eO
RIGINAL
ARTICLE
Use of Artificial Neural Networks to Predict Recurrent Lumbar Disk Herniation Parisa Azimi, MD,* Hassan R. Mohammadi, MD,* Edward C. Benzel, MD,w Sohrab Shahzadi, MD,* and Shirzad Azhari, MD*
Background: The aim of this study was to develop an artificial neural network (ANN) model to predict recurrent lumbar disk herniation (LDH). Methods: An ANN model and a logistic regression model were used to predict recurrent LDH. The age, sex, duration of symptoms, smoking status, recurrent LDH, level of herniation, type of herniation, sports activity; occupational lifting, occupational driving, duration of symptoms, visual analog scale (VAS), the Zung Depression Scale (ZDS), and the Japanese Orthopaedic Association (JOA) Score, were determined as the input variables for the established ANN model. The Macnab classification, VAS, and JOA were used for outcome assessment. ANNs on data from LDH patients, who underwent surgery, were trained to predict LDH using several input variables. The patients were divided into a recurrent LDH group (R group) and a primary LDH group (P group). Sensitivity analysis was applied to identify the relevant variables. The receiver-operating characteristic curve, accuracy rate of predicting, and Hosmer-Lemeshow statistics were considered for evaluating the 2 models. Results: A total of 402 patients were categorized into training, testing, and validation data sets consisting of 201, 101, and 100 cases, respectively. The recurrence rate was 8.7%, and the median time to recurrence was 26.2 months (SD = 4 mo). The VAS of leg/back pain and JOA were improved at 1-year follow-up (P < 0.05) and no significant difference was observed between the 2 groups. Surgical successful outcome was categorized as: excellent, 31.1%; good, 44.3%; fair, 18.9%; and poor, 5.7% at 1-year follow-up. Compared with the logistic regression model, the ANN model was associated with superior results: accuracy rate, 94.1%; Hosmer-Lemeshow statistic, 40.2%; and area under the curve, 0.83% of patients. Conclusion: The findings show that an ANNs can be used to predict the diagnostic statues of recurrent and nonrecurrent group of LDH patients before the first or index microdiscectomy.
Received for publication August 2, 2014; accepted October 13, 2014. From the *Department of Neurosurgery, Shahid Beheshti University of Medical Sciences, Tehran, Iran; and wDepartment of Neurosurgery, Cleveland Clinic Foundation, Cleveland, OH. The authors declare no conflict of interest. Reprints: Parisa Azimi, MD, Functional Neurosurgery Research Center of Shohada Tajrish Hospital, Shahid Beheshti University of Medical Sciences, Sharadari Street, Tajrish Square, Tehran 1989934148, Iran (e-mail: [email protected]). Copyright r 2014 Wolters Kluwer Health, Inc. All rights reserved.
J Spinal Disord Tech
Volume 28, Number 3, April 2015
Key Words: recurrent LDH, predict, ANN (J Spinal Disord Tech 2015;28:E161–E165)
L
umbar disk herniation (LDH) is a common disorder among adults, causing low back and/or leg pain. Symptoms can affect the lower back, buttocks, thigh, anal/genital region, and may emit into the foot and/or toe. Discectomy can be used to treat LDH and the majority of patients who undergo discectomy benefit from the procedure. However, recurrent LDH is a common cause of poor outcomes and a common indication for revision surgery. The reported recurrence rate incidence ranges from 0.5% to 21%.1–3 Many risk factors are associated with recurrence LDH such as male sex, trauma, annular competence, type of herniation, advanced age, high body mass index, protrusion type, smoking, and occupational lifting.4–8 Many studies have been conducted to determine the other factors that may increase the reherniation risk.9 However, many controversies exist regarding risk factors for prediction of recurrent LDH. Thus, if one could establish a marker to predict recurrent LDH in advance of the index first or index operation, clinical decision making would be enhanced. Medical software systems have been applied to develop prediction models to evaluate, diagnose, and treat of patients. These include logistic regression (LR) and artificial neural networks (ANNs) or neural networks. LR is a traditional predictive measure. ANNs are information processing algorithms that are used to establish the complex relationships between input and output variables by emulating information processing mechanisms of biological nervous systems. ANNs learn like people learn, by “observing” the sampling of variables. The learning process includes the assessment of feeding sample (observed) data through ANNs that modifies their structure based on the training algorithms. The trained network can then predict or estimate outputs for new input data. This ability makes ANNs useful tools for a decision maker concerned prospects. The detailed for the concept of ANNs has been published elsewhere.10–12 Relationships between prognostic factors and prediction of recurrent in LDH patients have not been previously investigated using ANNs. Hence, the aim of the study presented here was to develop an ANN model based on the age, visual analog scale (VAS) of leg/back pain, www.jspinaldisorders.com |
E161
J Spinal Disord Tech
Azimi et al
recurrent LDH, level of herniation, type of herniation, current smoking status, sports activity, occupational lifting exposure, occupational driving exposure, duration of symptoms, the Japanese Orthopaedic Association (JOA) Score, and Zung Depression Scale (ZDS). It also sought to determine whether ANNs perform better at predicting recurrent LDH, compared with LR in LDH patients.
