Identification, validation, and prediction of longitudinal recovery phenotypes in degenerative cervical myelopathy: analysis of prospective cohorts and a randomised controlled trial Authors: Karlo M. Pedro, Mohammed Ali Alvi, Christopher Lozano, Vishwathsen Karthikeyan, Alex B. Bak, Benjamin M. Davies, James S. Harrop, Jetan H. Badhiwala, Jefferson R. Wilson, Wendy Lou, Zihang Lu, Kevin E. Thorpe, Michael G. Fehlings Published: 2026 ABSTRACT SummaryBackgroundSymptom evolution in degenerative cervical myelopathy (DCM) is traditionally characterized by severity of neurological deficit. A recent expert consensus challenged this view, proposing that DCM fluctuates across neurological, functional and health-related quality of life measures. We examined whether joint changes across these three domains define distinct recovery patterns in a clinic-based cohort of adult patients with DCM.MethodsCohorts analysed were the Cervical Spondylotic Myelopathy-North America (CSM-NA, 2005–2007, N = 278), CSM-International (CSM-I, 2007–2011, N = 479) and the CSM-PROTECT Trial (2012–2017, N = 290). All participants were symptomatic adult DCM patients who underwent surgical treatment and completed at least three longitudinal assessments. Bayesian consensus clustering (BCC) was used to simultaneously model one-year changes in the modified Japanese Orthopaedic Association (mJOA), Neck Disability Index (NDI), and health-related quality of life (SF-36 Score). The optimal number of trajectory clusters was selected using maximum mean adjusted adherence. Multivariable logistic regression examined association between trajectory classes and achievement of minimal clinically important difference (MCID) across outcomes with internal validation performed using 100 bootstrap resamples.FindingsThe final analytical cohort consisted of 1047 patients diagnosed with DCM. The mean age was 56.8 years, and 60.7% were male. The mean symptom duration was 29.8 months, and most patients exhibited involvement of two spinal levels. Joint modelling of multiple longitudinal outcomes resulted in three distinct patient groups. Trajectory 1 (n = 177, 17%) was characterised by minimal postoperative improvement with persistently low neurological function and health-related quality-of-life scores. Trajectory 2 (n = 524, 50%) demonstrated the greatest reductions in neck pain disability. Trajectory 3 (n = 346, 33%) exhibited global recovery, with substantial improvements across neurological, functional, and quality-of-life domains. Trajectory 3 patients had markedly higher odds of achieving MCID in neurological function (odds ratio [OR] 10.90, 95% CI 6.15–19.32), physical quality of life (OR 7.32, 95% CI 3.53–15.22), and disability (OR 3.02, 95% CI 1.85–4.93) compared to patients in Trajectory 1. BCC-derived models reduced clustering uncertainty and improved model discrimination beyond conventional outcome predictors. Predictors of global recovery included an anterior surgical approach (OR 3.55, 95% CI 1.21–10.37) and higher baseline frailty (OR 1.42, 95% CI 1.14–1.76).InterpretationWe provide evidence that outcomes in DCM follow three distinct, multidimensional trajectories that are not captured by neurological severity alone. Simultaneous assessment of neurological, functional, and health-related quality-of-life outcomes is both valid and informative, enabling phenotypic stratification and improved longitudinal prognostication in DCM. External validation of these recovery phenotypes in independent DCM cohorts is warranted.FundingNone. Research in contextEvidence before this studyImplicit in characterizing severity in degenerative cervical myelopathy (DCM) is the use of the modified Japanese Orthopaedic Association (mJOA) score, which limits assessment to neurological dysfunction. This monocentric framework fails to reflect functional impact and quality of life, a limitation that has shaped how postoperative recovery in DCM has conventionally been modelled. We searched in PubMed for articles published since the database’s inception up to January 2026 using terms such as phenotype [“cervical myelopathy” OR “degenerative” OR “myelopathy”] AND [“subphenotype” OR “phenotype”] AND [“statistical models” OR “latent class analysis” OR “trajectory”]. We found that existing trajectory-based approaches rely predominantly on single outcome measures. A recent recommendation from an international DCM initiative advocates the combined assessment of neurological, disability, and quality-of-life outcomes as a core framework for evaluating postoperative recovery.Added value of this studyBuilding on this recent progress, we examined whether the combined assessment of mJOA, neck disability index (NDI), and SF-36 Physical Component Score - core elements of the proposed DCM outcome set - has construct validity and captures clinically meaningful recovery heterogeneity in a real-world DCM cohort. We applied joint trajectory modelling to evaluate simultaneous change across these interrelated outcomes, representing a novel application in spine research. This analysis identified three distinct multidimensional recovery trajectories, with most patients demonstrating predominant improvement in neck pain disability at one year. Models that included these clusters were robust to discrimination and prediction of postoperative improvement. To facilitate clinical translation, we developed a practical tool to assist clinicians in identifying expected recovery trajectories and individualizing DCM management.Implications of all the available evidenceTaken together with recent international recommendations, our study findings raise an important clinical question: whether DCM can continue to be meaningfully represented using a single outcome scale, or whether it should instead be recognised as a condition with inherently multidimensional domain involvement? Our results support the latter, demonstrating that combined neurological, functional, and quality-of-life assessment effectively characterises prognostic phenotypes in DCM. This multidimensional approach provides a valuable tool for subphenotyping in future clinical trials and treatment allocation of this common but debilitating condition. Nevertheless, external validation in independent DCM cohorts is required to confirm the reproducibility, generalizability, and clinical utility of these trajectory-derived phenotypes. Research in context Evidence before this study Implicit in characterizing severity in