Downloaded from http://jnnp.bmj.com/ on January 6, 2015 - Published by group.bmj.com
JNNP Online First, published on July 1, 2014 as 10.1136/jnnp-2014-307807 Neurosurgery
REVIEW
Artificial neural networks in neurosurgery Parisa Azimi,1 Hasan Reza Mohammadi,1 Edward C Benzel,2 Sohrab Shahzadi,1 Shirzad Azhari,1 Ali Montazeri3 ▸ Additional material is published online only. To view please visit the journal online (http://dx.doi.org/10.1136/ jnnp-2014-307807) 1
Department of Neurosurgery, Shahid Beheshti University of Medical Sciences, Tehran, Iran 2 Department of Neurosurgery, Cleveland Clinic Foundation, Cleveland, Ohio, USA 3 Mental Health Research Group, Health Metrics Research Centre, Iranian Institute for Health Sciences Research, ACECR, Tehran, Iran Correspondence to Dr Parisa Azimi, Functional Neurosurgery Research Center of Shohada Tajrish Hospital, Shahid Beheshti University of Medical Sciences, Sharadari St., Tajrish Square, Tehran 1989934148, Iran; [email protected] Received 5 February 2014 Revised 28 May 2014 Accepted 12 June 2014
ABSTRACT Artificial neural networks (ANNs) effectively analyze nonlinear data sets. The aimed was A review of the relevant published articles that focused on the application of ANNs as a tool for assisting clinical decision-making in neurosurgery. A literature review of all full publications in English biomedical journals (1993–2013) was undertaken. The strategy included a combination of key words ‘artificial neural networks’, ‘prognostic’, ‘brain’, ‘tumor tracking’, ‘head’, ‘tumor’, ‘spine’, ‘classification’ and ‘back pain’ in the title and abstract of the manuscripts using the PubMed search engine. The major findings are summarized, with a focus on the application of ANNs for diagnostic and prognostic purposes. Finally, the future of ANNs in neurosurgery is explored. A total of 1093 citations were identified and screened. In all, 57 citations were found to be relevant. Of these, 50 articles were eligible for inclusion in this review. The synthesis of the data showed several applications of ANN in neurosurgery, including: (1) diagnosis and assessment of disease progression in low back pain, brain tumours and primary epilepsy; (2) enhancing clinically relevant information extraction from radiographic images, intracranial pressure processing, low back pain and realtime tumour tracking; (3) outcome prediction in epilepsy, brain metastases, lumbar spinal stenosis, lumbar disc herniation, childhood hydrocephalus, trauma mortality, and the occurrence of symptomatic cerebral vasospasm in patients with aneurysmal subarachnoid haemorrhage; (4) the use in the biomechanical assessments of spinal disease. ANNs can be effectively employed for diagnosis, prognosis and outcome prediction in neurosurgery.
INTRODUCTION
To cite: Azimi P, Mohammadi HR, Benzel EC, et al. J Neurol Neurosurg Psychiatry Published Online First: [ please include Day Month Year] doi:10.1136/ jnnp-2014-307807
The ability to establish an accurate clinical diagnosis, appreciate clinical patterns ( pattern recognition), analyse and interpret images to facilitate decision-making and ultimately predict optimal treatment is important in choosing the most appropriate management strategy for neurosurgical disorders. Medical informatics has been applied to develop models such as logistic regression (LR) and artificial neural networks (ANNs). LR models are traditional predictive tools. ANNs are computational models based on the functioning of biological neural networks that can be used for non-linear statistical data modelling, with which the complex relationships between inputs and outputs (observed data) are modelled and patterns are revealed. The advantage of neural networks over conventional programming lies in their ability to solve problems that do not have an algorithmic solution or whose available solution is too complex to be readily determined. ANNs are well suited to
Azimi P, et al. J Neurol Neurosurg Psychiatry(or 2014;0:1–6. doi:10.1136/jnnp-2014-307807 Copyright Article author their employer) 2014. Produced
address solvable problems or dilemmas, such as prediction, clinical diagnosis determination, pattern recognition and image analysis and interpretation. The use of ANNs for clinical decision-making support systems began in the late 1980s1 2; however, there has been little use of this method in neurosurgery.1 2 The history and theory of ANNs has been reported in detail elsewhere.1 3 4 The advantages and disadvantages of ANNs have been previously reported.5 6 The advantages of ANNs include: (1) the ability to consider complicated non-linearities between predictors (input parameters) and outcomes, instead of assuming a linear relationship and/or normal distribution between them. Hence, ANNs can better simulate complex biological systems which have a non-linear and non-normal nature; (2) ANNs are universal estimators/predictors and can be applied to any type of data; (3) ANNs avoid the curse of dimensionality so that the estimation/approximation error becomes independent of the dimension of the input (number of variables); (4) robust performance regarding the handling of noisy or incomplete data; (5) ANNs can be trained using multiple training algorithms; (6) ease of construction and use, even with very little mathematical knowledge or model building experience. Moreover, software is currently readily available for building ANNs. The disadvantages of ANNs include: (1) the difficulty associated with the clinical interpretation of model parameters. Since ANNs, like black boxes, represent implicit relationships between input parameters and outcomes, the explicit relationships between predictors and outcomes is difficult to extract; (2) CIs associated with the predicted risks are difficult to determine; (3) the proclivity to overfit the model under training; (4) model development is empirical and few guidelines exist to determine the optimal ANN structure and training algorithms. The purpose of this review is to evaluate the application of ANNs to the clinical functions of diagnosis, prognosis and survival analysis, for neurosurgical applications. The methodology of ANNs was also evaluated.
METHODS A brief introduction to ANN methodology ANN methodology involves three basic steps, namely: data collection, data division and reduction to practice. Data collection is the process of selection of input and output variables according to the parameters defined by the study population. Data division is examination of data normality, in order to modify or delete the values that are obviously
by BMJ Publishing Group Ltd under licence.
1
Downloaded from http://jnnp.bmj.com/ on January 6, 2015 - Published by group.bmj.com
Neurosurgery Data synthesis The data obtained from each study were synthesised by providing descriptive tables reporting authors’ names, publication year, study setting, study sample, disease conditions (where relevant data were available) and the main findings or conclusions. The findings were then sorted and presented chronologically.
RESULTS Statistics A total of 1093 citations were identified and screened. Overall, 1035 papers were found irrelevant (other disease conditions=1031, letter to editor=2, commentary=1 and conference reports=2). Thus, 57 citations were examined for eligibility. Of these, animal studies were excluded (7 papers) and the remaining 50 papers were included in this review (figure 2). Mean size of the testing set, training set and average ratio of the training set-to-testing set were 2171.7 (4158.3), 1375 (SD=3466.2) and 4.9 (SD=13.6), respectively. In general, the application of ANNs can be divided into four major groups: diagnosis, prognosis, outcome prediction and biomechanical assessments. Here, the major findings are summarised and presented under the following headings.
Diagnosis
Figure 1 Schematic representation of artificial neural network (ANN) training procedure.
incongruent. Reduction to practice is the process of ANN application that includes choosing a suitable ANN structure and providing a suitable algorithm for a sample of the data, and then extending the model to all data. The associated flow chart is shown in figure 1.
Search strategy
ANNs have been used for the clinical diagnosis, image analysis and histopathology assessment, data interpretation in the intensive care settings and waveform analysis.7 In neurosurgery, ANNs have been successfully used for diagnosis in paediatric posterior fossa tumours,8 9 low back pain,10–12 cervical spine vertebra,13 scoliosis spinal deformity,14 15 brain tumours,16–21 primary generalised epilepsy using the analysis of EEGs22 and tumour and non-tumour cerebral disorders.23 ANNs have been used to interpret radiographic images,24 25 to enhance surgical decision-making for traumatic brain injury,26 to recognise and correctly diagnose patients with different facial pain syndromes,27 28 to discriminate between the essential tremor and the tremor in Parkinson’s disease (PD),29 for identification of epilepsy seizures and brain tumour,30 for discriminating between normal and PD participants31 and for real-time tumour tracking.32 Each of them is summarised in online supplementary table S1.
Prognosis ANNs have been used to accurately predict survival in patients with brain metastases treated with radiosurgery.2 They have been shown to predict the occurrence of symptomatic cerebral vasospasm in patients with aneurysmal subarachnoid
A literature search was performed using PubMed. The intention was to review all full publications that appeared in the English language biomedical journals. The search strategy included a combination of key words ‘artificial neural networks’, and ‘prognostic’, ‘brain’, ‘tumor tracking’, ‘head’, ‘tumor’, ‘spine’, ‘classification’ and ‘back pain’ in the title/abstract of publications. Since the first study of ANNs in neurosurgery was published in 1993, a time interval was set from 1993 to the present (2013). The initial search was carried out in early 2013 and was updated three times in 2013 (February, September and December).
