1 Introduction

Lumbar spinal stenosis (LSS) is a clinical syndrome associated with the reduction of the diameter of the spinal canal resulting in a reduction in the space available for neurons and vascular of the spine.¹ LSS can be classified according to aetiology in congenital, acquired, or both,² or according to anatomical localization in central, foraminal, and lateral. Causing an LSS is disc bulging, loss of disc height, facet joint arthropathy, osteophytes, yellow ligament formation and hypertrophy, all of which can lead to narrowing of the spinal canal, which could also be exacerbated by the displacement of a vertebral body anteriorly or posteriorly, relative to the adjacent vertebral body (spondylolisthesis). LSS is one of the most frequent reasons for access to orthopaedic and neurosurgical examinations and represents the most frequent cause of spinal surgery among individuals over 65 years of age.³

The most common clinical symptoms of LSS are strength deficit and back, buttocks, and lower extremity pain. This symptomatology might significantly impair the patient's ability to walk and global functional autonomy.⁴ Furthermore, pain can have an impact on mood, causing anxiety and depression also in relation to the impairment and limitation of personal independence.⁵

Surgery is effective in well selected patients with severe strength deficit of lower limbs and with back and leg pain resistant to conservative management. A Cochrane review pointed out that there is no evidence of efficacy for the conservative therapy alone, and that different types of surgery are used when there is no improvement to a conservative approach, without clear evidence of efficacy.

It is documented that a significant number of subjects affected by LSS who were treated surgically had substantially greater improvement in pain and motor function up to four years than those who were treated with conservative approaches.⁶ However, a significant proportion of those who undergo surgery experience residual symptoms especially when it comes to pain.⁷ A recent retrospective study conducted on the registry of Norway with more than 11.000 subjects, estimated that 33 % of subjects reported failure after surgery for LSS.⁸ Several research groups have sought to identify the predictive factors of efficacy of surgical treatment of LSS: biological, sociodemographic, occupational, and psychosocial.

A recent study reports that about 20 % of patients undergoing spinal surgery have residual symptoms in the postoperative period and are therefore candidates for re-surgery, pointing out that there is a weak consensus on the type of surgery, on the prognostic factors that could be helpful to identify the good responders to surgery, and on the criteria identifying risk of re-surgery.⁹ Moreover, psychosocial factors are determinants for the development of chronic low back pain, and therefore the evaluation of patients with LSS must not be based on the study of morphological/surgical factors alone but also of biopsychosocial ones.¹⁰

This significant variation in the clinical outcome following surgery may indicate that the greatest predictive markers for surgery in patients with LSS may not be revealed by traditional statistical methods. One reason could be that traditional statistics only consider the linear links between each element and the result, ignoring interactions and more intricate (non-linear) linkages. The application of machine learning (ML) techniques has evolved as a more sensitive way to discriminate between different classes of potential prognostic indicators beneficial for highly accurate outcome prediction in other disciplines, such as rehabilitation.¹¹

The aim of this study is to evaluate, using an artificial neural network (ANN), motor, functional and psychological profile of subjects with LSS and with indication to spinal surgical treatment, to identify the cohorts of patients who most benefit from the surgical approach in terms of biopsychosocial well-being and global functional autonomy recovery.

2 Material and methods

The research was designed as an observational cohort study. Subjects consecutively admitted to Neurosurgery Department of the San Salvatore Hospital in L'Aquila with a diagnosis of LSS and with surgical indication have been progressively screened. Study participants were enrolled on a voluntary basis after informing them about the contents and purposes of the study and signed the informed consent. The research was approved by the Ethics Committee of University of XXX (protocol n. 07/2022; 1^st March 2022).

Inclusion criteria were age ≥18 years; LSS caused by degenerative pathology of the lumbar spine such as herniated discs, degenerative disc disease, osteoarthritis with osteophytes, congenital reduction, and spondylolisthesis. Exclusion criteria were represented by LSS due to traumatic pathology or tumour of the spine and patients with diagnosis of other rheumatological, neurological or psychiatric diseases.

Patients in the preoperative phase, within two days of surgery, have been carried out a longitudinal clinical evaluation who investigated pain severity, motor, cognitive and behavioural performances, and the functional global independence. The same battery of clinical scales will be administered two months after surgery.

Demographics and clinical characteristics of the sample as well as the quality of life was measured in the preoperative phase.

2.2

2.2 Neural network analysis

The Artificial Neural Network (ANN) analysis was performed on SPSS v.23, using the Neural Network module (IBM Corp., Armonk, NY, USA). We used an ANN model already used successfully in previous studies (ARtificial Intelligent Assistant for Neural Network Analysis - ARIANNA-).^13,14 ARIANNA is based on a Multilayer Perceptron Procedure, and it is formed by the input layer (from which entered the above listed 24 variables), two hidden layers (of 5 elements each one), and a final output layer (the output of which was the predicted outcome) (Fig. 1). The architecture of the ANN was that of a Feed Forward Neural Network (FFNN), with data moving in only one direction, from the input nodes through the two hidden layers to the output node. For all the units in the hidden layers and for the output layer the activation function was a hyperbolic tangent. The chosen computational procedure was based on an online training.¹³ Because of the reliability problems of ANN,¹⁵ ARIANNA ran 10 times and the best results (in terms of higher percentage of predictions with an error <10 %) was used as its result (see Fig. 2).

