Utilizing machine learning to predict post-treatment outcomes in chronic non-specific neck pain patients undergoing cervical extension traction

Utilizing machine learning to predict post-treatment outcomes in chronic non-specific neck pain patients undergoing cervical extension traction

More and more, machine learning models are being widely developed and utilized in healthcare examination, diagnosis, and treatment decision algorithms in the efforts to reduce costs, systematize cumbersome processes, and develop prediction rules with the goals of improving selection and outcomes of patients undergoing various interventions25,26,27,28,29. To our knowledge, no investigations have used machine learning tools to look at the prediction of patient outcomes undergoing a multimodal rehabilitation program using CET to improve the sagittal plane alignment of the cervical spine in patients suffering from chronic non-specific neck pain (CNSNP) disorders. Since surgical7,9,12,30 and conservative10,11,13,22,23,24 analysis and rehabilitation interventions of sagittal cervical malalignment represents a large financial burden for societies, machine learning affords a unique area of opportunity to improve understanding, outcomes, and potentially reduce costs. In our investigation, by integrating Python, sci-kit-learn, and NumPy, we facilitated a rigorous and data-driven approach, allowing for a nuanced exploration of predictive modeling in the context of patient outcomes following a standardized treatment approach including CET22,23,24. By leveraging these sophisticated tools, our study aimed not only to forecast post-treatment outcomes (ARA C2–C7, NDI, and NRS) but also to provide a comprehensive understanding of the factors influencing the efficacy of the CET procedure. In the ensuing sections, we discuss the results of the linear regression models, shedding light on their performance and implications for clinical practice.

A groundbreaking aspect of our investigation lies in its pioneering use of machine learning to predict outcomes in CNSNP patients undergoing the CET procedure using a prospectively collected database that was large enough (570 consecutive patients) to develop and test the models. The study’s use of linear regression models to predict post-treatment outcomes demonstrates promising precision and accuracy. The R-squared value is crucial in assessing the accuracy of fit of the linear regression model. A higher R-squared value indicates that the model accounts for a larger proportion of the variability in the outcome variable. The identified R-squared values (ranging from 0.298 to 0.549) indicate a moderate to strong explanatory power for the variability in predicting the outcomes in CNSNP intensity (NPRS), neck pain disability (NDI), and the cervical lordotic angle (ARA C2–C7), respectively. The variability explained by the models, provides valuable insights into the predictive capabilities of the linear regression approach. For the cervical lordotic angle, disability index, and pain score, the respective R-squared values imply that a significant portion of the variance in these post-treatment outcomes can be attributed to the predictors included in the models. This suggests that the linear regression models offer meaningful insights into the factors influencing the outcomes of patients undergoing this multi-modal CET application program of rehabilitation.

For the cervical lordotic angle, the mean squared error (MSE) of 6.628 implies that, on average, predicted values deviate by 6.628 units (°) from actual values of curvature change. The R-squared value of 0.549 indicates that the model can explain approximately 54.9% of post-treatment lordotic angle variability, reflecting a moderate to strong association between predictors and observed angles. The precision in predicting the change in cervical lordotic angle may be attributed to the relatively straightforward biomechanical nature of this outcome, where the chosen predictors have a more direct impact on the cervical spine’s alignment. We recognize that enhancing the model’s predictive accuracy could be possible by adding other important mechanical factors. For instance, biomechanically, the cervical and thoracic spines are interrelated7,8,12,30, adding thoracic alignment features such as thoracic kyphosis magnitude12, T1 vertebra sagittal slope12,30, the thoracic inlet morphology30, and the relationship between cervical lordosis and T1 slope (T1slope—ARAC2–C7 ≤ 20°)7,12 as input variables presents an important opportunity for enhancing the predictive capabilities of the model. Additionally, adding other spine regions as mechanical features such as spinopelvic parameters may significantly enhance the model’s robustness and accuracy, considering the correlation between the full spine alignment and the cervical lordotic angle12.