METHODS Patients and Data Collection This study included a consecutive sampling of 402 patients (186 females and 216 males) who underwent microdiscectomy via a conventional open discectomy by experienced surgeons for LDH between February 2008 and May 2012 in a large teaching hospital in Tehran, Iran. Of these patients, 35 underwent reoperation or nonoperative treatment for recurrent LDH. The diagnosis of LDH was performed based on clinical symptoms, neurological examinations, and imaging studies—including plain radiography, computed tomography, and magnetic resonance imaging of the lumbar spine. All patients had the typical symptoms of LDH, such as back pain and pain in one or both legs. In all cases, >1 spine surgeon confirmed the diagnosis. The disk level(s) were localized with magnetic resonance imaging or computed tomography. Indications for microdiscectomy were intractable pain that had not responded to conservative treatment for 6 to 8 weeks. There were no restrictions on patient selection with regard to types of LDH, age, or other characteristics. The patients were divided into a recurrent LDH group (R group) and a primary LDH group (P group). Recurrent LDH was defined as a disk herniation at the same level, regardless of ipsilateral or contralateral herniation, in a patient who experienced a pain-free interval of at least 6 months after prior surgery.8 All patients completed a demographic questionnaire that assessed age, sex and body weight, walking distance (m), duration of symptoms (mo), a 100 mm VAS to measure both leg pain and back pain,13 the JOA Score for assessing low back pain (total 29 points, a high JOA Score indicates a better clinical outcome),14 level of herniation, type of herniation, current smoking, sports activity, occupational lifting profile, occupational driving profile, and the ZDS (a 20-item as a total score ranging from 20 to 80—a patient with a score of Z55 would indicate depression) before surgery.15 The exclusion criteria were spinal anomalies, spondylolisthesis, and polyneuropathy. The Macnab classification system was used for outcome assessment, with 4 success categories used to portray outcome: excellent, good, fair, and poor at 1-year follow-up assessment.16
ANN Model The ANN was created via standard methodology, using the Statistical Package for the Social Sciences (SPSS) software program. In this study, a multilayer perceptron (MLP) model was selected. MLP-ANNs consist of a series of nodes arranged in 3 layers: an input layer, a hidden layer, and an output layer. The MLP-
E162 | www.jspinaldisorders.com
Volume 28, Number 3, April 2015
ANN used observed data consisting of inputs (age, VAS of leg/back pain, JOA, ZDS, duration of symptoms, recurrent LDH, level of herniation, type of herniation, current smoking, sports activity; occupational lifting profiles, occupational driving profile, duration of symptoms) and outputs (recurrent and nonrecurrent LDH) to define (learn) the complex relationship between inputs and outputs. Patients were divided by a 2:1:1 ratio to create training, testing, and validation. Once the MLPANN was trained, it was then used to estimate (predict) results (outputs) from new sets of input data.17,18
LR Traditional statistical analysis of parameter significance was performed with standard LR on the same data set used for the ANN.
Statistical Analysis All statistical analyses were performed using the PASW Statistics 18, version 18 (SPSS Inc., 2009, Chicago, IL). Student t tests were performed for continuous variables, whereas w2 analyses and Fisher exact tests were used for categorical variables depending on sample size. For each individual parameter and for comparing the ANN model and LR model, receiver-operating characteristic curves were created and used to calculate specificities, positive predictive value, and negative predictive value at 95% sensitivity. The area under the curve (AUC) was calculated from the receiver-operating characteristic analysis to evaluate of the discrimination capability. For each pair of ANN and LR models (trained and tested on the same data sets), AUC, Hosmer-Lemeshow statistics for assessment of goodness-of-fit of the models, and accuracy rate were calculated and compared by t tests for the validation group (n = 100 patients).12
Ethics The research project was approved by the Ethics Committee of Shahid Beheshti University of Medical Sciences, Tehran, Iran.
RESULTS The demographics of the LDH patients and their scores on the JOA and the ZDS are shown in Table 1. A total of 402 (186 male, 216 female; median age, 49.6 ± 10.4 y) patients were divided into training (n = 201), testing (n = 101), and validation (n = 100) data sets. Interrelationships between predictor variables (input nodes), hidden variables (8 of them in 1 hidden layer), and recurrent LDH/nonrecurrent LDH (output nodes) are illustrated in Figure 1. The results of comparing the ANN model and LR model and individual parameters are shown in Tables 2 and 3. The sensitivity analysis determined that no significant statistically important variable was selected by the ANN. Compared with the LR model the ANN model had a better accuracy rate in 94.1% of patients, a better Hosmer-Lemeshow statistic in 40.2% of patients, and a better AUC in 0.83% of patients. Copyright
r
2014 Wolters Kluwer Health, Inc. All rights reserved.
J Spinal Disord Tech
Volume 28, Number 3, April 2015
Predicting Recurrent LDH Using ANN
TABLE 1. Demographic Data and Preoperative Status of Patients With Lumbar Disk Herniation [n = 402 (186 Male, 216 Female)] Mean (SD) Characteristics Age (y) Range Sex [male; n (%)] Smoking [n (%)] Body weight (kg) Body mass index (BMI) Sports activity [n (%)] Occupational lifting [n (%)] Occupational driving [n (%)] Symptoms Duration of symptoms (mo) Range VAS of leg pain (mm) Range VAS of back pain (mm) Range JOA score Range ZDS Range Level of herniation [n (%)] L1–L2 L2–L3 L3–L4 L4–L5 L5–S1 Type of herniation [n (%)] Sequestration Transligamentous extrusion Subligamentous extrusion Protrusion
P Group* (n = 367)
R Group* (n = 35)
49.2 (10.6) 17–79 172 (46.9) 121 (32.9) 83.1 (10.7) 25.9 (5.1) 63 (17.2) 66 (17.9) 32 (8.7)
50.1 (10.2) 22–82 14 (40.0) 18 (51.4) 86.5 (9.8) 26.5 (4.9) 5 (14.3) 7 (20.0) 4 (11.4)
14.3 (11.7) 1–24 57.3 (19.2) 14–100 51.4 (23.9) 19–100 6.1 (3.4) 3 to 14 42.7 (9.5) 21–69
15.3 (10.4) 1–30 58.4 (18.1) 15–100 52.1 (24.1) 21–100 6.3 (3.5) 3 to 14 44.1 (10.1) 23–69
P# 0.675 0.078 0.062 0.171 0.165 0.276 0.268 0.311 0.560 0.261 0.277 0.145 0.187 0.123
4 13 43 178 130
(1.1) (3.5) (11.6) (48.5) (35.3)
0 1 (2.8) 3 (8.6) 19 (54.3) 12 (34.3)
106 129 95 37
(28.8) (35.1) (25.9) (10.1)
13 9 7 6
0.131 (37.1) (25.7) (20.0) (17.1)
Values are mean (SD) and n (%). *R group, a recurrent LDH group; P group, a primary LDH group. #Statistical significance of difference between groups is tested by Student t tests for continuous variables, whereas w2 analyses and Fisher exact tests were used for categorical variables depending on sample size. LDH indicates lumbar disk herniation; JOA, Japanese Orthopaedic Association; VAS, visual analog scale; ZDS, Zung Depression Scale.