Inclusion and exclusion criteria All research articles using the ANNs in neurosurgery were included. Papers were excluded if the topic was about other disease conditions or the manuscript dealt with animal studies. 2
Figure 2
Relevant manuscript selection process.
Azimi P, et al. J Neurol Neurosurg Psychiatry 2014;0:1–6. doi:10.1136/jnnp-2014-307807
Downloaded from http://jnnp.bmj.com/ on January 6, 2015 - Published by group.bmj.com
Neurosurgery haemorrhage.33 ANNs have also been used to identify patients with a high risk of hypotension during spinal anaesthesia.34 Also, they have been used to determine the prognosis of head trauma,35 low back pain36 37 and brain metastases.38 Finally, they have been used to develop novel computational tools for deep brain stimulation systems39 and intracranial pressure processing.40 A summary of the studies is shown in online supplementary table S2.
Outcome prediction Online supplementary table S3 summarises the studies that used ANNs for outcome prediction. ANNs have been used to predict outcome in epilepsy surgery,41 lumbar spinal stenosis (LSS),3 lumbar disc herniation (LDH),4 childhood hydrocephalus using endoscopic third ventriculostomy,1 trauma mortality26 42–53 and brain metastasis.2 38
The use of ANNs for the biomechanical assessments of spinal disease ANNs have been used in the biomechanical assessment of spinal disease via the interpretation of joint moments, spinal loads and muscle forces.54–56 They have also been used for the optimisation of the design of spinal pedicle screws,57 to determine the reliability of the patient pain drawing in lumbar spine disorders58 and the prediction of low bone mineral density.59 These findings are summarised in online supplementary table S4.
DISCUSSION Application The present work is the first identifiable literature review of the application of neural networks in neurosurgery. The findings provide a summary of relevant publications and a roadmap to guide future research relating to the application of ANN in neurosurgery. Specifically, the findings from the literature indicated that using ANN models in neurosurgery could be used to enhance the practice of neurosurgery in several ways—including disease prognosis, outcome prediction and biomechanical study design. In fact, ANN models have been successfully applied to various areas of medicine, such as diagnostic systems, biomedical analysis, image analysis and discovery of new drugs, and to solve drug development challenges and policies.42 60 As such, for instance, when using X-ray, CT, MRI or radioisotope scans in neurosurgery, it is possible to use human observations and direct digitised images as inputs to apply the ANN models.7 In this particular example, the application of ANNs was used to develop automated decision-making support to assist with the classification of brain tumours. In addition, medical images could be analysed, processed and characterised by ANNs. By selecting an appropriate training set and learning process, ANNs become appropriate for recognition of unusual images.24 There are many types of ANNs such as multilayer perceptron (MLP), radial basis function, etc. The MLP type ANN model is one of the most widely applied models. Based on the training algorithms, ANN models are comprised of an input layer, a hidden layer and an output layer (representing ANN architecture). Forty nine of the 50 studies use the MLP type; however, in some studies, has been described in more detail with training algorithms as shown in the tables. It can be argued that the use of ANN modelling in neurosurgery is underutilised. However, ANNs are used for simulating a complicated and unclear relationship between parameters of the system under study. Since this relationship is unknown and different systems have different parameter inter-relationships, the introduction of a universal ANN remains a controversial issue. Therefore, finding a suitable Azimi P, et al. J Neurol Neurosurg Psychiatry 2014;0:1–6. doi:10.1136/jnnp-2014-307807
ANN is usually done by first choosing a network structure such as MLP and then modifying it by network pruning or network growing until a network with the best achievable performance is found. In addition, the weighting of initial network structure values can influence the estimation error. Choosing suitable initial values will lead to a more rapid convergence to the desired network during ANN training, with associated improved performance. Since the nature of the system under study is unclear, these values are chosen by trial and error. The experience of the ANN trainer can be very helpful in this process. In addition, the ANN models often outperform linear models due to the simplicity of the model, but the SE is within an acceptable range. However, it seems to be considered as a sensitivity analysis using a variety of artificial intelligence methodologies commonly used in the final production tool for the clinical decision-making process. This review has identified papers that have sample sizes ranging from 37 to over 35 000. One might inquire about the minimum adequate size for a training set or about the relationships between the size of the training set and the performance of the model. Certainly the sample size significantly affects the training and final performance of the network. The greater the number of samples, the more effective the training. This translates to an improved simulation of the system under study by the trained ANN. In addition, the complexity level of a system plays a major role in the determination of the sample size. Greater complexity obligates a larger sample size. Thus, all new received data can advance the performance of the ANN estimator. Consequently, the need arises for a bank of samples. However, as a guide, it should be noted that when the input of more new samples has no notable impact on the lowering of the estimation error, there exists no need for more samples. The classification of neurosurgery disorders into different prognostic groups is an important clinical and research endeavour due to the potential impact on patient treatment and clinical trials design. For instance, AANs have been used for the classification of brain tumours. The associated studies showed that the proposed method could assist radiologists in formulating optimal decisions for classifying brain tumours.18 19 Also, ANNs have a huge potential in the arena of pattern recognition, which includes tasks such as the determination of the diagnosis and the identification of clinical anomalies, laboratory analysis and neurosurgery image processing. Prognosis is extremely important in planning appropriate treatment strategies and follow-up assessments. Accurate identification of high-risk patients may facilitate targeted aggressive adjuvant therapy that may help cure the disease and prolong survival. ANNs, with their ability to exploit non-linear relationships between variables, are particularly suitable for analysis of complex cancer data.2 7 There is an escalating interest in developing satisfactory models for outcome prediction in clinical practice. The ANN model is being increasingly employed for such purposes. Many researchers have constructed predictive models for outcome prediction after traumatic brain injury. Most of these attempts focused on outcomes with two possibilities such as survival versus death or good outcomes versus poor outcomes.26 42–46 48 49 However, one study focused on ANN prediction of outcomes in five possible categories based on the Glasgow Outcome Scale (death, persistent vegetative state, severe disability, moderate disability and good recovery) after moderate-to-severe head injury.48 This strategy has also been applied to brain metastasis, LSS and LDH.2–4 Spinal stability is maintained by a variety of anatomic structures that have evolved to provide resistance against deforming 3
Downloaded from http://jnnp.bmj.com/ on January 6, 2015 - Published by group.bmj.com
Neurosurgery forces. It has been associated with excessive mechanical loads on the human spine during physical activities such as lifting and measurements of force. Since the direct in vivo measurements of spinal loads and muscle forces are invasive, investigators have decided to apply ANNs for spinal biomechanical assessment by using different possible models.54 Thus, it is argued that an understanding of biomechanical assessments using ANNs is essential in spine surgery.54–57
Potential future application of ANN in neurosurgery Arguably, ANNs are potentially more successful than traditional statistical models in predicting clinical outcomes. It is believed that complex medical diagnostic and treatment decisions will be increasingly based on ANNs for neurosurgical disorders in the next few years. Already, ANNs have been successfully applied to various areas of medicine, such as image analysis and drug development.42 60 This experience can also be used in the domain of neurosurgical disorders. Several computer-aided diagnosis projects can be performed for neurosurgery disorders, including the use of ANNs for predicting outcomes for operative versus non-operative treatment strategies for disorders and as a discriminating classifier for tasks regarding medical diagnosis for the early detection of diseases based on clinical and imaging findings. A variety of well-known instruments exist for assessing performance status or functionality in neurosurgery disorders. However, no existing standard cut-off points exist for the data acquired by these instruments for the assessment of operative versus non-operative outcomes in these patients. If standard cut-off points are subsequently made available, such instruments may be used as output variables for the development of an estimator software package. The structure and details of an ANN that does not require further training can be saved and used to create a decision-making software tool. At present, the literature lacks examples of the clinical application of ANNs in neurosurgery, although ANNs have been employed in the research arena. An improved portable computing technology diminishes the barrier that once existed. To develop and modify an ANN model as a universally available tool, it is recommended that an international centre be considered for the following issues: A. The improvement of data collection for training, such as online data captured from the process, the data division and preprocessing. The area for ANN development that harbours the most hope is associated with their ability to act as ‘data refineries’ for large and complex data sets. Their ability to make decisions based on the experience of thousands of individual records is particularly attractive. An ANN tool that correctly trained, containing the post hoc experience of the clinical features and outcome of complex sample sets, which makes for more accurate prediction, is a potentially very valuable clinical tool. B. The choice of suitable network architecture, careful selection of some internal parameters that control the optimisation methodology and stopping criteria. The development of models using ANNs is associated with the fact that there set methods do not exist for building the architecture of the network. The most common type of ANN, though, is the feed-forward back propagation MLP. C. Improving the user-friendly interface between the user and the computer. D. Model validation. External validation is necessary to obtain an accurate measure of performance outside the development case. 4
E. To evaluate ANNs as tools for performing randomised clinical trials (RCT). We know of no RCT examining the impact of ANN output on clinical actions or patient outcomes. The tool will be regularly reassessed as new information becomes available. For example, some medical applications of ANN tools and related techniques are presented in web services.61 62 With the explosion in clinical data over the past few years, we anticipate a growing need for modern machine learning techniques such as ANNs to properly make use of the information at our disposal and realise its full benefit for clinical care. To optimise the ANN software products or programs using Matlab or other toolboxes, it is recommended that an international centre be considered for the collection of real-world outcomes. Prognostic factors can be used as input variables. The number of patients in ANN models can be increased to develop and modify the model. We encourage other researchers to share their data.
