint, its likelihood allotment. A parent nodule’s NPT defines the comparative probability of every condition (value); a kid node’s NPT defines the comparative likelihood of every state provisional on each mixture of conditions of its parents. One of the attributes of BNs is their ability of programming the idea of causality. For example, the curve from A to B points out that node A leads to node B. Officially, the connection between two nodules is founded on Bayes’ law (Pearl, J., 1988):
Where:
- P(X | E) is known as the posterior allocation and symbolizes the likelihood of X specified proof E;
- P (E) is known as the prior allocation and symbolizes the likelihood of X before facts E is specified;
- P (E | X) is known as the likelihood purpose and indicates the likelihood of E presuming X is factual. The BN supposes that the model elements themselves are accidental and follow a previous allocation, given depending on model designer/builders’ previous knowledge. The previous distribution will be modernized once investigational data is accessible and turns out to be posterior allocation (Chen et al., 2008).
Building the model
The Bayesian mold will be developed and assessed within three stages having two datasets. The initial stage is for variable assortment that entails 189 crash statistics and its matching 6478 usual traffic condition information. The second stage builds the model by concluding the acyclic guided graph after making it simpler by means of parent dissociating method and then produces the provisional likelihood tables for every element. The final stage appraises the performance of the model through a divide dataset that lays out a plan for executing the model in actual circumstances (Hossain, M., &Muroma, C., 2012).
All information processing and presentation examination will be completed employing STATA program. Information will be arbitrarily divided into three stratified files. Three iterations of mold preparation and corroboration were performed, every time approximating model elements by means of two-thirds of the information and measuring its performance through the other third. Data driven substantiation mold performance will be measured by means of receiver operating feature (ROC) arcs.
Data Set
The graphical form will be tried through a test from persons having traumas. Prospectively, statistics will be removed and reviewed from the TARN catalog that comprises hospital-reported disturbing injury information. The statistics set comprise complete data on the entire key in factors for wounded trauma patients. After realizing that the files were ordered with regard to Location, the entire trauma patients regardless of age were incorporated for review if patients who reach any infirmary in Wales and England alive within the previous four years and suit any of the ensuing criteria: Bereavement from wound at any spot during admittance, stay in sanatorium for over 3 days needs thorough or high reliance care and needs to be moved to other infirmaries for expert care or for an Intensive attention/ elevated reliance unit bed. TARN involving hospitals will be incorporated in the research dataset if they contained a Glasgow Coma Score (GCS) below 15 at provision or any head damage having AIS harshness code 3 and beyond. Only situations having known last result will be chosen.
Ethical Approval
The projected research will integrate secondary techniques of exploration, hence TARN database has no patient detectors, and research result may be susceptible to moral limitations. The investigator will thus strive to involve in practices and processes that stick to the three usual moral considerations: beneficence, autonomy, and fairness.
Timetable & strategy for the future research
Even though Bayesian methods hold much guarantee for foretelling patient findings, they have not been established to be tough to sound; Streib et al., 16 for instance, confirmed that their general performance was considerably decreased with growing levels of noise. For an ICU background, when attempting to build up a model for result forecast that spans numerous clinical elements, every being modernized frequently, that huge amount of potential contributions entering into a Bayesian implication engine is difficult not just from a computational intricacy perspective but as well from a burdening “sound” stance that consecutively compromises the implication engine’s presentation. One means to decrease the affluence of physiologic factors to a lesser amount of more noteworthy values that can subsequently be reviewed over time is to employ primary component analysis (PCA) ( Crump et al., 2011).
Researches of this type are extremely resource challenging and several times information on all the factors are not accessible during the era of modeling. As a result, a modeling technique that can have room for future fresh factors as well as information from new statistics within course of moment without needing reconstruction or recalibrating the entire model is extremely attractive (Hossain, M., & Muroma, C., 2012).The plain Bayesian network built using the outcomes of PCA for the collective set of patients shows the viability of using a further intricate Bayesian net to more precisely predict result for serious care patients (Crump et al., 2011).