The median operative time per level was about 70 minutes (range, 50–90 min). The recurrence rate was 8.7% (35 of 402 patients), and the median time to recurrence was 26.2 months (SD = 4.1 mo). The median VAS scale for leg pain improved from 4.28 to 0.89 (P < 0.05). The median VAS scale for back pain improved from 4.31 to 1.2 (P < 0.05) and the median JOA improved from 21.5 to 16.4 (P < 0.05). No significant difference was observed between the 2 groups at 1-year follow-up. The surgical outcome for the 2 groups was: excellent, 31.1%; good, 44.3%; fair, 18.9%; and poor, 5.7% at 1-year follow-up. No significant difference was observed between the 2 groups based on surgical successful outcome.
DISCUSSION Individual parameters such as VAS of leg/back pain, JOA, ZDS, duration of symptoms, surgical recurrence, type Copyright
r
2014 Wolters Kluwer Health, Inc. All rights reserved.
FIGURE 1. Artificial neural network output diagram with insets for each layer. Output figure generated by PASW Statistics 18, version 18 (SPSS Inc., 2009, Chicago, IL). Input layer: bias (input layer bias); JOA (the Japanese Orthopaedic Association Score); ZDS (Zung Depression Scale); S (smoking); VAS.L (visual analog scale of leg pain); gender; VAS.B (visual analog scale of back pain); age; OL (occupational lifting); OD (occupational driving); SA (sports activity); BMI (body mass index); LH (level of hearniation); TH (type of herniation); DR [duration of symptoms (mo)]. Hidden layer: H (1:1), H (1:2), H (1:3), H (1:4), H (1:5), H (1:6), H (1:7), and H (1:8); bias (hidden layer bias). Output layer: recurrent lumbar disk herniation (LDH) and primary LDH.
of herniation, smoking, and age can assist to better understanding recurrence predictors. However, the findings from the study presented here indicate that the combination of these parameters, in an ANN model, could be applied to estimate risk of recurrent disk herniation with a high degree of accuracy. The full structure details of the resulting ANN can be saved and then be applied to new or subsequent www.jspinaldisorders.com |
E163
J Spinal Disord Tech
Azimi et al
Volume 28, Number 3, April 2015
TABLE 2. Comparison of the AUC and Predictive Values of ANN and LR Models, and Individual Parameters to Predict Recurrent Lumbar Disk Herniation (n = 402) Parameters/Model ANN LR JOA ZDS S VAS.L Sex VAS.B Age OL OD SA BMI LH TH DR
AUC (%)
P*
Specificity (%)w
PPV (%)w
NPV (%)w
0.84 0.77 0.57 0.56 0.54 0.53 0.52 0.51 0.51 0.49 0.48 0.47 0.47 0.46 0.45 0.44
0.001 0.001 0.24 0.28 0.29 0.32 0.34 0.37 0.41 0.43 0.52 0.53 0.53 0.67 0.67 0.68
46 34 22 19 16 11 9 6 6 5 5 5 4 4 3 3
69 64 57 55 55 54 54 53 53 53 52 52 51 51 51 50
88 82 58 56 54 49 44 42 31 21 20 19 19 18 17 17
*Asymptotic significance on receiver-operating characteristic curve analysis. wSpecificity and predictive values at 95% sensitivity. ANN indicates artificial neural network; AUC, area under the receiver-operating characteristic; BMI, body mass index; DR, duration of symptoms (mo); JOA, the Japanese Orthopaedic Association Score; LH, level of herniation; LR, logistic regression; NPV, negative predictive value; OD, occupational driving; OL, occupational lifting; PPV, positive predictive value; S, smoking; SA, sports activity; TH, type of herniation; VAS.B, visual analog scale of back pain; VAS.L, visual analog scale of leg pain; ZDS, Zung Depression Scale.