Limitations of ANNs This review has limitations. One major limitation is the fact that it covers a very heterogeneous data analysis and modelling technique applied across a very heterogeneous (and idiosyncratic) set of clinical conditions. Therefore, it cannot provide an in-depth and focused synthesis of the literature. However, since this is the first attempt to summarise the evidence regarding ANNs in neurosurgery, future reviews might reasonably concentrate on one specific issue. For example, as suggested, upcoming reviews study very small and specific applications of ANNs to clinical practice (eg, “ANNs for the interpretation of radiographic images in traumatic brain injury (TBI)”) and present the various approaches and strategies used, as well as differences in the implementation of each ANN for that purpose, along with specific strengths and weaknesses. In addition, one should first note that the search strategy was limited to the key words in the title/ abstract of the publications. Thus, we might have missed some papers. Second, this work restricted the query search for articles in PubMed. Third, non-English publications were not considered in this study. We believe research regarding the application of ANNs in neurosurgery have also been published in other languages. Finally, there are possibly countless applications of ANNs in the field of neurosurgery. However, a common criticism of ANNs is that they require significant training and ‘know-how’ for real-world operations. This is not unexpected, since any learning machine is obligatorily associated with such limitations. These are imminently solvable. ANNs will never replace human experts, but they can help in screening and also be used by experts to double-check their diagnosis. Perhaps, more importantly, though, is the fact that ANNs can be used to identify variables that experts do not ‘see’, thus enhancing the diagnostic acumen of the expert.63 At present, there are no articles published that describe the use of ANN outputs in clinical practice. This is due to several factors, such as their position in the decision-making process, the lack of a support substructure, difficulties with assessment of performance, etc. It is difficult to begin to assess the true value of ANNs on clinical practice as no RCT data exist that specifically examine this. ANNs will never replace human expert decision-makers, but they can assist in double-checking the routine decision-making process.
CONCLUSION Owing to its high accuracy, ANNs can be effectively employed for diagnosis, prognosis and outcome prediction in neurosurgery. Azimi P, et al. J Neurol Neurosurg Psychiatry 2014;0:1–6. doi:10.1136/jnnp-2014-307807
Downloaded from http://jnnp.bmj.com/ on January 6, 2015 - Published by group.bmj.com
Neurosurgery Acknowledgements The authors thank the staff of the Neurosurgery Unit at Imam-Hossain Hospital, Tehran, Iran. Contributors PA was involved in the conception and design, acquisition of the data and drafting of the article. PA and AM were involved in the analysis and interpretation of data. ECB was involved in the study supervision. All authors critically revised the article and reviewed the submitted version of the manuscript. PA approved the final version of the manuscript on behalf of all authors.
24 25
26
Competing interests None. Ethics approval The research was approved by the Ethics Committee of ShahidBeheshti University of Medical Sciences, Tehran, Iran. Provenance and peer review Not commissioned; externally peer reviewed.
27 28
29
REFERENCES 1
2
3
4
5 6
7 8
9 10
11
12 13 14 15 16
17
18
19
20
21
22
23
Azimi P, Mohammadi HR. Prediction of successful ETV outcome in childhood hydrocephalus: an artificial neural networks analysis. J Neurosurg Pediatr 2014;13:426–32. Oermann EK, Kress MAS, Collins BT, et al. Predicting survival in patients with brain metastases treated with radiosurgery using artificial neural networks. Neurosurgery 2013;72:944–51. Azimi P, Benzel EC, Shahzadi S, et al. Use of artificial neural networks to predict surgical satisfaction in patients with lumbar spinal canal stenosis. J Neurosurg Spine 2014;20:300–5. Azimi P, Benzel EC, Shahzadi S, et al. The prediction of successful surgery outcome in lumbar disc herniation based on artificial neural networks. J Neurosurg Sci In press. Dreiseitl S, Ohno-Machado L. Logistic regression and artificial neural network classification models: a methodology review. J Biomed Inform 2002;35:352–9. Tu JV. Advantages and disadvantages of using artificial neural networks versus logistic regression for predicting medical outcomes. J Clin Epidemiol 1996;49:1225–31. Ramesh AN, Kambhampati C, Monson JR, et al. Artificial intelligence in medicine. Ann R Coll Surg Engl 2004;86:334–8. Arle JE, Morriss C, Wang ZJ, et al. Prediction of posterior fossa tumor type in children by means of magnetic resonance image properties, spectroscopy, and neural networks. J Neurosurg 1997;86:755–61. Bidiwala S, Pittman T. Neural network classification of pediatric posterior fossa tumors using clinical and imaging data. Pediatr Neurosurg 2004;40:8–15. Bishop JB, Szpalski M, Ananthraman SK, et al. Classification of low back pain from dynamic motion characteristics using an artificial neural network. Spine 1997;22:2991–8. Liszka-Hackzell JJ, Martin DP. Categorization and analysis of pain and activity in patients with low back pain using a neural network technique. J Med Syst 2002;26:337–47. Sari M, Gulbandilar E, Cimbiz A. Prediction of low back pain with two expert systems. J Med Syst 2010;36:1523–7. Stanley RJ, Long R. A radius of curvature-based approach to cervical spine vertebra image analysis. Biomed Sci Instrum 2001;37:385–90. Lin H. Identification of spinal deformity classification with total curvature analysis and artificial neural network. IEEE Trans Biomed Eng 2008;55:376–82. Jaremko JL, Poncet P, Ronsky J, et al. Estimation of spinal deformity in scoliosis from torso surface cross sections. Spine 2001;26:1583–91. Iglesias J, Esparza J, Scherf M. Diagnosis of human oligodendrogliomas with the help of the NeuroShell Easy Classifier neural network. Anal Quant Cytol Histol 2000;22:383–92. Kolles H, von-Wangenheim A, Vince GH, et al. Automated grading of astrocytomas based on histomorphometric analysis of Ki-67 and Feulgen stained paraffin sections. Classification results of neuronal networks and discriminant analysis. Anal Cell Pathol 1995;8:101–16. Petalidis LP, Oulas A, Backlund M, et al. Improved grading and survival prediction of human astrocytic brain tumors by artificial neural network analysis of gene expression microarray data. Mol Cancer Ther 2008;7:1013–24. Sachdeva J, Kumar V, Gupta I, et al. A dual neural network ensemble approach for multiclass brain tumor classification. Int J Numer Method Biomed Eng 2012;28:1107–20. Sieben G, Praet M, Roels H, et al. The development of a decision support system for the pathological diagnosis of human cerebral tumours based on a neural network classifier. Acta Neurochir (Wien) 1994;129:193–7. Usenius JP, Tuohimetsä S, Vainio P, et al. Automated classification of human brain tumours by neural network analysis using in vivo 1H magnetic resonance spectroscopic metabolite phenotypes. Neuroreport 1996;7:1597–600. Oğulata SN, Sahin C, Erol R. Neural network-based computer-aided diagnosis in classification of primary generalized epilepsy by EEG signals. J Med Syst 2009;33:107–12. Poptani H, Kaartinen J, Gupta RK, et al. Diagnostic assessment of brain tumours and non-neoplastic brain disorders in vivo using proton nuclear magnetic resonance
Azimi P, et al. J Neurol Neurosurg Psychiatry 2014;0:1–6. doi:10.1136/jnnp-2014-307807
30 31
32
33
34
35
36
37
38 39
40
41 42 43
44
45
46
47 48 49
50
51