The issue with Bayesian Network that if our facts concerning nodes is not whole, how can we evaluate their absolute relations? Moreover, not all the associations are at all times essential. However, within the Bayesian system method, the absolute net of dependencies is just assumed with little validation.
Future effort will consider techniques for better symbolizing time-series information so that rather than employing data taken at separate positions in time, various values that signify a window of era could be applied. For instance, heart rate might have been recorded numerous occasions within a specified hour, whereas a laboratory worth such as sodium may just have been gathered once or twice within a day. Additionally, networks maybe developed that offer modernized result likelihood on an everyday basis as novel fundamental sign data turn out to be accessible in actual time (Crump et al., 2011).
References
Abell ́an, J., Moral, S.(2005). Upper entropy of credal sets. Applications to credal classification. International Journal of Approximate Reasoning 39(2-3), 235-255.
Abdel-Aty, M., Pande, A., Das, A., &Knibbe, W. J. (2008).Assessing safety on Dutch freeways with data from infrastructure-based intelligent transportation systems.Transportation Research Record, 2083(2083), 153–161. http://doi.org/10.3141/2083-18
Ainbinder, D.J., Mazzoli, R.A., Raymond, W.R., Hansen, E.A., Sanford, E.G. (2003)Ophthalmic care of the combat casualty, Ocular trauma scores. Section 2, chapter 6.
Baker, S. P., O’Neill, B., Haddon, W., & Long, W. B. (1974). The injury severity score: a method for describing patients with multiple injuries and evaluating emergency care. The Journal of Trauma, 14(3), 187–196.
Bouamra, O., &Lesko, M. M. (2014). Outcome prediction modelling for trauma patients: a German perspective. Critical Care, 18(5), 616.
Bruijns, S. R., Guly, H. R., Bouamra, O., Lecky, F., & Lee, W. a. (2013). The value of traditional vital signs, shock index, and age-based markers in predicting trauma mortality.The Journal of Trauma and Acute Care Surgery, 74(6), 1432-7.
Centers for Disease Control and Prevention. Injury Prevention & Control: Data & Statistics (WISQARS) 2014; Available at; http://www.cdc.gov/injury/wisqars/ index.html. Accessed May 26, 2016.
Chaudhary, R., Upendran, M., Campion, N., Yeung, A., Blanch, R., Morgan-Warren, P., … Scott, R. (2015). The role of computerised tomography in predicting visual outcome in ocular trauma patients.Eye (Lond), 29(7), 867–871.
Chawda, M. N., Hildebrand, F., Pape, H. C., &Giannoudis, P. V. (2004). Predicting outcome after multiple trauma: Which scoring system?.International Journal of the Care of the Injured, 35, 347-358.
Chen, W., Xiong, Y., Tsui, K.-L., & Wang, S. (2008). A Design-Driven Validation Approach Using Bayesian Prediction Models. Journal of Mechanical Design, 130(2), 021101.
Chesnut, R. M., Marshall, L. F., Klauber, M. R., Blunt, B. A., Baldwin, N., Eisenberg, H. M., Jane, J.A., Marmarou, A. A., Foulkes, M. A. (1993). The role of secondary brain injury in determining outcome from severe head injury.The Journal of Trauma, 34(2), 216-22.
Committee on Medical Aspects of Automative Safety. (1971). Rating the Severity of Tissue Damage – I. The Abbreviated Scale. Jama, 215(Jan 11), 277-280.
Cooper, D. J., Myles, P. S., McDermott, F. T., Murray, L. J., Laidlaw, J., Cooper, G., Tremayne, A. B., Bernard, S. S., Ponsford, J. (2004). Prehospital hypertonic saline resuscitation of patients with hypotension and severe traumatic brain injury: a randomized controlled trial. JAMA : The Journal of the American Medical Association, 291(11), 1350-1357.