Close

Fig. 1 ANN structures. The 24 input variables were analysed via neural network analysis with two hidden levels (five layers each) to provide a single layer output for disability estimation.

Export to PPT

Close

Fig. 2 Flowchart of patients' recruitment.

Export to PPT

3 Results

108 subjects accepted to participate in this clinical study and were initially assessed from March 1, 2022, to Decembre 31, 2023. Of these, the surgery was postponed from the enrolment in 5 subjects, and in the other 13 the follow-up evaluation after surgery was not completed. Ninety subjects also completed the post-surgery assessment, and their data were analysed. Their demographic and clinical conditions have been reported in Table 1. Clinical scale scores assessed at baseline and post-surgery, with the p-value of their paired comparison, were reported in Table 2. The results of linear regression analysis identify three variables into the model: Oswestry disability score, VAS Pain and FIM motor score (Table 3). The accuracy of predicting outcome with an error <10 % of this approach was 54.4 % with a mean absolute error of 11.2 ± 9.3 %.

The results of the regression and ANN predictions are reported in Fig. 3 compared to the assessed Oswestry score after surgery. The mean absolute error of ANN prediction was 7.0 ± 8.5 % and the number of predictions with a discrepancy lower than 10 points on 100 was 81.1 %. Some of the variables assessed at baseline highlighted by ANN as having a weigh >5 % resulted also significantly correlated with the outcome (Oswestry disability score: R = 0.372, p < 0.001; Roland and Morris score: R = 0.314, p < 0.001; Quebec back pain score: R = 0.359, p < 0.001; VAS Pain R = 0.379, p < 0.001; FIM motor score: R = −0.255, p = 0.015). The Oswestry disability score post intervention was found slightly higher in women than in men (30.4 ± 22.4 % vs. 21.7 ± 13.5 %, p = 0.086) but it was already higher at baseline (53.8 ± 19.4 % vs. 44.2 ± 19.0 %, p = 0.013). The higher weights on ANN output were observed for the Oswestry Disability Score (0.076); in the Quebec Back Pain Score (0.073), in the FIM Motor score (0.073), for the height (0.072), the time from symptom onset (0.071), the Roland & Morris (0.07), and in the age (0.07). All results of the normalised with respect to the highest weight (the sum of all weights is 1) are reported in Fig. 4.

Close

Fig. 3 ANN (filled dots) and Linear regression (empty dots) predictions (y axis) with respect to Oswestry score assessed post intervention (x axis). The black solid line represents the perfect prediction and the dotted grey line the range of predictions within ±10 points of error.

Export to PPT

Close

Fig. 4 Results of ANN expressed like normalised importance with respect to the highest weight.

Export to PPT

4 Discussion

The aim of the present study was to investigate the motor, functional and psychological profile of subjects with LSS with indication to surgery approach, to identify the cohorts of patients who might benefit more from the surgical treatment in terms of functional independence recovery, using an ANN able to model complex relationships. The use of ANN decreased the mean absolute error of the predictions from 11.2 points of Oswestry disability score to 7, increasing the accuracy of a prediction with a range of ±10 points from 54 % up to 81 %. The use of multiple variables into the ANN provided an accurate estimation of surgery outcome in a wide group of patients with LSS, more than linear regression. Specifically, both clinical and demographic characteristics showed a weight in surgery success. Disability, pain, and motor function were the most determining clinical factors to predict surgical success in terms of improvement of these parameters. Regarding demographics characteristics, the height, the time from symptom onset and the age were the most relevant prognostic factors about surgical efficacy followed by the body mass and gender.

Finally, regarding clinical data, the Oswestry Disability Score, the Quebec Back Pain Score, the FIM Motor score and the Roland & Morris scales showed a higher weight in the estimation of surgical success. These tools, despite slight differences, measure and describe the disability and its impact on activity of daily living. In the ANN analysis, differently to the multiple regression that can flatten the effect of scales that measure the same outcome, each tool is weighed to obtain the best possible outcome value.

Our results are in accordance with a recent study of Hou et al., 2024 that identified the initial high disability score, measured by ODI, was an independent predictor to allow minimal clinical important difference at ODI measure after a follow-up period of at least 1 year.¹⁶

The same results were already obtained by Singh et al., 2023 that identified low preoperative score and poor muscle health the only independent factors to not achieve the MCID at Oswestry at 1-year follow up.¹⁷

Therefore, considering the possible collateral effects due to surgery intervention (ranging from 10 % to 24 % of cases)¹⁸ the opportunity of conservative treatment (i.e. pharmacological intervention, physical therapy, rehabilitation program) options should be considered in patients with slight disability and pain.¹⁹

Despite many research and manuscript reporting results on prognostic factors for LSS, there are few studies applying artificial intelligence and machine learning regarding lumbar stenosis surgery.