spectroscopy and artificial neural networks. J Cancer Res Clin Oncol 1999;125:343–9. Jiang J, Trundle P, Ren J. Medical image analysis with artificial neural networks. Comput Med Imaging Graph 2010;34:617–31. Sharma N, Ray AK, Sharma S, et al. Segmentation and classification of medical images using texture-primitive features: application of BAM-type artificial neural network. J Med Phys 2008;33:119–26. Li YC, Liu L, Chiu WT, et al. Neural network modeling for surgical decisions on traumatic brain injury patients. Int J Med Inform 2000;57:1–9. Limonadi FM, McCartney S, Burchiel KJ. Design of an artificial neural network for diagnosis of facial pain syndromes. Stereotact Funct Neurosurg 2006;84:212–20. McCartney S, Weltin M, Burchiel KJ. Use of an artificial neural network for diagnosis of facial pain syndromes: an update. Stereotact Funct Neurosurg 2013;92:44–52. Hossen A. A neural network approach for feature extraction and discrimination between Parkinsonian tremor and essential tremor. Technol Health Care 2013;21:345–56. Sharanreddy M, Kulkarni PK. Automated EEG signal analysis for identification of epilepsy seizures and brain tumour. J Med Eng Technol 2013;37:511–19. Muniz AM, Liu W, Liu H, et al. Assessment of the effects of subthalamic stimulation in Parkinson disease patients by artificial neural network. Conf Proc IEEE Eng Med Biol Soc 2009;2009:5673–6. Seregni M, Pella A, Riboldi M, et al. Real-time tumor tracking with an artificial neural networks-based method: a feasibility study. Phys Med 2013;29:48–59. Dumont TM, Rughani AI, Tranmer BI. Prediction of symptomatic cerebral vasospasm after aneurysmal subarachnoid hemorrhage with an artificial neural network: feasibility and comparison with logistic regression models. World Neurosurg 2011;75:57–63. Lin CS, Chiu JS, Hsieh MH, et al. Predicting hypotensive episodes during spinal anesthesia with the application of artificial neural networks. Comput Methods Programs Biomed 2008;92:193–7. Sinha M, Kennedy CS, Ramundo ML. Artificial neural network predicts CT scan abnormalities in pediatric patients with closed head injury. J Trauma 2001;50:308–12. Dickey JP, Pierrynowski MR, Bednar DA, et al. Relationship between pain and vertebral motion in chronic low-back pain subjects. Clin Biomech (Bristol, Avon) 2002;17:345–52. Parsaeian M, Mohammad K, Mahmoudi M, et al. Comparison of logistic regression and artificial neural network in low back pain prediction: second national health survey. Iran J Public Health 2012;41:86–92. Viani GA, da Silva LG, Stefano EJ. Prognostic indexes for brain metastases: which is the most powerful? Int J Radiat Oncol Biol Phys 2012;83:e325–30. Chaturvedi A, Luján JL, McIntyre CC. Artificial neural network based characterization of the volume of tissue activated during deep brain stimulation. J Neural Eng 2013;10:056023. Mariak Z, Swiercz M, Krejza J, et al. Intracranial pressure processing with artificial neural networks: classification of signal properties. Acta Neurochir (Wien) 2000;142:407–11. Arle JE, Perrine K, Devinsky O, et al. Neural network analysis of preoperative variables and outcome in epilepsy surgery. J Neurosurg 1999;90:998–1004. Cross SS, Harrison RF, Kennedy RL. Introduction to neural networks. Lancet 1995;346:1075–9. DiRusso SM, Sullivan T, Holly C, et al. An artificial neural network as a model for prediction of survival in trauma patients: validation for a regional trauma area. J Trauma 2000;49:212–23. Edwards DF, Hollingsworth H, Zazulia AR, et al. Artificial neural networks improve the prediction of mortality in intracerebral hemorrhage. Neurology 1999;53:351–7. Eftekhar B, Mohammad K, Ardebili HE, et al. Comparison of artificial neural network and logistic regression models for prediction of mortality in head trauma based on initial clinical data. BMC Med Inform Decis Mak 2005;5:3. Hsu MH, Li YC, Chiu WT, et al. Outcome prediction after moderate and severe head injury using an artificial neural network. Stud Health Technol Inform 2005;116:241–5. Izenberg SD, Williams MD, Luterman A. Prediction of trauma mortality using a neural network. Am Surg 1997;63:275–81. Rughani AI, Dumont TM, Lu Z, et al. Use of an artificial neural network to predict head injury outcome. J Neurosurg 2010;113:585–90. Shi HY, Hwang SL, Lee KT, et al. In-hospital mortality after traumatic brain injury surgery: a nationwide population-based comparison of mortality predictors used in artificial neural network and logistic regression models. J Neurosurg 2013;118:746–52. Lang EW, Pitts LH, Damron SL, et al. Outcome after severe head injury: an analysis of prediction based upon comparison of neural network versus logistic regression analysis. Neurol Res 1997;19:274–80. Erol FS, Uysal H, Ergün U, et al. Prediction of minor head injured patients using logistic regression and MLP neural network. J Med Syst 2005;29:205–15.
5
Downloaded from http://jnnp.bmj.com/ on January 6, 2015 - Published by group.bmj.com
Neurosurgery 52 53
54
55
56
6
Davis DP, Peay J, Sise MJ, et al. The impact of prehospital endotracheal intubation on outcome in moderate to severe traumatic brain injury. J Trauma 2005;58:933–9. Abouzari M, Rashidi A, Zandi-Toghani M, et al. Chronic subdural hematoma outcome prediction using logistic regression and an artificial neural network. Neurosurg Rev 2009;32:479–84. Arjmand N, Ekrami O, Shirazi-adl A, et al. relative performances of artificial neural network and regression mapping tools in evaluation of spinal loads and muscle forces during static lifting. J Biomech 2013;46:1454–62. Wang L, Buchanan TS. Prediction of joint moments using a neural network model of muscle activations from EMG signals. IEEE Trans Neural Syst Rehabil Eng 2002;10:30–7. Nussbaum MA, Chaffin DB. Pattern classification reveals intersubject group differences in lumbar muscle recruitment during static loading. Clin Biomech (Bristol, Avon) 1997;12:97–106.
57
58 59
60 61 62 63
Amaritsakul Y, Chao CK, Lin J. Multiobjective optimization design of spinal pedicle screws using neural networks and genetic algorithm: mathematical models and mechanical validation. Comput Math Methods Med 2013;2013:462875. Mann NH, Brown MD, Hertz DB, et al. Initial-impression diagnosis using low-back pain patient pain drawings. Spine 1993;18:41–53. Ongphiphadhanakul B, Rajatanavin R, Chailurkit L, et al. Prediction of low bone mineral density in postmenopausal women by artificial neural network model compared to logistic regression model. J Med Assoc Thai 1997;80:508–15. http://library.thinkquest.org/C007395/tqweb/history.html, 2013. http://neuroxl.com/applications/medicine/neural-networks-in-medicine/index.htm, 2013. http://www.phil.gu.se/ann/annworld.html, 2013. Rughani AI, Dumont TM, Tranmer BI. Editorial: predicting surgical satisfaction using artificial neural networks. J Neurosurg Spine 2014;20:298–9.
Azimi P, et al. J Neurol Neurosurg Psychiatry 2014;0:1–6. doi:10.1136/jnnp-2014-307807
Downloaded from http://jnnp.bmj.com/ on January 6, 2015 - Published by group.bmj.com
Artificial neural networks in neurosurgery Parisa Azimi, Hasan Reza Mohammadi, Edward C Benzel, Sohrab Shahzadi, Shirzad Azhari and Ali Montazeri J Neurol Neurosurg Psychiatry published online July 1, 2014
Updated information and services can be found at: http://jnnp.bmj.com/content/early/2014/07/01/jnnp-2014-307807
Supplementary Material
Supplementary material can be found at: http://jnnp.bmj.com/content/suppl/2014/07/01/jnnp-2014-307807.DC1. html These include:
References
This article cites 59 articles, 1 of which you can access for free at: http://jnnp.bmj.com/content/early/2014/07/01/jnnp-2014-307807#BIBL
Email alerting service
Receive free email alerts when new articles cite this article. Sign up in the box at the top right corner of the online article.