Corani, G., Antonucci, A., &Zaffalon, M. (2012). Bayesian Networks with Imprecise Probabilities: Theory and Application to Classification. Intelligent Systems Reference Library, 23, 49-93.
Crump, C., Silvers, C. T., Wilson, B., Schlachta-Fairchild, L., Lingley-Papadopoulos, C. a, & Ashley, J. S. (2011). Using principal component analysis to aid Bayesian Network development for prediction of critical care patient outcomes. The Journal of Trauma, 71(6), 1841-1849.
Dent, D. L., Croce, M. A., Menke, P. G., Young, B. H., Hinson, M. S., Kudsk, K. A., … Fabian, T. C. (1995). Prognostic factors after acute subdural hematoma. The Journal of Trauma, 39(1), 33–36. Retrieved from http://www.ncbi.nlm.nih.gov/pubmed/7636908
Ferson S. (2005). Bayesian methods in risk assessment. Unpublished Report prepared for th Bureau de Recherches Geologiques et Minieres (BRGM), New York.
Gehrt, T. B., Wisbech, Carstensen, T. B., Ornbøl, E., Fink, P. K., Kasch, H., &Frostholm, L. (2015).The Role of Illness Perceptions in Predicting Outcome After Acute Whiplash Trauma: A Multicenter 12-month Follow-up Study. The Clinical journal of pain (Vol. 31).
Hossain, M., &Muromachi, Y. (2012). A Bayesian network based framework for real-time crash prediction on the basic freeway segments of urban expressways. Accident Analysis and Prevention, 45, 373-381.
Imen, R. B., Olfa, C., Kamilia, C., Meriam, B., Hichem, K., Adel, C., Mabrouk, B., Noureddine, R. (2015). Factors predicting early outcome in patients admitted at emergency department with severe head trauma. Journal of Acute Disease, 4(1), 68-72.
Jain, V., Chari, R., Maslovitz, S., Farine, D., Maternal Fetal Medicine Committee, Bujold, E., … Sanderson, F. (2015). Guidelines for the Management of a Pregnant Trauma Patient. Journal of Obstetrics and Gynaecology Canada, 37(6), 553-574.
Joseph, D., Shusdock, I., Staff, I., & Butler, K. (2013). 671: Predicting Outcomes in Elderly Critically Ill Trauma Patients. Critical Care Medicine, 41(12). Retrieved from
Kao, L.-J., & Chen, H.-F. (2012). Applying Hierarchical Bayesian Neural Network in Failure Time Prediction. Mathematical Problems in Engineering.
Kong, G., Xu, D. L., Yang, J. B., Yin, X., Wang, T., Jiang, B., & Hu, Y. (2016).Belief rule-based inference for predicting trauma outcome. Knowledge-Based Systems, 95, 35–44.
Kuhn, F., Maisiak, R., Mann, L., Mester, V., Morris, R., & Witherspoon, C. D. (2002). The ocular trauma score (OTS). Ophthalmology Clinics of North America. Krishnan, M., Beck, S., Havelock, W., Eeles, E., Hubbard, R. E., & Johansen, A. (2014). Predicting outcome after hip fracture: Using a frailty index to integrate comprehensive geriatric assessment results. Age and Ageing, 43(1), 122-126.
Liebman, B., Strating, R. P., Van Wieringen, W., Mulder, W., Oomen, J. L. T., & Engel, A. F. (2010). Risk modelling of outcome after general and trauma surgery (the IRIS score).British Journal of Surgery, 97(1), 128–133. http://doi.org/10.1002/bjs.6808
Malone, D. L., Kuhls, D., Napolitano, L. M., McCarter, R., &Scalea, T. (2001). Back to basics: validation of the admission systemic inflammatory response syndrome score in predicting outcome in trauma. The Journal of Trauma, 51(3), 458-463.
