Role of Machine Learning in Management of Degenerative Spondylolisthesis

A Systematic Review

Sherif El-Daw, MD; Ahmad El-Tantawy, MD; Tarek Aly, MD; Mohamed Ramadan, MD


Curr Orthop Pract. 2021;32(3):302-308. 

In This Article


The most frequent type of spondylolisthesis that is found in adults is degenerative,[41] and its management is an everyday challenge for spinal physicians. In these patients, a combination of degenerated intervertebral disc, facet joint arthropathy, and stenosis of the spinal canal usually leads to one vertebra slipping over the other. The condition is more common in women with radiographic prevalence of about 40% in Caucasian populations and is usually located in the lower lumbar region. The most common complaints are lower back ache, lower limb radicular pain, intermittent claudication, and neurological deficits.[42–45]

Surgical treatment in degenerative spondylolisthesis is indicated if conservative treatments fail or when there is progressive slippage or serious neurological deficits. Surgical treatment includes decompression and stabilization. In a review of the literature, there were many articles that discussed treatment methods,[46] but only a small number of them were controlled trials.

Many hospitals use digital recording systems, but the use of such electronic health records (EHR) usually is limited to a local level. This quantity of data could detect differences in practices and give a chance to standardize processes. Electronic medical records could provide an opportunity to analyze medical data in different ways. Advancement of computer technology allows easier extraction of information from recorded data and uses these data to populate clinical registries. Imaging data that are present in the picture archiving and communication system also can be uploaded and analyzed easily through the machine learning algorithms that are used. The new digitized medical equipment allows coordination of clinical and imaging data to make the analysis of the patient's data complete.

The number of publications concerned with machine learning with regard to treatment of the spine has increased rapidly over the last few years, but few studies have focused on surgery for degenerative spondylolisthesis. Artificial intelligence uses computer programs to predict outcomes through simulating cognitive capabilities.[47,48] Artificial intelligence and machine learning enable algorithms to transfer large amounts of data that are found in the databases to mathematical forms to produce an understandable expected result.[49,50] The use of machine learning to handle large datasets can result in new correlations that could benefit medical practice.[51–54]

Spinal surgery is continuously looking for ways to decrease complications and improve surgical outcomes because of the high risks that are encountered. The main aims for the use of machine learning are to decrease the load of disease management on healthcare providers and patients and also to decrease the time for reaching a diagnosis.

Machine learning can be divided into four types of learning: supervised (as logistic regression, SVM, neural networks, and Naive Bayes approaches), unsupervised, reinforcement, or semisupervised.[55] Both supervised and unsupervised techniques need algorithms that are designed to formulate understandable results.

In the search that was conducted in this study, eight articles that discussed the role of artificial intelligence and machine learning in the management of spondylolisthesis were detected; five of them were concerned with diagnosis. The problem we faced in evaluating these articles was the differences in the modeling approaches that were applied, the algorithms that were used, the ways of inserting or getting data, and the methods of assessment of expected results. Demographic data, patients' clinical histories, and investigations often were used to detect the surgical results. SVM proved to be the best algorithms in the machine learning materials.[56]

Unfortunately, the studies that were reviewed did not address the way to define the type of spondylolisthesis or to compare conservative treatments with surgical treatments of degenerative spondylolisthesis.

All articles that discussed the diagnosis included spondylolisthesis with other vertebral column disorders, especially disc herniation. They also included many mathematical equations that could be difficult to be understand by clinicians.

Ansari et al.[23] used machine learning for the diagnosis of spinal pathologies. They used the neural networks and SVM and classified the pathologies in three classes (normal, disk hernia, and spondylolisthesis). They concluded that the generalized regression network performance was better than the neural network and SVM (accuracy of 92.05%), and the SVM was more effective than the neural network.

Karabulut and Ibrikci[24] studied the same vertebral pathologies with the aim to develop a new method that can classify models of these patients automatically to help healthcare personnel to make the right decisions. They mentioned that the best results were found with the oversampling technique of SMOTE and a composite classification algorithm.