Topic Collections
Articles on similar topics can be found in the following collections Epilepsy and seizures (738) CNS cancer (168) Neurooncology (221) Pain (neurology) (656) Musculoskeletal syndromes (489) Hydrocephalus (127)
Notes
To request permissions go to: http://group.bmj.com/group/rights-licensing/permissions To order reprints go to: http://journals.bmj.com/cgi/reprintform To subscribe to BMJ go to: http://group.bmj.com/subscribe/
JNNP Online First, published on July 1, 2014 as 10.1136/jnnp-2014-307807 Neurosurgery
REVIEW
Artificial neural networks in neurosurgery Parisa Azimi,1 Hasan Reza Mohammadi,1 Edward C Benzel,2 Sohrab Shahzadi,1 Shirzad Azhari,1 Ali Montazeri3 ▸ Additional material is published online only. To view please visit the journal online (http://dx.doi.org/10.1136/ jnnp-2014-307807) 1
Department of Neurosurgery, Shahid Beheshti University of Medical Sciences, Tehran, Iran 2 Department of Neurosurgery, Cleveland Clinic Foundation, Cleveland, Ohio, USA 3 Mental Health Research Group, Health Metrics Research Centre, Iranian Institute for Health Sciences Research, ACECR, Tehran, Iran Correspondence to Dr Parisa Azimi, Functional Neurosurgery Research Center of Shohada Tajrish Hospital, Shahid Beheshti University of Medical Sciences, Sharadari St., Tajrish Square, Tehran 1989934148, Iran; [email protected] Received 5 February 2014 Revised 28 May 2014 Accepted 12 June 2014
ABSTRACT Artificial neural networks (ANNs) effectively analyze nonlinear data sets. The aimed was A review of the relevant published articles that focused on the application of ANNs as a tool for assisting clinical decision-making in neurosurgery. A literature review of all full publications in English biomedical journals (1993–2013) was undertaken. The strategy included a combination of key words ‘artificial neural networks’, ‘prognostic’, ‘brain’, ‘tumor tracking’, ‘head’, ‘tumor’, ‘spine’, ‘classification’ and ‘back pain’ in the title and abstract of the manuscripts using the PubMed search engine. The major findings are summarized, with a focus on the application of ANNs for diagnostic and prognostic purposes. Finally, the future of ANNs in neurosurgery is explored. A total of 1093 citations were identified and screened. In all, 57 citations were found to be relevant. Of these, 50 articles were eligible for inclusion in this review. The synthesis of the data showed several applications of ANN in neurosurgery, including: (1) diagnosis and assessment of disease progression in low back pain, brain tumours and primary epilepsy; (2) enhancing clinically relevant information extraction from radiographic images, intracranial pressure processing, low back pain and realtime tumour tracking; (3) outcome prediction in epilepsy, brain metastases, lumbar spinal stenosis, lumbar disc herniation, childhood hydrocephalus, trauma mortality, and the occurrence of symptomatic cerebral vasospasm in patients with aneurysmal subarachnoid haemorrhage; (4) the use in the biomechanical assessments of spinal disease. ANNs can be effectively employed for diagnosis, prognosis and outcome prediction in neurosurgery.
INTRODUCTION
To cite: Azimi P, Mohammadi HR, Benzel EC, et al. J Neurol Neurosurg Psychiatry Published Online First: [ please include Day Month Year] doi:10.1136/ jnnp-2014-307807
The ability to establish an accurate clinical diagnosis, appreciate clinical patterns ( pattern recognition), analyse and interpret images to facilitate decision-making and ultimately predict optimal treatment is important in choosing the most appropriate management strategy for neurosurgical disorders. Medical informatics has been applied to develop models such as logistic regression (LR) and artificial neural networks (ANNs). LR models are traditional predictive tools. ANNs are computational models based on the functioning of biological neural networks that can be used for non-linear statistical data modelling, with which the complex relationships between inputs and outputs (observed data) are modelled and patterns are revealed. The advantage of neural networks over conventional programming lies in their ability to solve problems that do not have an algorithmic solution or whose available solution is too complex to be readily determined. ANNs are well suited to
Azimi P, et al. J Neurol Neurosurg Psychiatry(or 2014;0:1–6. doi:10.1136/jnnp-2014-307807 Copyright Article author their employer) 2014. Produced
address solvable problems or dilemmas, such as prediction, clinical diagnosis determination, pattern recognition and image analysis and interpretation. The use of ANNs for clinical decision-making support systems began in the late 1980s1 2; however, there has been little use of this method in neurosurgery.1 2 The history and theory of ANNs has been reported in detail elsewhere.1 3 4 The advantages and disadvantages of ANNs have been previously reported.5 6 The advantages of ANNs include: (1) the ability to consider complicated non-linearities between predictors (input parameters) and outcomes, instead of assuming a linear relationship and/or normal distribution between them. Hence, ANNs can better simulate complex biological systems which have a non-linear and non-normal nature; (2) ANNs are universal estimators/predictors and can be applied to any type of data; (3) ANNs avoid the curse of dimensionality so that the estimation/approximation error becomes independent of the dimension of the input (number of variables); (4) robust performance regarding the handling of noisy or incomplete data; (5) ANNs can be trained using multiple training algorithms; (6) ease of construction and use, even with very little mathematical knowledge or model building experience. Moreover, software is currently readily available for building ANNs. The disadvantages of ANNs include: (1) the difficulty associated with the clinical interpretation of model parameters. Since ANNs, like black boxes, represent implicit relationships between input parameters and outcomes, the explicit relationships between predictors and outcomes is difficult to extract; (2) CIs associated with the predicted risks are difficult to determine; (3) the proclivity to overfit the model under training; (4) model development is empirical and few guidelines exist to determine the optimal ANN structure and training algorithms. The purpose of this review is to evaluate the application of ANNs to the clinical functions of diagnosis, prognosis and survival analysis, for neurosurgical applications. The methodology of ANNs was also evaluated.
METHODS A brief introduction to ANN methodology ANN methodology involves three basic steps, namely: data collection, data division and reduction to practice. Data collection is the process of selection of input and output variables according to the parameters defined by the study population. Data division is examination of data normality, in order to modify or delete the values that are obviously
by BMJ Publishing Group Ltd under licence.
1
Downloaded from http://jnnp.bmj.com/ on January 6, 2015 - Published by group.bmj.com
Neurosurgery Data synthesis The data obtained from each study were synthesised by providing descriptive tables reporting authors’ names, publication year, study setting, study sample, disease conditions (where relevant data were available) and the main findings or conclusions. The findings were then sorted and presented chronologically.
RESULTS Statistics A total of 1093 citations were identified and screened. Overall, 1035 papers were found irrelevant (other disease conditions=1031, letter to editor=2, commentary=1 and conference reports=2). Thus, 57 citations were examined for eligibility. Of these, animal studies were excluded (7 papers) and the remaining 50 papers were included in this review (figure 2). Mean size of the testing set, training set and average ratio of the training set-to-testing set were 2171.7 (4158.3), 1375 (SD=3466.2) and 4.9 (SD=13.6), respectively. In general, the application of ANNs can be divided into four major groups: diagnosis, prognosis, outcome prediction and biomechanical assessments. Here, the major findings are summarised and presented under the following headings.
Diagnosis
Figure 1 Schematic representation of artificial neural network (ANN) training procedure.
incongruent. Reduction to practice is the process of ANN application that includes choosing a suitable ANN structure and providing a suitable algorithm for a sample of the data, and then extending the model to all data. The associated flow chart is shown in figure 1.
Search strategy
ANNs have been used for the clinical diagnosis, image analysis and histopathology assessment, data interpretation in the intensive care settings and waveform analysis.7 In neurosurgery, ANNs have been successfully used for diagnosis in paediatric posterior fossa tumours,8 9 low back pain,10–12 cervical spine vertebra,13 scoliosis spinal deformity,14 15 brain tumours,16–21 primary generalised epilepsy using the analysis of EEGs22 and tumour and non-tumour cerebral disorders.23 ANNs have been used to interpret radiographic images,24 25 to enhance surgical decision-making for traumatic brain injury,26 to recognise and correctly diagnose patients with different facial pain syndromes,27 28 to discriminate between the essential tremor and the tremor in Parkinson’s disease (PD),29 for identification of epilepsy seizures and brain tumour,30 for discriminating between normal and PD participants31 and for real-time tumour tracking.32 Each of them is summarised in online supplementary table S1.
Prognosis ANNs have been used to accurately predict survival in patients with brain metastases treated with radiosurgery.2 They have been shown to predict the occurrence of symptomatic cerebral vasospasm in patients with aneurysmal subarachnoid
A literature search was performed using PubMed. The intention was to review all full publications that appeared in the English language biomedical journals. The search strategy included a combination of key words ‘artificial neural networks’, and ‘prognostic’, ‘brain’, ‘tumor tracking’, ‘head’, ‘tumor’, ‘spine’, ‘classification’ and ‘back pain’ in the title/abstract of publications. Since the first study of ANNs in neurosurgery was published in 1993, a time interval was set from 1993 to the present (2013). The initial search was carried out in early 2013 and was updated three times in 2013 (February, September and December).