Nijboer JMM, van der Naalt J, Duis H-JT. (2010). Patients beyond salvation?Variouscategories of trauma patients with aminimal Glasgow Coma Score.Injury, 41:52-7.
Osler, T. (1993). Injury severity scoring: perspectives in development and future directions. Am. J. Surg., 165, 43S-51S.
Osler, T., Glance, L., Buzas, J. S., Mukamel, D., Wagner, J., & Dick, A. (2008). A trauma mortality prediction model based on the anatomic injury scale. Annals of Surgery, 247(6), 1041-8.
Palocsay, S., Stevens, S., Brookshire, R., Sacco, W., Copes, W., Buckman,R., Smith, J. (1996). Presentation Stream: Modelling Towards Policy Outcomes. European Journal of Operational Research, 93, 369-386.
Pearl, J. (1988). Probabilistic Reasoning in Intelligent Systems. Morgan Kauffmann San Mateo. http://doi.org/10.2307/2026705
Perel, P., Roberts, I., Bouamra, O., Woodford, M., Mooney, J., & Lecky, F. (2009). Intracranial bleeding in patients with traumatic brain injury: a prognostic study. BMC Emergency Medicine, 9, 15.
Pressley, C. M., Fry, W. R., Philp, A. S., Berry, S. D., & Smith, R. S. (2012). Predicting outcome of patients with chest wall injury.American Journal of Surgery, 204(6), 910-914.
Rainey, E. E., Petrey, L. B., Reynolds, M., Agtarap, S., & Warren, A. M. (2014). Psychological factors predicting outcome after traumatic injury: the role of resilience. The American Journal of Surgery, 208(4), 517-523.
Rosenfeld, J. V. (2010). How confident can we be in predicting outcome in patients with a minimal Glasgow Coma Score? Re: Nijboer et al. Patients beyond salvation? Various categories of trauma patients with a minimal Glasgow Coma Score.Injury, 41(1), 50-51.
Schmidt, G. W., Broman, A. T., Hindman, H. B., & Grant, M. P. (2008).Vision Survival after Open Globe Injury Predicted by Classification and Regression Tree Analysis.Ophthalmology, 115(1), 202–209.
Streib FE, Dehmer M, Bakir GH, et al. Influence of noise on the inference of dynamic Bayesian networks from short time series. In: Proceedings of World Academy of Science, Engineering and Technology. 2005;70 –74.
Strobl, C., Boulesteix, A.-L., Zeileis, A., &Hothorn, T. (2007). Bias in random forest variable importance measures: illustrations, sources and a solution. BMC Bioinformatics, 8, 25.
Tanoue, K., Aida, N., & Matsui, K. (2013). Apparent diffusion coefficient values predict outcomes of abusive head trauma. ActaPaediatr, 102(8), 805-808.
Teasdale, G. M., Pettigrew, L. E. L., Wilson, J. T. L., Murray, G., &Jennett, B. (1998). Analyzing outcome of treatment of severe head injury: A review and update on advancing the use of the Glasgow Outcome Scale. Journal of Neurotrauma, 15(8), 587-597
Wilson, J. T., Pettigrew, L. E., & Teasdale, G. M. (1998). Structured interviews for the Glasgow Outcome Scale and the extended Glasgow Outcome Scale: guidelines for their use. Journal of Neurotrauma, 15(8), 573–585.
Woodcock, M. G. L., Scott, R. A. H., Huntbach, J., &Kirkby, G. R. (2006). Mass and Shape as Factors in Intraocular Foreign Body Injuries. Ophthalmology, 113(12), 2262–2269.
Walcott, B. P., Khanna, A., Kwon, C.-S., Phillips, H. W., Nahed, B. V, &Coumans, J.-V. (2014). Time interval to surgery and outcomes following the surgical treatment of acute traumatic subdural hematoma.Journal of Clinical Neuroscience : Official Journal of the Neurosurgical Society of Australasia, 21(12), 2107-2111.