The cervical lordosis has been reported to have a relatively broad range of normative alignment values ranging from mildly kyphotic to a deep lordosis and is influence by many variables including type of population being investigated12,31, measurement methodology32, advanced age12, and may be influenced by ethnicity12. In the current investigation we choose the cutoff for normal versus abnormal cervical lordosis amount of 20° as it has been identified as a statistically significant value with good sensitivity and specificity in separating normal vs. chronic neck pain populations with little to no postural AHT31. Furthermore, the 20° cervical lordosis mark (accounting for differences in measurement techniques) is the average value of post treatment outcomes in chronic pain patients undergoing CET procedures and this value seems to be related to improved short-term and long-term pain, disability and neurophysiological findings22,23,24. Finally, in the current prospective population of 570 patients with CNSNP treated with CET, the actual post treatment cervical lordosis reached the 20° magnitude; further evidencing this as a minimum benchmark of rehabilitation care.

According to the model results, the feature that had greatest influence on the amount of post-treatment change in cervical lordotic curvature was the pre-treatment cervical curve magnitude (ARA C2–7), with a feature coefficient (FC) of 2.78. This seems logical in as much as the more abnormal a neck curve is to begin with, it will result in having a positive affect increasing the potential for curve correction following CET application. Furthermore, features that positively influenced the amount of cervical curve change included frequency (FC of 1.84) and duration (FC of 1.54) of CET application; again, this seems logical as the more often you apply a procedure, in theory, the more benefit should occur. In contrast, three features were identified to negatively influence the amount of cervical lordotic change and these included presenting neck disability score (FC of − 0.45), age (FC of − 0.97), and compliance (FC of − 0.43). These negative feature coefficients, suggest that higher initial NDI scores, older age, and a greater lack of adherence to the recommended program all have the potential to decrease the outcome of curve correction. The finding of increased initial neck disability influencing the outcomes of patients suffering from neck injuries and neck pain has been previously documented in the motor vehicle crash collision literature; which adds credibility to our modelling predictions33,34.

Regarding the neck disability index or NDI, the minimal detectable change (MDC) for the NDI is 6.9 and the minimal clinically important difference (MCID) for the NDI is 5.535. Herein, we used two thresholds for the MCID of the NDI of 5.5 and 6 points and we identified that 82.1% and 71.4% of our patients achieved these respective MCID values and above after our treatment program. Our analysis indicates that our model exhibits a MSE of 2.091, suggesting relatively low prediction errors in estimating post-treatment NDI scores. The R-squared value of 0.305 signifies that 30.5% of the variability in post-treatment NDI can be explained by the model, indicating a substantial proportion of captured variance. Our comprehensive analysis of feature importance not only elucidates the individual contributions of each factor but also aids in understanding the intricate dynamics of the CET procedure’s effectiveness. According to the model results, the features that had the greatest positive influence on the amount of post-treatment change in neck disability were the pre-treatment cervical curve magnitude (ARA C2–7 with FC of 0.42) and the initial disability level (NDI with FC of 0.33). The positive influence of these features (ARA and NDI) indicate more or faster benefit in those patients presenting with a greater loss of cervical curvature and with an initial higher disability level both indicating greater room for improvement. In contrast, the NDI is notably negatively influenced by the treatment variables of frequency, duration, and compliance % (missed sessions) of CET application, with feature coefficients of − 0.42, − 0.56, and − 0.20, respectively. Previously, it has been demonstrated that a greater dose of care, of up to 16–18 sessions results in a greater improvement in patient outcomes in chronic spine related conditions thus validating our current model’s predictions36,37.

In terms of the global burden of disease, chronic neck pain is known to be one of the greatest contributors1. Underpinning the understanding and outcomes for patients suffering from this disorder is the influence of any specific variable that might be mitigated in order to improve the odds of recovery; obviously, some patient variables can’t be changed (age, sex, etc.) The NPRS has an MDC of 2.6 and the MCID has been reported to be 1.535. We chose a threshold for the MCID of 2 points to indicate the percentage of patients who achieve the MCID at follow-up treatment on the NPRS and we identified that 96.67% of patients achieved this minimal or above amount of improvement. The high percentage of patients achieving MCID in the post treatment response pain score underscores the clinical relevance of this treatment protocol in reducing pain severity using the NPRS pain score. Furthermore, our modelling findings of a low MSE of 0.342 suggests relatively accurate predictions for the change in pain intensity post treatment with CET; while the R-squared value of 0.298 indicates a moderate association between predictors and observed pain score change. Notably, we identified that the NPRS pain score change is negatively influenced by treatment frequency (FC of − 0.139) and duration (FC of − 0.240) of CET application, indicating that the fewer number of interventional sessions over a less amount of time has a negative impact on a patient achieving greater pain improvements. The finding of increasing dose of treatment application in mitigating post-treatment pain has been previously reported36,37, however, this is the first time it has been reported in a population receiving a CET procedure.