Inclusion and exclusion criteria All research articles using the ANNs in neurosurgery were included. Papers were excluded if the topic was about other disease conditions or the manuscript dealt with animal studies. 2
Figure 2
Relevant manuscript selection process.
Azimi P, et al. J Neurol Neurosurg Psychiatry 2014;0:1–6. doi:10.1136/jnnp-2014-307807
Downloaded from http://jnnp.bmj.com/ on January 6, 2015 - Published by group.bmj.com
Neurosurgery haemorrhage.33 ANNs have also been used to identify patients with a high risk of hypotension during spinal anaesthesia.34 Also, they have been used to determine the prognosis of head trauma,35 low back pain36 37 and brain metastases.38 Finally, they have been used to develop novel computational tools for deep brain stimulation systems39 and intracranial pressure processing.40 A summary of the studies is shown in online supplementary table S2.
Outcome prediction Online supplementary table S3 summarises the studies that used ANNs for outcome prediction. ANNs have been used to predict outcome in epilepsy surgery,41 lumbar spinal stenosis (LSS),3 lumbar disc herniation (LDH),4 childhood hydrocephalus using endoscopic third ventriculostomy,1 trauma mortality26 42–53 and brain metastasis.2 38
The use of ANNs for the biomechanical assessments of spinal disease ANNs have been used in the biomechanical assessment of spinal disease via the interpretation of joint moments, spinal loads and muscle forces.54–56 They have also been used for the optimisation of the design of spinal pedicle screws,57 to determine the reliability of the patient pain drawing in lumbar spine disorders58 and the prediction of low bone mineral density.59 These findings are summarised in online supplementary table S4.
DISCUSSION Application The present work is the first identifiable literature review of the application of neural networks in neurosurgery. The findings provide a summary of relevant publications and a roadmap to guide future research relating to the application of ANN in neurosurgery. Specifically, the findings from the literature indicated that using ANN models in neurosurgery could be used to enhance the practice of neurosurgery in several ways—including disease prognosis, outcome prediction and biomechanical study design. In fact, ANN models have been successfully applied to various areas of medicine, such as diagnostic systems, biomedical analysis, image analysis and discovery of new drugs, and to solve drug development challenges and policies.42 60 As such, for instance, when using X-ray, CT, MRI or radioisotope scans in neurosurgery, it is possible to use human observations and direct digitised images as inputs to apply the ANN models.7 In this particular example, the application of ANNs was used to develop automated decision-making support to assist with the classification of brain tumours. In addition, medical images could be analysed, processed and characterised by ANNs. By selecting an appropriate training set and learning process, ANNs become appropriate for recognition of unusual images.24 There are many types of ANNs such as multilayer perceptron (MLP), radial basis function, etc. The MLP type ANN model is one of the most widely applied models. Based on the training algorithms, ANN models are comprised of an input layer, a hidden layer and an output layer (representing ANN architecture). Forty nine of the 50 studies use the MLP type; however, in some studies, has been described in more detail with training algorithms as shown in the tables. It can be argued that the use of ANN modelling in neurosurgery is underutilised. However, ANNs are used for simulating a complicated and unclear relationship between parameters of the system under study. Since this relationship is unknown and different systems have different parameter inter-relationships, the introduction of a universal ANN remains a controversial issue. Therefore, finding a suitable Azimi P, et al. J Neurol Neurosurg Psychiatry 2014;0:1–6. doi:10.1136/jnnp-2014-307807
ANN is usually done by first choosing a network structure such as MLP and then modifying it by network pruning or network growing until a network with the best achievable performance is found. In addition, the weighting of initial network structure values can influence the estimation error. Choosing suitable initial values will lead to a more rapid convergence to the desired network during ANN training, with associated improved performance. Since the nature of the system under study is unclear, these values are chosen by trial and error. The experience of the ANN trainer can be very helpful in this process. In addition, the ANN models often outperform linear models due to the simplicity of the model, but the SE is within an acceptable range. However, it seems to be considered as a sensitivity analysis using a variety of artificial intelligence methodologies commonly used in the final production tool for the clinical decision-making process. This review has identified papers that have sample sizes ranging from 37 to over 35 000. One might inquire about the minimum adequate size for a training set or about the relationships between the size of the training set and the performance of the model. Certainly the sample size significantly affects the training and final performance of the network. The greater the number of samples, the more effective the training. This translates to an improved simulation of the system under study by the trained ANN. In addition, the complexity level of a system plays a major role in the determination of the sample size. Greater complexity obligates a larger sample size. Thus, all new received data can advance the performance of the ANN estimator. Consequently, the need arises for a bank of samples. However, as a guide, it should be noted that when the input of more new samples has no notable impact on the lowering of the estimation error, there exists no need for more samples. The classification of neurosurgery disorders into different prognostic groups is an important clinical and research endeavour due to the potential impact on patient treatment and clinical trials design. For instance, AANs have been used for the classification of brain tumours. The associated studies showed that the proposed method could assist radiologists in formulating optimal decisions for classifying brain tumours.18 19 Also, ANNs have a huge potential in the arena of pattern recognition, which includes tasks such as the determination of the diagnosis and the identification of clinical anomalies, laboratory analysis and neurosurgery image processing. Prognosis is extremely important in planning appropriate treatment strategies and follow-up assessments. Accurate identification of high-risk patients may facilitate targeted aggressive adjuvant therapy that may help cure the disease and prolong survival. ANNs, with their ability to exploit non-linear relationships between variables, are particularly suitable for analysis of complex cancer data.2 7 There is an escalating interest in developing satisfactory models for outcome prediction in clinical practice. The ANN model is being increasingly employed for such purposes. Many researchers have constructed predictive models for outcome prediction after traumatic brain injury. Most of these attempts focused on outcomes with two possibilities such as survival versus death or good outcomes versus poor outcomes.26 42–46 48 49 However, one study focused on ANN prediction of outcomes in five possible categories based on the Glasgow Outcome Scale (death, persistent vegetative state, severe disability, moderate disability and good recovery) after moderate-to-severe head injury.48 This strategy has also been applied to brain metastasis, LSS and LDH.2–4 Spinal stability is maintained by a variety of anatomic structures that have evolved to provide resistance against deforming 3
Downloaded from http://jnnp.bmj.com/ on January 6, 2015 - Published by group.bmj.com
Neurosurgery forces. It has been associated with excessive mechanical loads on the human spine during physical activities such as lifting and measurements of force. Since the direct in vivo measurements of spinal loads and muscle forces are invasive, investigators have decided to apply ANNs for spinal biomechanical assessment by using different possible models.54 Thus, it is argued that an understanding of biomechanical assessments using ANNs is essential in spine surgery.54–57
Potential future application of ANN in neurosurgery Arguably, ANNs are potentially more successful than traditional statistical models in predicting clinical outcomes. It is believed that complex medical diagnostic and treatment decisions will be increasingly based on ANNs for neurosurgical disorders in the next few years. Already, ANNs have been successfully applied to various areas of medicine, such as image analysis and drug development.42 60 This experience can also be used in the domain of neurosurgical disorders. Several computer-aided diagnosis projects can be performed for neurosurgery disorders, including the use of ANNs for predicting outcomes for operative versus non-operative treatment strategies for disorders and as a discriminating classifier for tasks regarding medical diagnosis for the early detection of diseases based on clinical and imaging findings. A variety of well-known instruments exist for assessing performance status or functionality in neurosurgery disorders. However, no existing standard cut-off points exist for the data acquired by these instruments for the assessment of operative versus non-operative outcomes in these patients. If standard cut-off points are subsequently made available, such instruments may be used as output variables for the development of an estimator software package. The structure and details of an ANN that does not require further training can be saved and used to create a decision-making software tool. At present, the literature lacks examples of the clinical application of ANNs in neurosurgery, although ANNs have been employed in the research arena. An improved portable computing technology diminishes the barrier that once existed. To develop and modify an ANN model as a universally available tool, it is recommended that an international centre be considered for the following issues: A. The improvement of data collection for training, such as online data captured from the process, the data division and preprocessing. The area for ANN development that harbours the most hope is associated with their ability to act as ‘data refineries’ for large and complex data sets. Their ability to make decisions based on the experience of thousands of individual records is particularly attractive. An ANN tool that correctly trained, containing the post hoc experience of the clinical features and outcome of complex sample sets, which makes for more accurate prediction, is a potentially very valuable clinical tool. B. The choice of suitable network architecture, careful selection of some internal parameters that control the optimisation methodology and stopping criteria. The development of models using ANNs is associated with the fact that there set methods do not exist for building the architecture of the network. The most common type of ANN, though, is the feed-forward back propagation MLP. C. Improving the user-friendly interface between the user and the computer. D. Model validation. External validation is necessary to obtain an accurate measure of performance outside the development case. 4
E. To evaluate ANNs as tools for performing randomised clinical trials (RCT). We know of no RCT examining the impact of ANN output on clinical actions or patient outcomes. The tool will be regularly reassessed as new information becomes available. For example, some medical applications of ANN tools and related techniques are presented in web services.61 62 With the explosion in clinical data over the past few years, we anticipate a growing need for modern machine learning techniques such as ANNs to properly make use of the information at our disposal and realise its full benefit for clinical care. To optimise the ANN software products or programs using Matlab or other toolboxes, it is recommended that an international centre be considered for the collection of real-world outcomes. Prognostic factors can be used as input variables. The number of patients in ANN models can be increased to develop and modify the model. We encourage other researchers to share their data.