cases with no need for further training. Although ANNs show promise, the study sample was small, and therefore the technique will need to be repeated with larger, multicenter data sets and external validation to convincingly demonstrate its validity and predictive power. To date, there are no reports describing ANNs as a predictor for recurrent LDH. However, risk factors for recurrent LDH have been reported to include male sex,5,8 degenerated disk,5 smoking,7,8 trauma,8 annular competence, type of herniation,4 advanced age, high body mass index, protrusion type,6 and occupational lifting.7 Many studies have been conducted to determine the additional factors that may increase reherniation risk. These factors include the type of LDH, the amount of disk material removed, alcohol consumption, and the duration of restricted activities. Few are agreed upon by all. For the analysis of risk factors in recurrent LDH, large-scale multicenter prospective studies will be required in the future.9 Herein, we have presented the first attempt at ANN prediction of recurrent LDH. The results obtained from the sensitivity analysis clearly suggest that an ANN model could be used confidently to predict recurrent LDH. This
TABLE 3. Comparison of ANN and LR Models for Predict Successful Recurrent Lumbar Disk Herniation (n = 100*) Accuracy rate (%) AUC H-L statistics
ANN (95% CI)
LR (95% CI)
P
94.1 (93.2–97.1) 0.83 (0.86–0.88) 40.2 (37.3–49.6)
86.4 (84.5–91.4) 0.76 (0.74–0.78) 55.4 (50.4–59.6)
< 0.001 < 0.001 < 0.001
*Patients of validation group. ANN indicates artificial neural network; AUC, area under the receiver-operating characteristic; CI, confidence interval; H-L statistics, Hosmer-Lemeshow statistics; LR, logistic regression.
E164 | www.jspinaldisorders.com
simple, elegant method may be useful to reduce disk herniation recurrence. The discrimination of R group and P group from each other may be able to substantially help in therapy planning. Using an increased number of patients and additional parameters for input layer of ANN, the accuracy of the ANN model is enhanced regarding the optimization of the clinical decision-making process.19 Hence, the ANN model introduced in this study appears to be an acceptable test to predict recurrent LDH. There were some limitations in our study. First, we likely did not identify all possibly significant variables to predict recurrent LDH. Future studies of this model may consider the effect of a more detailed database that contains more input variables (such as clinical and imaging details), as well as more detailed result data. Second, the number of patients was relatively small. Third, further randomized clinical trials are needed to establish benefit and to confirm these findings in ANN model. Finally, ANNs will never replace human expert decision makers, but they can assist in double-checking and enhancing the routine decisionmaking process.
CONCLUSION The findings presented here show that ANNs can be used to effectively assist with the prediction of recurrent LDH before the performance of the initial or index microdiscectomy. ACKNOWLEDGMENTS The authors thank the staff of the Neurosurgery Unit at Imam-Hossain Hospital, Tehran, Iran. Copyright
r
2014 Wolters Kluwer Health, Inc. All rights reserved.
J Spinal Disord Tech
Volume 28, Number 3, April 2015
REFERENCES 1. Ahsan K, Najmus S, Hossain A, et al. Discectomy for primary and recurrent prolapse of lumbar intervertebral discs. J Orthop Surg (Hong Kong). 2012;20:7–10. 2. Aizawa T, Ozawa H, Kusakabe T, et al. Reoperation for recurrent lumbar disc herniation: a study over a 20-year period in a Japanese population. J Orthop Sci. 2012;17:107–113. 3. Ambrossi GL, McGirt MJ, Sciubba DM, et al. Recurrent lumbar disc herniation after single-level lumbar discectomy: incidence and health care cost analysis. Neurosurgery. 2009;65:574–578. 4. Carragee EJ, Han MY, Suen PW, et al. Clinical outcomes after lumbar discectomy for sciatica: the effects of fragment type and anular competence. J Bone Joint Surg Am. 2003;85:102–108. 5. Cinotti G, Roysam GS, Eisenstein SM, et al. Ipsilateral recurrent lumbar disc herniation: prospective, controlled study. J Bone Joint Surg Br. 1998;80:825–832. 6. Kim JM, Lee SH, Ahn Y, et al. Recurrence after successful percutaneous endoscopic lumbar discectomy. Minim Invasive Neurosurg. 2007;50:82–85. 7. Miwa S, Yokogawa A, Kobayashi T, et al. Risk factors of recurrent lumbar disc herniation: a single center study and review of the literature. J Spinal Disord Tech. 2013. [Epub ahead of print]. 8. Suk KS, Lee HM, Moon SH, et al. Recurrent lumbar disc herniation: results of operative management. Spine. 2001;26: 672–676. 9. Shin BJ. Risk factors for recurrent lumbar disc herniations. Asian Spine J. 2014;8:211–215.
Copyright
r
2014 Wolters Kluwer Health, Inc. All rights reserved.
Predicting Recurrent LDH Using ANN
10. 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. 2013. [Epub ahead of print]. 11. 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;20:300–305. 12. Azimi P, Mohammadi HR. Predicting endoscopic third ventriculostomy success in childhood hydrocephalus: an artificial neural networks analysis. J Neurosurg Pediatr. 2014;13:426–432. 13. Price DD, McGrath PA, Rafii A, et al. The validation of visual analogue scales as ratio scale measures for chronic and experimental pain. Pain. 1983;17:45–56. 14. Azimi P, Mohammadi HR, Montazeri A. An outcome measure of functionality and pain in patients with lumbar disc herniation: a validation study of the Japanese Orthopedic Association (JOA) score. J Orthop Sci. 2012;17:341–345. 15. Zung WW, Richards CB, Short MJ. Self-rating depression scale in an outpatient clinic. Further validation of the SDS. Arch Gen Psychiatry. 1965;13:508–515. 16. Macnab I. Chapter 14: pain and disability in degenerative disc disease. Clin Neurosurg. 1973;20:193–196. 17. Cross SS, Harrison RF, Kennedy RL. Introduction to neural networks. Lancet. 1995;346:1075–1079. 18. Rughani AI, Dumont TM, Lu Z, et al. Use of an artificial neural network to predict head injury outcome. J Neurosurg. 2010;113: 585–590. 19. Azimi P, Mohammadi HR, Benzel EC, et al. Artificial neural networks in neurosurgery. J Neurol Neurosurg Psychiatry. 2015;86: 251–256.