Akben[25] used the Naive Bayes classifier to analyze the data in his study. He also used datasets of three classes: healthy, hernia, and spondylolisthesis. The results were evaluated according to single or combined use of the attributes, such as pelvic incidence, pelvic tilt, sacral slope, slippage degree, and lordosis angle. He found that use of three or more attributes results in classification success much more than classification with the use of a single attribute. The success rate will be about 84.5% in assigning classification when using three or more attributes. However, similar to previous researchers, he noted that the result is inaccurate regarding successful diagnosis of a disorder because the number of subjects in the groups were not equal. So, the high classification success rate of spondylolisthesis (accuracy rate 96.13%) may be misleadingly high because of the large number of spondylolisthesis patients compared with hernia patients. He also found that the success rate of the hernia-spondylolisthesis classification was very high. So, he stated that spondylolisthesis can be differentiated easily from the healthy and hernia. He mentioned that patients with spondylolisthesis easily can be distinguished from the patients who have a hernia or patients who are healthy by using the parameter of the grade of spondylolisthesis. The other parameters, such as shape and spinopelvic orientation, may be misleading for healthy-spondylolisthesis and hernia-spondylolisthesis classification with a success rate of 98.10%, but those parameters can be used for a hernia-spondylolisthesis classification since all attributes are informative compared with healthy-spondylolisthesis classification.

Unal et al.[26] used pairwise fuzzy C-means (FCM) based feature weighting method for classification of the spinal pathologies that aimed to improve the classification performance of algorithms that are used and to change the nonlinearly separable datasets to linearly separable datasets. They showed that the technique was strong in the classification of the spinal pathologies' dataset. Unal et al.[26] concluded that the best way for this dataset was a combination of FCM and Naive Bayes that give classification accuracy of 97.42%, and the combination of FCM and SVM gave classification accuracy of 96.45%. This accuracy percentage was much higher than using SVM alone (accuracy of 82% to 85%).[57]

One of the points of disagreement, is whether to add fusion to decompression in patients who were suffering from degenerative spondylolisthesis. Most importantly, there is disagreement regarding how to identify the patients that may develop instability after decompression alone.

Reviewing the randomized controlled trials, it was found that instrumentation of the spine usually was associated with a high fusion rate, but the patient outcome may not have been affected. Because the assessment tools that were used in those studies were nonvalidated, it was very difficult to get a strong conclusion about comparing patient outcomes.[58,59] Norton et al.[37] extracted demographic data and outcome measures for all patients who underwent surgery for degenerative spondylolisthesis from 2001 to 2010. The complication rate was 22.5% (in 48,911 surgically treated patients). The complication rate differed according to the approach that was used and was restricted to patients with added fusion. SIDs contain data that can track patients over time and are used to study readmission and reoperation rates. Using that dataset in patients who have lumbar spondylolisthesis, a study found that the reoperation rate was 17% at 5 yr (either with adding fusion to decompression or without).[38]

Despite databases that contain large amounts of data, there are still some problems in using these datasets. Gologorsky et al.[60] mentioned that the indication for surgery by every surgeon was not always compatible with the diagnosis codes found in the dataset. Thus, larger datasets should be reviewed critically when informing policy. Also, the databases are not perfect if we want to compare surgical procedures because they usually do not contain patient-reported outcomes. The American Association of Neurological Surgeons has proposed a database for quality outcome for surgery of spondylolisthesis. Investigators found that 82% of the patients (of 441 patients) who were treated with fusion and decompression were equivalent to those reported in spinal instrumented laminectomy in short-term follow-up.[61,62]

Reduction in pain and return of function are the most important aspects of spinal care, and such information can be provided only by the patients. So, an advanced system has been developed called Patient-Reported Outcomes Measurement Information System and has been evaluated for lumbar spinal disorders.[63] This system utilizes computer technology to reduce the size of the questionnaire answered by the patients. Some researchers developed techniques to assess opioid consumption in lumbar spondylolisthesis patients. They found that preoperative chronic opioid use was found to be a strong indicator for postoperative opioid use.[64]

Regarding nonhome discharge, Ogink et al.[39] developed a machine learning algorithm to predict patients who are likely to have an extended length of stay and subsequent adverse events. The parameters they used included age, gender, BMI, preoperative white blood cell count, preoperative creatinine, diabetes, surgical procedure, number of levels operated, and ASA class. They mentioned that the Bayes machine is considered the best model depending on discrimination, calibration, and overall model performance. These calibration metrics are essential to determine whether the predictive models or scoring systems are useful.

Limitations and Future Perspectives

There are some limitations in machine learning techniques as there are in any promising scientific method. First, the errors in the use of different coding in spinal surgery outcomes may prejudice the outcomes and the algorithm that are used.[65] Second, some variables, such as health insurance state, employment type, and preoperative patient's scores that are considered to be important predictors for discharge placement, may not be present in all datasets. Third, most databases are formed from data from patients in the US or European countries, and this data may not be suitable for application to patients in other areas in the world. Encouraging all hospitals to use electronic medical files would improve the machine learning output in the future.