Limitations of ANNs This review has limitations. One major limitation is the fact that it covers a very heterogeneous data analysis and modelling technique applied across a very heterogeneous (and idiosyncratic) set of clinical conditions. Therefore, it cannot provide an in-depth and focused synthesis of the literature. However, since this is the first attempt to summarise the evidence regarding ANNs in neurosurgery, future reviews might reasonably concentrate on one specific issue. For example, as suggested, upcoming reviews study very small and specific applications of ANNs to clinical practice (eg, “ANNs for the interpretation of radiographic images in traumatic brain injury (TBI)”) and present the various approaches and strategies used, as well as differences in the implementation of each ANN for that purpose, along with specific strengths and weaknesses. In addition, one should first note that the search strategy was limited to the key words in the title/ abstract of the publications. Thus, we might have missed some papers. Second, this work restricted the query search for articles in PubMed. Third, non-English publications were not considered in this study. We believe research regarding the application of ANNs in neurosurgery have also been published in other languages. Finally, there are possibly countless applications of ANNs in the field of neurosurgery. However, a common criticism of ANNs is that they require significant training and ‘know-how’ for real-world operations. This is not unexpected, since any learning machine is obligatorily associated with such limitations. These are imminently solvable. ANNs will never replace human experts, but they can help in screening and also be used by experts to double-check their diagnosis. Perhaps, more importantly, though, is the fact that ANNs can be used to identify variables that experts do not ‘see’, thus enhancing the diagnostic acumen of the expert.63 At present, there are no articles published that describe the use of ANN outputs in clinical practice. This is due to several factors, such as their position in the decision-making process, the lack of a support substructure, difficulties with assessment of performance, etc. It is difficult to begin to assess the true value of ANNs on clinical practice as no RCT data exist that specifically examine this. ANNs will never replace human expert decision-makers, but they can assist in double-checking the routine decision-making process.
CONCLUSION Owing to its high accuracy, ANNs can be effectively employed for diagnosis, prognosis and outcome prediction in neurosurgery. Azimi P, et al. J Neurol Neurosurg Psychiatry 2014;0:1–6. doi:10.1136/jnnp-2014-307807
Downloaded from http://jnnp.bmj.com/ on January 6, 2015 - Published by group.bmj.com
Neurosurgery Acknowledgements The authors thank the staff of the Neurosurgery Unit at Imam-Hossain Hospital, Tehran, Iran. Contributors PA was involved in the conception and design, acquisition of the data and drafting of the article. PA and AM were involved in the analysis and interpretation of data. ECB was involved in the study supervision. All authors critically revised the article and reviewed the submitted version of the manuscript. PA approved the final version of the manuscript on behalf of all authors.
24 25
26
Competing interests None. Ethics approval The research was approved by the Ethics Committee of ShahidBeheshti University of Medical Sciences, Tehran, Iran. Provenance and peer review Not commissioned; externally peer reviewed.
27 28
29
REFERENCES 1
2
3
4
5 6
7 8
9 10
11
12 13 14 15 16
17
18
19
20
21
22
23
Azimi P, Mohammadi HR. Prediction of successful ETV outcome in childhood hydrocephalus: an artificial neural networks analysis. J Neurosurg Pediatr 2014;13:426–32. Oermann EK, Kress MAS, Collins BT, et al. Predicting survival in patients with brain metastases treated with radiosurgery using artificial neural networks. Neurosurgery 2013;72:944–51. Azimi P, Benzel EC, Shahzadi S, et al. Use of artificial neural networks to predict surgical satisfaction in patients with lumbar spinal canal stenosis. J Neurosurg Spine 2014;20:300–5. Azimi P, Benzel EC, Shahzadi S, et al. The prediction of successful surgery outcome in lumbar disc herniation based on artificial neural networks. J Neurosurg Sci In press. Dreiseitl S, Ohno-Machado L. Logistic regression and artificial neural network classification models: a methodology review. J Biomed Inform 2002;35:352–9. Tu JV. Advantages and disadvantages of using artificial neural networks versus logistic regression for predicting medical outcomes. J Clin Epidemiol 1996;49:1225–31. Ramesh AN, Kambhampati C, Monson JR, et al. Artificial intelligence in medicine. Ann R Coll Surg Engl 2004;86:334–8. Arle JE, Morriss C, Wang ZJ, et al. Prediction of posterior fossa tumor type in children by means of magnetic resonance image properties, spectroscopy, and neural networks. J Neurosurg 1997;86:755–61. Bidiwala S, Pittman T. Neural network classification of pediatric posterior fossa tumors using clinical and imaging data. Pediatr Neurosurg 2004;40:8–15. Bishop JB, Szpalski M, Ananthraman SK, et al. Classification of low back pain from dynamic motion characteristics using an artificial neural network. Spine 1997;22:2991–8. Liszka-Hackzell JJ, Martin DP. Categorization and analysis of pain and activity in patients with low back pain using a neural network technique. J Med Syst 2002;26:337–47. Sari M, Gulbandilar E, Cimbiz A. Prediction of low back pain with two expert systems. J Med Syst 2010;36:1523–7. Stanley RJ, Long R. A radius of curvature-based approach to cervical spine vertebra image analysis. Biomed Sci Instrum 2001;37:385–90. Lin H. Identification of spinal deformity classification with total curvature analysis and artificial neural network. IEEE Trans Biomed Eng 2008;55:376–82. Jaremko JL, Poncet P, Ronsky J, et al. Estimation of spinal deformity in scoliosis from torso surface cross sections. Spine 2001;26:1583–91. Iglesias J, Esparza J, Scherf M. Diagnosis of human oligodendrogliomas with the help of the NeuroShell Easy Classifier neural network. Anal Quant Cytol Histol 2000;22:383–92. Kolles H, von-Wangenheim A, Vince GH, et al. Automated grading of astrocytomas based on histomorphometric analysis of Ki-67 and Feulgen stained paraffin sections. Classification results of neuronal networks and discriminant analysis. Anal Cell Pathol 1995;8:101–16. Petalidis LP, Oulas A, Backlund M, et al. Improved grading and survival prediction of human astrocytic brain tumors by artificial neural network analysis of gene expression microarray data. Mol Cancer Ther 2008;7:1013–24. Sachdeva J, Kumar V, Gupta I, et al. A dual neural network ensemble approach for multiclass brain tumor classification. Int J Numer Method Biomed Eng 2012;28:1107–20. Sieben G, Praet M, Roels H, et al. The development of a decision support system for the pathological diagnosis of human cerebral tumours based on a neural network classifier. Acta Neurochir (Wien) 1994;129:193–7. Usenius JP, Tuohimetsä S, Vainio P, et al. Automated classification of human brain tumours by neural network analysis using in vivo 1H magnetic resonance spectroscopic metabolite phenotypes. Neuroreport 1996;7:1597–600. Oğulata SN, Sahin C, Erol R. Neural network-based computer-aided diagnosis in classification of primary generalized epilepsy by EEG signals. J Med Syst 2009;33:107–12. Poptani H, Kaartinen J, Gupta RK, et al. Diagnostic assessment of brain tumours and non-neoplastic brain disorders in vivo using proton nuclear magnetic resonance
Azimi P, et al. J Neurol Neurosurg Psychiatry 2014;0:1–6. doi:10.1136/jnnp-2014-307807
30 31
32
33
34
35
36
37
38 39
40
41 42 43
44
45
46
47 48 49
50
51