www.jspinaldisorders.com |
E165
RIGINAL
ARTICLE
Use of Artificial Neural Networks to Predict Recurrent Lumbar Disk Herniation Parisa Azimi, MD,* Hassan R. Mohammadi, MD,* Edward C. Benzel, MD,w Sohrab Shahzadi, MD,* and Shirzad Azhari, MD*
Background: The aim of this study was to develop an artificial neural network (ANN) model to predict recurrent lumbar disk herniation (LDH). Methods: An ANN model and a logistic regression model were used to predict recurrent LDH. The age, sex, duration of symptoms, smoking status, recurrent LDH, level of herniation, type of herniation, sports activity; occupational lifting, occupational driving, duration of symptoms, visual analog scale (VAS), the Zung Depression Scale (ZDS), and the Japanese Orthopaedic Association (JOA) Score, were determined as the input variables for the established ANN model. The Macnab classification, VAS, and JOA were used for outcome assessment. ANNs on data from LDH patients, who underwent surgery, were trained to predict LDH using several input variables. The patients were divided into a recurrent LDH group (R group) and a primary LDH group (P group). Sensitivity analysis was applied to identify the relevant variables. The receiver-operating characteristic curve, accuracy rate of predicting, and Hosmer-Lemeshow statistics were considered for evaluating the 2 models. Results: A total of 402 patients were categorized into training, testing, and validation data sets consisting of 201, 101, and 100 cases, respectively. The recurrence rate was 8.7%, and the median time to recurrence was 26.2 months (SD = 4 mo). The VAS of leg/back pain and JOA were improved at 1-year follow-up (P < 0.05) and no significant difference was observed between the 2 groups. Surgical successful outcome was categorized as: excellent, 31.1%; good, 44.3%; fair, 18.9%; and poor, 5.7% at 1-year follow-up. Compared with the logistic regression model, the ANN model was associated with superior results: accuracy rate, 94.1%; Hosmer-Lemeshow statistic, 40.2%; and area under the curve, 0.83% of patients. Conclusion: The findings show that an ANNs can be used to predict the diagnostic statues of recurrent and nonrecurrent group of LDH patients before the first or index microdiscectomy.
Received for publication August 2, 2014; accepted October 13, 2014. From the *Department of Neurosurgery, Shahid Beheshti University of Medical Sciences, Tehran, Iran; and wDepartment of Neurosurgery, Cleveland Clinic Foundation, Cleveland, OH. The authors declare no conflict of interest. Reprints: Parisa Azimi, MD, Functional Neurosurgery Research Center of Shohada Tajrish Hospital, Shahid Beheshti University of Medical Sciences, Sharadari Street, Tajrish Square, Tehran 1989934148, Iran (e-mail: [email protected]). Copyright r 2014 Wolters Kluwer Health, Inc. All rights reserved.
J Spinal Disord Tech
Volume 28, Number 3, April 2015
Key Words: recurrent LDH, predict, ANN (J Spinal Disord Tech 2015;28:E161–E165)
L
umbar disk herniation (LDH) is a common disorder among adults, causing low back and/or leg pain. Symptoms can affect the lower back, buttocks, thigh, anal/genital region, and may emit into the foot and/or toe. Discectomy can be used to treat LDH and the majority of patients who undergo discectomy benefit from the procedure. However, recurrent LDH is a common cause of poor outcomes and a common indication for revision surgery. The reported recurrence rate incidence ranges from 0.5% to 21%.1–3 Many risk factors are associated with recurrence LDH such as male sex, trauma, annular competence, type of herniation, advanced age, high body mass index, protrusion type, smoking, and occupational lifting.4–8 Many studies have been conducted to determine the other factors that may increase the reherniation risk.9 However, many controversies exist regarding risk factors for prediction of recurrent LDH. Thus, if one could establish a marker to predict recurrent LDH in advance of the index first or index operation, clinical decision making would be enhanced. Medical software systems have been applied to develop prediction models to evaluate, diagnose, and treat of patients. These include logistic regression (LR) and artificial neural networks (ANNs) or neural networks. LR is a traditional predictive measure. ANNs are information processing algorithms that are used to establish the complex relationships between input and output variables by emulating information processing mechanisms of biological nervous systems. ANNs learn like people learn, by “observing” the sampling of variables. The learning process includes the assessment of feeding sample (observed) data through ANNs that modifies their structure based on the training algorithms. The trained network can then predict or estimate outputs for new input data. This ability makes ANNs useful tools for a decision maker concerned prospects. The detailed for the concept of ANNs has been published elsewhere.10–12 Relationships between prognostic factors and prediction of recurrent in LDH patients have not been previously investigated using ANNs. Hence, the aim of the study presented here was to develop an ANN model based on the age, visual analog scale (VAS) of leg/back pain, www.jspinaldisorders.com |
E161
J Spinal Disord Tech
Azimi et al
recurrent LDH, level of herniation, type of herniation, current smoking status, sports activity, occupational lifting exposure, occupational driving exposure, duration of symptoms, the Japanese Orthopaedic Association (JOA) Score, and Zung Depression Scale (ZDS). It also sought to determine whether ANNs perform better at predicting recurrent LDH, compared with LR in LDH patients.