The observed patterns suggest that input factors included in the models have a more pronounced impact on lordotic angle than on pain and disability index. This could indicate the complexity of pain and disability in CNSNP is influenced by a broader spectrum of variables not fully captured in the current model. For instance, existing research consistently underscores the significant role of cognitive factors in influencing the levels of disability and pain experienced by patients with persistent chronic pain disorders38. Moreover, the stress-buffering model posits that social support plays a constructive role in enhancing health outcomes, shielding individuals from the detrimental impacts of stress. Studies have provided evidence supporting the notion that higher levels of perceived social support and justice correlate with reduced pain severity and decreased pain-related disability among individuals dealing with chronic pain-related psychosocial conditions39. Furthermore, a recent meta-analysis identified that psychological interventions delivered by physiotherapists were more effective than standard physiotherapy for CNSNP, but the effect sizes were small-to-medium and these investigations did not include an analysis of any RCT’s using CET devices to rehabilitate the cervical lordosis40. Consequently, the current project model struggles to capture the entirety of variability in the complex measures of neck disability and neck pain, resulting in lower R-squared values. This highlights the challenges in predicting outcomes that are inherently influenced by diverse and interconnected factors, emphasizing the need for a more comprehensive and nuanced approach in future predictive modeling efforts.

Interestingly, we identified no predictive value for the amount of forward head posture [(FHP) measured as AHT C2–C7 herein] relative to the outcomes of cervical lordosis, NDI and NRS pain intensity and this seemingly contradicts the prevailing literature7,10,11,12,13. However, in our current sample of patients the mean initial magnitude of FHP was minimal (22.4 mm) and this is below the threshold of 40 mm that has been reported as the optimum cut point that explains increasing NDI and NRS pain scores7,41. Furthermore, the small amount of FHP in our population is unlikely to biomechanically affect the magnitude of cervical lordosis8; these two reasons likely explain our lack of predictive values for the amount of FHP in our modelling. However, since the integration of machine learning into spinal rehabilitation has emerged as an important tool for advancing personalized healthcare, several of our model’s results are clinically relevant. First, the capacity to predict treatment outcomes based on individual demographic and clinical patient characteristics has implications for selecting the proper CET application parameters. Spinal rehabilitation specialists, equipped with insights into these factors that influence post-treatment changes, can make better-informed decisions regarding treatment plans, enhance the precision of treatment strategies, and potentially enhance patient satisfaction and adherence. Second, understanding the role that patient age, compliance to a defined course of action, treatment frequency and duration needs, coupled with the amount of initial pain and disability at presentation can help mitigate patient confusion and increase patient participation in their healthcare goals with the effort to focus on things that can be controlled by the patient and provider.

As with any study, it is essential to acknowledge the limitations that may influence the generalizability of the findings. The study’s focus on a specific patient population with CNSNP and the potential influence of unmeasured or latent variables introduces an element of uncertainty. The acknowledgement of these limitations opens the door to future research endeavors. Additionally, for all pre-treatment evaluations, we relied on a single evaluation (i.e., “snapshot medicine”). To further improve the accuracy of prediction models we could gather pre-treatment information at several time points to obtain more reliable self-reporting information. For example, neck disability or pain intensity diaries over 7 days could be implemented in routine clinical care. Future directions for research may involve additional predictors specific to pain intensity and disability (psycho-social and emotional variables), which showed a slightly lower R-squared value compared to lordotic angle. Additionally, the incorporation of full spine and other spine alignment data along with more advanced machine-learning techniques could help capture non-linear relationships that may be present in the data, further refining future predictive models. Finally, our current investigation did not include medium to long-term follow-up measures for stability of patient improvements over time. However, previous studies suggest that the improvements in cervical spine alignment and consequent pain and disability improvements can last up to 2 years post-rehabilitation, with minimal loss of benefits5,23,42. This durability is likely due to the structural changes in cervical alignment induced by the CET treatment, which need to be monitored in future longitudinal studies.

link

Leave a Reply

Your email address will not be published. Required fields are marked *