spectroscopy and artificial neural networks. J Cancer Res Clin Oncol 1999;125:343–9. Jiang J, Trundle P, Ren J. Medical image analysis with artificial neural networks. Comput Med Imaging Graph 2010;34:617–31. Sharma N, Ray AK, Sharma S, et al. Segmentation and classification of medical images using texture-primitive features: application of BAM-type artificial neural network. J Med Phys 2008;33:119–26. Li YC, Liu L, Chiu WT, et al. Neural network modeling for surgical decisions on traumatic brain injury patients. Int J Med Inform 2000;57:1–9. Limonadi FM, McCartney S, Burchiel KJ. Design of an artificial neural network for diagnosis of facial pain syndromes. Stereotact Funct Neurosurg 2006;84:212–20. McCartney S, Weltin M, Burchiel KJ. Use of an artificial neural network for diagnosis of facial pain syndromes: an update. Stereotact Funct Neurosurg 2013;92:44–52. Hossen A. A neural network approach for feature extraction and discrimination between Parkinsonian tremor and essential tremor. Technol Health Care 2013;21:345–56. Sharanreddy M, Kulkarni PK. Automated EEG signal analysis for identification of epilepsy seizures and brain tumour. J Med Eng Technol 2013;37:511–19. Muniz AM, Liu W, Liu H, et al. Assessment of the effects of subthalamic stimulation in Parkinson disease patients by artificial neural network. Conf Proc IEEE Eng Med Biol Soc 2009;2009:5673–6. Seregni M, Pella A, Riboldi M, et al. Real-time tumor tracking with an artificial neural networks-based method: a feasibility study. Phys Med 2013;29:48–59. Dumont TM, Rughani AI, Tranmer BI. Prediction of symptomatic cerebral vasospasm after aneurysmal subarachnoid hemorrhage with an artificial neural network: feasibility and comparison with logistic regression models. World Neurosurg 2011;75:57–63. Lin CS, Chiu JS, Hsieh MH, et al. Predicting hypotensive episodes during spinal anesthesia with the application of artificial neural networks. Comput Methods Programs Biomed 2008;92:193–7. Sinha M, Kennedy CS, Ramundo ML. Artificial neural network predicts CT scan abnormalities in pediatric patients with closed head injury. J Trauma 2001;50:308–12. Dickey JP, Pierrynowski MR, Bednar DA, et al. Relationship between pain and vertebral motion in chronic low-back pain subjects. Clin Biomech (Bristol, Avon) 2002;17:345–52. Parsaeian M, Mohammad K, Mahmoudi M, et al. Comparison of logistic regression and artificial neural network in low back pain prediction: second national health survey. Iran J Public Health 2012;41:86–92. Viani GA, da Silva LG, Stefano EJ. Prognostic indexes for brain metastases: which is the most powerful? Int J Radiat Oncol Biol Phys 2012;83:e325–30. Chaturvedi A, Luján JL, McIntyre CC. Artificial neural network based characterization of the volume of tissue activated during deep brain stimulation. J Neural Eng 2013;10:056023. Mariak Z, Swiercz M, Krejza J, et al. Intracranial pressure processing with artificial neural networks: classification of signal properties. Acta Neurochir (Wien) 2000;142:407–11. Arle JE, Perrine K, Devinsky O, et al. Neural network analysis of preoperative variables and outcome in epilepsy surgery. J Neurosurg 1999;90:998–1004. Cross SS, Harrison RF, Kennedy RL. Introduction to neural networks. Lancet 1995;346:1075–9. DiRusso SM, Sullivan T, Holly C, et al. An artificial neural network as a model for prediction of survival in trauma patients: validation for a regional trauma area. J Trauma 2000;49:212–23. Edwards DF, Hollingsworth H, Zazulia AR, et al. Artificial neural networks improve the prediction of mortality in intracerebral hemorrhage. Neurology 1999;53:351–7. Eftekhar B, Mohammad K, Ardebili HE, et al. Comparison of artificial neural network and logistic regression models for prediction of mortality in head trauma based on initial clinical data. BMC Med Inform Decis Mak 2005;5:3. Hsu MH, Li YC, Chiu WT, et al. Outcome prediction after moderate and severe head injury using an artificial neural network. Stud Health Technol Inform 2005;116:241–5. Izenberg SD, Williams MD, Luterman A. Prediction of trauma mortality using a neural network. Am Surg 1997;63:275–81. Rughani AI, Dumont TM, Lu Z, et al. Use of an artificial neural network to predict head injury outcome. J Neurosurg 2010;113:585–90. Shi HY, Hwang SL, Lee KT, et al. In-hospital mortality after traumatic brain injury surgery: a nationwide population-based comparison of mortality predictors used in artificial neural network and logistic regression models. J Neurosurg 2013;118:746–52. Lang EW, Pitts LH, Damron SL, et al. Outcome after severe head injury: an analysis of prediction based upon comparison of neural network versus logistic regression analysis. Neurol Res 1997;19:274–80. Erol FS, Uysal H, Ergün U, et al. Prediction of minor head injured patients using logistic regression and MLP neural network. J Med Syst 2005;29:205–15.
5
Downloaded from http://jnnp.bmj.com/ on January 6, 2015 - Published by group.bmj.com
Neurosurgery 52 53
54
55
56
6
Davis DP, Peay J, Sise MJ, et al. The impact of prehospital endotracheal intubation on outcome in moderate to severe traumatic brain injury. J Trauma 2005;58:933–9. Abouzari M, Rashidi A, Zandi-Toghani M, et al. Chronic subdural hematoma outcome prediction using logistic regression and an artificial neural network. Neurosurg Rev 2009;32:479–84. Arjmand N, Ekrami O, Shirazi-adl A, et al. relative performances of artificial neural network and regression mapping tools in evaluation of spinal loads and muscle forces during static lifting. J Biomech 2013;46:1454–62. Wang L, Buchanan TS. Prediction of joint moments using a neural network model of muscle activations from EMG signals. IEEE Trans Neural Syst Rehabil Eng 2002;10:30–7. Nussbaum MA, Chaffin DB. Pattern classification reveals intersubject group differences in lumbar muscle recruitment during static loading. Clin Biomech (Bristol, Avon) 1997;12:97–106.
57
58 59
60 61 62 63
Amaritsakul Y, Chao CK, Lin J. Multiobjective optimization design of spinal pedicle screws using neural networks and genetic algorithm: mathematical models and mechanical validation. Comput Math Methods Med 2013;2013:462875. Mann NH, Brown MD, Hertz DB, et al. Initial-impression diagnosis using low-back pain patient pain drawings. Spine 1993;18:41–53. Ongphiphadhanakul B, Rajatanavin R, Chailurkit L, et al. Prediction of low bone mineral density in postmenopausal women by artificial neural network model compared to logistic regression model. J Med Assoc Thai 1997;80:508–15. http://library.thinkquest.org/C007395/tqweb/history.html, 2013. http://neuroxl.com/applications/medicine/neural-networks-in-medicine/index.htm, 2013. http://www.phil.gu.se/ann/annworld.html, 2013. Rughani AI, Dumont TM, Tranmer BI. Editorial: predicting surgical satisfaction using artificial neural networks. J Neurosurg Spine 2014;20:298–9.
Azimi P, et al. J Neurol Neurosurg Psychiatry 2014;0:1–6. doi:10.1136/jnnp-2014-307807
Downloaded from http://jnnp.bmj.com/ on January 6, 2015 - Published by group.bmj.com
Artificial neural networks in neurosurgery Parisa Azimi, Hasan Reza Mohammadi, Edward C Benzel, Sohrab Shahzadi, Shirzad Azhari and Ali Montazeri J Neurol Neurosurg Psychiatry published online July 1, 2014
Updated information and services can be found at: http://jnnp.bmj.com/content/early/2014/07/01/jnnp-2014-307807
Supplementary Material
Supplementary material can be found at: http://jnnp.bmj.com/content/suppl/2014/07/01/jnnp-2014-307807.DC1. html These include:
References
This article cites 59 articles, 1 of which you can access for free at: http://jnnp.bmj.com/content/early/2014/07/01/jnnp-2014-307807#BIBL
Email alerting service
Receive free email alerts when new articles cite this article. Sign up in the box at the top right corner of the online article.
Topic Collections
Articles on similar topics can be found in the following collections Epilepsy and seizures (738) CNS cancer (168) Neurooncology (221) Pain (neurology) (656) Musculoskeletal syndromes (489) Hydrocephalus (127)
Notes
To request permissions go to: http://group.bmj.com/group/rights-licensing/permissions To order reprints go to: http://journals.bmj.com/cgi/reprintform To subscribe to BMJ go to: http://group.bmj.com/subscribe/