METHODS Patients and Data Collection This study included a consecutive sampling of 402 patients (186 females and 216 males) who underwent microdiscectomy via a conventional open discectomy by experienced surgeons for LDH between February 2008 and May 2012 in a large teaching hospital in Tehran, Iran. Of these patients, 35 underwent reoperation or nonoperative treatment for recurrent LDH. The diagnosis of LDH was performed based on clinical symptoms, neurological examinations, and imaging studies—including plain radiography, computed tomography, and magnetic resonance imaging of the lumbar spine. All patients had the typical symptoms of LDH, such as back pain and pain in one or both legs. In all cases, >1 spine surgeon confirmed the diagnosis. The disk level(s) were localized with magnetic resonance imaging or computed tomography. Indications for microdiscectomy were intractable pain that had not responded to conservative treatment for 6 to 8 weeks. There were no restrictions on patient selection with regard to types of LDH, age, or other characteristics. The patients were divided into a recurrent LDH group (R group) and a primary LDH group (P group). Recurrent LDH was defined as a disk herniation at the same level, regardless of ipsilateral or contralateral herniation, in a patient who experienced a pain-free interval of at least 6 months after prior surgery.8 All patients completed a demographic questionnaire that assessed age, sex and body weight, walking distance (m), duration of symptoms (mo), a 100 mm VAS to measure both leg pain and back pain,13 the JOA Score for assessing low back pain (total 29 points, a high JOA Score indicates a better clinical outcome),14 level of herniation, type of herniation, current smoking, sports activity, occupational lifting profile, occupational driving profile, and the ZDS (a 20-item as a total score ranging from 20 to 80—a patient with a score of Z55 would indicate depression) before surgery.15 The exclusion criteria were spinal anomalies, spondylolisthesis, and polyneuropathy. The Macnab classification system was used for outcome assessment, with 4 success categories used to portray outcome: excellent, good, fair, and poor at 1-year follow-up assessment.16
ANN Model The ANN was created via standard methodology, using the Statistical Package for the Social Sciences (SPSS) software program. In this study, a multilayer perceptron (MLP) model was selected. MLP-ANNs consist of a series of nodes arranged in 3 layers: an input layer, a hidden layer, and an output layer. The MLP-
E162 | www.jspinaldisorders.com
Volume 28, Number 3, April 2015
ANN used observed data consisting of inputs (age, VAS of leg/back pain, JOA, ZDS, duration of symptoms, recurrent LDH, level of herniation, type of herniation, current smoking, sports activity; occupational lifting profiles, occupational driving profile, duration of symptoms) and outputs (recurrent and nonrecurrent LDH) to define (learn) the complex relationship between inputs and outputs. Patients were divided by a 2:1:1 ratio to create training, testing, and validation. Once the MLPANN was trained, it was then used to estimate (predict) results (outputs) from new sets of input data.17,18
LR Traditional statistical analysis of parameter significance was performed with standard LR on the same data set used for the ANN.
Statistical Analysis All statistical analyses were performed using the PASW Statistics 18, version 18 (SPSS Inc., 2009, Chicago, IL). Student t tests were performed for continuous variables, whereas w2 analyses and Fisher exact tests were used for categorical variables depending on sample size. For each individual parameter and for comparing the ANN model and LR model, receiver-operating characteristic curves were created and used to calculate specificities, positive predictive value, and negative predictive value at 95% sensitivity. The area under the curve (AUC) was calculated from the receiver-operating characteristic analysis to evaluate of the discrimination capability. For each pair of ANN and LR models (trained and tested on the same data sets), AUC, Hosmer-Lemeshow statistics for assessment of goodness-of-fit of the models, and accuracy rate were calculated and compared by t tests for the validation group (n = 100 patients).12
Ethics The research project was approved by the Ethics Committee of Shahid Beheshti University of Medical Sciences, Tehran, Iran.
RESULTS The demographics of the LDH patients and their scores on the JOA and the ZDS are shown in Table 1. A total of 402 (186 male, 216 female; median age, 49.6 ± 10.4 y) patients were divided into training (n = 201), testing (n = 101), and validation (n = 100) data sets. Interrelationships between predictor variables (input nodes), hidden variables (8 of them in 1 hidden layer), and recurrent LDH/nonrecurrent LDH (output nodes) are illustrated in Figure 1. The results of comparing the ANN model and LR model and individual parameters are shown in Tables 2 and 3. The sensitivity analysis determined that no significant statistically important variable was selected by the ANN. Compared with the LR model the ANN model had a better accuracy rate in 94.1% of patients, a better Hosmer-Lemeshow statistic in 40.2% of patients, and a better AUC in 0.83% of patients. Copyright
r
2014 Wolters Kluwer Health, Inc. All rights reserved.
J Spinal Disord Tech
Volume 28, Number 3, April 2015
Predicting Recurrent LDH Using ANN
TABLE 1. Demographic Data and Preoperative Status of Patients With Lumbar Disk Herniation [n = 402 (186 Male, 216 Female)] Mean (SD) Characteristics Age (y) Range Sex [male; n (%)] Smoking [n (%)] Body weight (kg) Body mass index (BMI) Sports activity [n (%)] Occupational lifting [n (%)] Occupational driving [n (%)] Symptoms Duration of symptoms (mo) Range VAS of leg pain (mm) Range VAS of back pain (mm) Range JOA score Range ZDS Range Level of herniation [n (%)] L1–L2 L2–L3 L3–L4 L4–L5 L5–S1 Type of herniation [n (%)] Sequestration Transligamentous extrusion Subligamentous extrusion Protrusion
P Group* (n = 367)
R Group* (n = 35)
49.2 (10.6) 17–79 172 (46.9) 121 (32.9) 83.1 (10.7) 25.9 (5.1) 63 (17.2) 66 (17.9) 32 (8.7)
50.1 (10.2) 22–82 14 (40.0) 18 (51.4) 86.5 (9.8) 26.5 (4.9) 5 (14.3) 7 (20.0) 4 (11.4)
14.3 (11.7) 1–24 57.3 (19.2) 14–100 51.4 (23.9) 19–100 6.1 (3.4) 3 to 14 42.7 (9.5) 21–69
15.3 (10.4) 1–30 58.4 (18.1) 15–100 52.1 (24.1) 21–100 6.3 (3.5) 3 to 14 44.1 (10.1) 23–69
P# 0.675 0.078 0.062 0.171 0.165 0.276 0.268 0.311 0.560 0.261 0.277 0.145 0.187 0.123
4 13 43 178 130
(1.1) (3.5) (11.6) (48.5) (35.3)
0 1 (2.8) 3 (8.6) 19 (54.3) 12 (34.3)
106 129 95 37
(28.8) (35.1) (25.9) (10.1)
13 9 7 6
0.131 (37.1) (25.7) (20.0) (17.1)
Values are mean (SD) and n (%). *R group, a recurrent LDH group; P group, a primary LDH group. #Statistical significance of difference between groups is tested by Student t tests for continuous variables, whereas w2 analyses and Fisher exact tests were used for categorical variables depending on sample size. LDH indicates lumbar disk herniation; JOA, Japanese Orthopaedic Association; VAS, visual analog scale; ZDS, Zung Depression Scale.