De Barros and colleagues (2023) proposed a machine learning approach for computing the probability of spinal surgical recommendations for LSS based on patient demographic factors, previous therapeutic history, symptoms and physical examinations and imaging findings.²⁰ Their model had a good prediction accuracy in proposing a binary outcome: recommendation of neurosurgical intervention or not. They found that the MRI severity score was the most important factor for classifying patients. A similar approach was used by Mourad and colleagues (2022) that used artificial intelligence to predict doctor's recommendation for a neurosurgical approach.²¹ They also found that imaging features are the most important index to indicate right recommendations, followed by motor deficits and back pain. Also, Khan et al. (2021), used a binary output from a machine learning algorithm, finding as main prognostic factors the functional status (assessed by the modified Japanese Orthopaedic Association score), male gender, duration of myelopathy, and the presence of comorbidities.²² The accuracy of their model was about 74 %.

An important difference of our study was that we used a continuous outcome related to the disability. It is fundamental to explain to the patients how much he/she could recover with the neurological surgery and not only to suggest or not the surgery. The accuracy of our approach was 81.1 %, higher than that observed using conventional statistics.

Our study also highlighted that, in addition to clinical measures, some demographic characteristics seem to be involved in surgical outcome: the height, the symptom onset, and the age. These characteristics indicate a greater surgical success compared to previously reported clinical parameters. As regards the age, notoriously, LSS is a common condition in elderly patients and represents a common indication of spinal surgery at an advanced age.²³ Also, the ANN analysis confirms that age is a parameter of surgery success. Parallelly, obese and overweight persons are at a higher risk of developing LSS²⁴ and this characteristic also impacts on surgical results in terms of surgical failure. Finally, our results about a patient's height impact on surgery success were aligned with data of Coeuret-Pellicer et al. (2010) a predictor for back surgery.²⁵

A recent study conducted with the Norwegian registry for spine surgery on 11538 subjects with spine surgery for LSS, identified preoperative duration of back pain for longer than 12 months, previous spinal surgery, and age above 70 years as the strongest predictors for increased odds of failure after surgery (ODI<31).⁸

Interestingly our data highlight that anatomical-pathological diagnosis (i.e., stenosis or hernia), anxiety-depressive comorbidity, smoking habits, and other comorbidities (Hypertension, Dyslipidemia, etc) have a marginal impact on the ANN output.

Regarding the correlation between preoperative anxiety and depression and poor surgical outcomes is conflicting. Some authors provided evidence about the relation between anxiety and depression with satisfaction of LSS surgery.^5,26,27

From another side the depression level before and after surgery is not related to patients’ outcomes (pain, disability, and quality of life) after spine surgery, as recently affirmed by a large study of Canadian Spine network.²⁸ Our ANN analysis corroborates these findings. Similarly Held et al. stated that presurgical level of anxiety and depression is not related to the ODDs to achieve good outcomes after LSS surgery.²⁷

So, in light of our research, surgeons should not be concerned that depression and anxiety will reduce outcomes after spine surgery. Therefore, data analysis with ARIANNA methodology, despite the high complexity of the neural network, constitutes a significant opportunity in clinical practice for an informed consent to the patients that is not limited to a suggestion (or not) of surgery, but includes a provision about the probable recovery from the disability.

5 Conclusion

The indication for surgery compared to conservative approach in the treatment of LSS remains unsolved in clinical practice guidelines.¹⁸ The ANN can support decision making, providing clinical and demographic characteristics of patients who may or may not benefit from surgery. Specifically, disability, pain, symptom onset, heights, and age analysed via ANN are greatly involved in surgery success in terms of pain reduction and global functional independence recovery.

Ethical

The research was approved by the Ethics Committee of University of L'Aquila (protocol n. 07/2022; 1^st March 2022).

Founding

This research was funded by the Santa Lucia Foundation and the Italian Ministry of Health, name of grant: “NEURO-METAVERSE: Application in Neurorehabilitation and Neuroscience of Metaverse Technologies as Virtual Reality and Artificial Intelligence”.

Patent

The research was approved by the Ethics Committee of University of L'Aquila (protocol n. 07/2022; 1^st March 2022).

Patient Consent

Not applicable.

CRediT authorship contribution statement

Alex Martino Cinnera: Data curation, Writing – original draft, Writing – review & editing. Giovanni Morone: Conceptualization, Methodology, Writing – review & editing. Marco Iosa: Data analysis, Writing – review & editing. Stefano Bonomi: Data curation, Writing – original draft, Writing – review & editing. Rocco Salvatore Calabrò: Writing – review & editing. Paolo Tonin: Writing – review & editing. Antonio Cerasa: Writing – review & editing. Alessandro Ricci: Writing – review & editing. Irene Ciancarelli: Conceptualization, Methodology, Writing – review & editing.