The median operative time per level was about 70 minutes (range, 50–90 min). The recurrence rate was 8.7% (35 of 402 patients), and the median time to recurrence was 26.2 months (SD = 4.1 mo). The median VAS scale for leg pain improved from 4.28 to 0.89 (P < 0.05). The median VAS scale for back pain improved from 4.31 to 1.2 (P < 0.05) and the median JOA improved from 21.5 to 16.4 (P < 0.05). No significant difference was observed between the 2 groups at 1-year follow-up. The surgical outcome for the 2 groups was: excellent, 31.1%; good, 44.3%; fair, 18.9%; and poor, 5.7% at 1-year follow-up. No significant difference was observed between the 2 groups based on surgical successful outcome.
DISCUSSION Individual parameters such as VAS of leg/back pain, JOA, ZDS, duration of symptoms, surgical recurrence, type Copyright
r
2014 Wolters Kluwer Health, Inc. All rights reserved.
FIGURE 1. Artificial neural network output diagram with insets for each layer. Output figure generated by PASW Statistics 18, version 18 (SPSS Inc., 2009, Chicago, IL). Input layer: bias (input layer bias); JOA (the Japanese Orthopaedic Association Score); ZDS (Zung Depression Scale); S (smoking); VAS.L (visual analog scale of leg pain); gender; VAS.B (visual analog scale of back pain); age; OL (occupational lifting); OD (occupational driving); SA (sports activity); BMI (body mass index); LH (level of hearniation); TH (type of herniation); DR [duration of symptoms (mo)]. Hidden layer: H (1:1), H (1:2), H (1:3), H (1:4), H (1:5), H (1:6), H (1:7), and H (1:8); bias (hidden layer bias). Output layer: recurrent lumbar disk herniation (LDH) and primary LDH.
of herniation, smoking, and age can assist to better understanding recurrence predictors. However, the findings from the study presented here indicate that the combination of these parameters, in an ANN model, could be applied to estimate risk of recurrent disk herniation with a high degree of accuracy. The full structure details of the resulting ANN can be saved and then be applied to new or subsequent www.jspinaldisorders.com |
E163
J Spinal Disord Tech
Azimi et al
Volume 28, Number 3, April 2015
TABLE 2. Comparison of the AUC and Predictive Values of ANN and LR Models, and Individual Parameters to Predict Recurrent Lumbar Disk Herniation (n = 402) Parameters/Model ANN LR JOA ZDS S VAS.L Sex VAS.B Age OL OD SA BMI LH TH DR
AUC (%)
P*
Specificity (%)w
PPV (%)w
NPV (%)w
0.84 0.77 0.57 0.56 0.54 0.53 0.52 0.51 0.51 0.49 0.48 0.47 0.47 0.46 0.45 0.44
0.001 0.001 0.24 0.28 0.29 0.32 0.34 0.37 0.41 0.43 0.52 0.53 0.53 0.67 0.67 0.68
46 34 22 19 16 11 9 6 6 5 5 5 4 4 3 3
69 64 57 55 55 54 54 53 53 53 52 52 51 51 51 50
88 82 58 56 54 49 44 42 31 21 20 19 19 18 17 17
*Asymptotic significance on receiver-operating characteristic curve analysis. wSpecificity and predictive values at 95% sensitivity. ANN indicates artificial neural network; AUC, area under the receiver-operating characteristic; BMI, body mass index; DR, duration of symptoms (mo); JOA, the Japanese Orthopaedic Association Score; LH, level of herniation; LR, logistic regression; NPV, negative predictive value; OD, occupational driving; OL, occupational lifting; PPV, positive predictive value; S, smoking; SA, sports activity; TH, type of herniation; VAS.B, visual analog scale of back pain; VAS.L, visual analog scale of leg pain; ZDS, Zung Depression Scale.
cases with no need for further training. Although ANNs show promise, the study sample was small, and therefore the technique will need to be repeated with larger, multicenter data sets and external validation to convincingly demonstrate its validity and predictive power. To date, there are no reports describing ANNs as a predictor for recurrent LDH. However, risk factors for recurrent LDH have been reported to include male sex,5,8 degenerated disk,5 smoking,7,8 trauma,8 annular competence, type of herniation,4 advanced age, high body mass index, protrusion type,6 and occupational lifting.7 Many studies have been conducted to determine the additional factors that may increase reherniation risk. These factors include the type of LDH, the amount of disk material removed, alcohol consumption, and the duration of restricted activities. Few are agreed upon by all. For the analysis of risk factors in recurrent LDH, large-scale multicenter prospective studies will be required in the future.9 Herein, we have presented the first attempt at ANN prediction of recurrent LDH. The results obtained from the sensitivity analysis clearly suggest that an ANN model could be used confidently to predict recurrent LDH. This
TABLE 3. Comparison of ANN and LR Models for Predict Successful Recurrent Lumbar Disk Herniation (n = 100*) Accuracy rate (%) AUC H-L statistics
ANN (95% CI)
LR (95% CI)
P
94.1 (93.2–97.1) 0.83 (0.86–0.88) 40.2 (37.3–49.6)
86.4 (84.5–91.4) 0.76 (0.74–0.78) 55.4 (50.4–59.6)
< 0.001 < 0.001 < 0.001
*Patients of validation group. ANN indicates artificial neural network; AUC, area under the receiver-operating characteristic; CI, confidence interval; H-L statistics, Hosmer-Lemeshow statistics; LR, logistic regression.
E164 | www.jspinaldisorders.com
simple, elegant method may be useful to reduce disk herniation recurrence. The discrimination of R group and P group from each other may be able to substantially help in therapy planning. Using an increased number of patients and additional parameters for input layer of ANN, the accuracy of the ANN model is enhanced regarding the optimization of the clinical decision-making process.19 Hence, the ANN model introduced in this study appears to be an acceptable test to predict recurrent LDH. There were some limitations in our study. First, we likely did not identify all possibly significant variables to predict recurrent LDH. Future studies of this model may consider the effect of a more detailed database that contains more input variables (such as clinical and imaging details), as well as more detailed result data. Second, the number of patients was relatively small. Third, further randomized clinical trials are needed to establish benefit and to confirm these findings in ANN model. Finally, ANNs will never replace human expert decision makers, but they can assist in double-checking and enhancing the routine decisionmaking process.
CONCLUSION The findings presented here show that ANNs can be used to effectively assist with the prediction of recurrent LDH before the performance of the initial or index microdiscectomy. ACKNOWLEDGMENTS The authors thank the staff of the Neurosurgery Unit at Imam-Hossain Hospital, Tehran, Iran. Copyright
r
2014 Wolters Kluwer Health, Inc. All rights reserved.
J Spinal Disord Tech
Volume 28, Number 3, April 2015
REFERENCES 1. Ahsan K, Najmus S, Hossain A, et al. Discectomy for primary and recurrent prolapse of lumbar intervertebral discs. J Orthop Surg (Hong Kong). 2012;20:7–10. 2. Aizawa T, Ozawa H, Kusakabe T, et al. Reoperation for recurrent lumbar disc herniation: a study over a 20-year period in a Japanese population. J Orthop Sci. 2012;17:107–113. 3. Ambrossi GL, McGirt MJ, Sciubba DM, et al. Recurrent lumbar disc herniation after single-level lumbar discectomy: incidence and health care cost analysis. Neurosurgery. 2009;65:574–578. 4. Carragee EJ, Han MY, Suen PW, et al. Clinical outcomes after lumbar discectomy for sciatica: the effects of fragment type and anular competence. J Bone Joint Surg Am. 2003;85:102–108. 5. Cinotti G, Roysam GS, Eisenstein SM, et al. Ipsilateral recurrent lumbar disc herniation: prospective, controlled study. J Bone Joint Surg Br. 1998;80:825–832. 6. Kim JM, Lee SH, Ahn Y, et al. Recurrence after successful percutaneous endoscopic lumbar discectomy. Minim Invasive Neurosurg. 2007;50:82–85. 7. Miwa S, Yokogawa A, Kobayashi T, et al. Risk factors of recurrent lumbar disc herniation: a single center study and review of the literature. J Spinal Disord Tech. 2013. [Epub ahead of print]. 8. Suk KS, Lee HM, Moon SH, et al. Recurrent lumbar disc herniation: results of operative management. Spine. 2001;26: 672–676. 9. Shin BJ. Risk factors for recurrent lumbar disc herniations. Asian Spine J. 2014;8:211–215.
Copyright
r
2014 Wolters Kluwer Health, Inc. All rights reserved.
Predicting Recurrent LDH Using ANN
10. 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. 2013. [Epub ahead of print]. 11. 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;20:300–305. 12. Azimi P, Mohammadi HR. Predicting endoscopic third ventriculostomy success in childhood hydrocephalus: an artificial neural networks analysis. J Neurosurg Pediatr. 2014;13:426–432. 13. Price DD, McGrath PA, Rafii A, et al. The validation of visual analogue scales as ratio scale measures for chronic and experimental pain. Pain. 1983;17:45–56. 14. Azimi P, Mohammadi HR, Montazeri A. An outcome measure of functionality and pain in patients with lumbar disc herniation: a validation study of the Japanese Orthopedic Association (JOA) score. J Orthop Sci. 2012;17:341–345. 15. Zung WW, Richards CB, Short MJ. Self-rating depression scale in an outpatient clinic. Further validation of the SDS. Arch Gen Psychiatry. 1965;13:508–515. 16. Macnab I. Chapter 14: pain and disability in degenerative disc disease. Clin Neurosurg. 1973;20:193–196. 17. Cross SS, Harrison RF, Kennedy RL. Introduction to neural networks. Lancet. 1995;346:1075–1079. 18. Rughani AI, Dumont TM, Lu Z, et al. Use of an artificial neural network to predict head injury outcome. J Neurosurg. 2010;113: 585–590. 19. Azimi P, Mohammadi HR, Benzel EC, et al. Artificial neural networks in neurosurgery. J Neurol Neurosurg Psychiatry. 2015;86: 251–256.
www.jspinaldisorders.com |
E165