degenerative cervical myelopathy (DCM) is the use of the modified Japanese Orthopaedic Association (mJOA) score, which limits assessment to neurological dysfunction. This monocentric framework fails to reflect functional impact and quality of life, a limitation that has shaped how postoperative recovery in DCM has conventionally been modelled. We searched in PubMed for articles published since the database’s inception up to January 2026 using terms such as phenotype [“cervical myelopathy” OR “degenerative” OR “myelopathy”] AND [“subphenotype” OR “phenotype”] AND [“statistical models” OR “latent class analysis” OR “trajectory”]. We found that existing trajectory-based approaches rely predominantly on single outcome measures. A recent recommendation from an international DCM initiative advocates the combined assessment of neurological, disability, and quality-of-life outcomes as a core framework for evaluating postoperative recovery. Added value of this study Building on this recent progress, we examined whether the combined assessment of mJOA, neck disability index (NDI), and SF-36 Physical Component Score - core elements of the proposed DCM outcome set - has construct validity and captures clinically meaningful recovery heterogeneity in a real-world DCM cohort. We applied joint trajectory modelling to evaluate simultaneous change across these interrelated outcomes, representing a novel application in spine research. This analysis identified three distinct multidimensional recovery trajectories, with most patients demonstrating predominant improvement in neck pain disability at one year. Models that included these clusters were robust to discrimination and prediction of postoperative improvement. To facilitate clinical translation, we developed a practical tool to assist clinicians in identifying expected recovery trajectories and individualizing DCM management. Implications of all the available evidence Taken together with recent international recommendations, our study findings raise an important clinical question: whether DCM can continue to be meaningfully represented using a single outcome scale, or whether it should instead be recognised as a condition with inherently multidimensional domain involvement? Our results support the latter, demonstrating that combined neurological, functional, and quality-of-life assessment effectively characterises prognostic phenotypes in DCM. This multidimensional approach provides a valuable tool for subphenotyping in future clinical trials and treatment allocation of this common but debilitating condition. Nevertheless, external validation in independent DCM cohorts is required to confirm the reproducibility, generalizability, and clinical utility of these trajectory-derived phenotypes. Introduction Degenerative cervical myelopathy (DCM) is the most common cause of acquired, subacute spinal cord dysfunction worldwide, affecting ∼2% of adults.1 Clinical presentation varies widely, from mild paresthesia to severe multisystem deficits, and patients report some of the lowest health-related quality of life (HRQoL) scores among chronic neurological conditions.2 Surgical decompression is the only effective treatment to halt disease progression, but postoperative outcomes are highly variable, influenced by baseline neurological status, frailty, and surgical approach.3,4 Stratifying patients by recovery trajectories is therefore critical to enable personalized care and improve decision-making. The inherent heterogeneity in DCM limits the utility of traditional unidimensional assessment tools such as the modified Japanese Orthopaedic Association (mJOA) score.5 The mJOA does not capture pain, an outcome prioritised by patients, thus failing to fully reflect the recovery experience.6 To address this gap, the 2024 RECODE-DCM initiative endorsed a core outcome set comprising the mJOA, Neck Disability Index (NDI), and Short Form-36 (SF-36) Health Survey.7 Together, these measures allow for parallel assessment of neurological function, pain-related disability, and quality of life. This multidomain framework aligns with both contemporary spinal cord injury (SCI) research and the WHO’s International Classification of Functioning, Disability and Health (ICF), which models disease burden across impairment, activity limitation, and participation restriction.8 To date, no study has jointly modelled longitudinal changes in mJOA, NDI, and SF-36 in DCM. Existing work has focused on these outcomes independently, neglecting the interdependence and frequent discordance observed in clinical recovery.9, 10, 11 Moreover, univariate trajectory models assume within-class homogeneity, limiting their ability to detect divergent or non-linear recovery patterns across multiple domains.12 These oversimplifications may misclassify patients and reduce clinical applicability of trajectory phenotyping. Given the complex interplay of neurological, functional, and psychosocial factors influencing recovery in DCM, joint modelling approaches are needed to capture co-evolving recovery patterns.13 Bayesian consensus clustering (BCC) offers a robust solution by enabling simultaneous modelling of multiple longitudinal outcomes while accommodating discordance and temporal heterogeneity among outcomes. In this study, we apply BCC to identify distinct recovery phenotypes following surgical decompression for DCM and compare its performance with conventional trajectory modelling methods. Methods Study design and ethics This retrospective analysis utilised patient-level data from three major multicenter international studies on DCM: AO-Spine Cervical Spondylotic Myelopathy-North America (CSM-NA) (2005–2007),14 AO Spine Cervical Spondylotic Myelopathy – International (CSM-I) (2007–2011)15 and Efficacy of Riluzole in Surgical Treatment of Cervical Spondylotic Myelopathy (CSM-PROTECT) trial (2012–2017).16 These trials collectively enrolled 1047 adult patients with a uniform diagnosis of DCM who underwent surgical decompression. Details of the regions, number of participating sites, along with key inclusion and exclusion criteria for each cohort are provided in Supplementary Table S1. Clinical variables collected across all studies included age, sex, race, symptom duration, surgical approach, number of surgical levels, underlying pathology, and frailty status. Race was included as a demographic covariate because of its established association with specific DCM pathologies, particularly OPLL. Frailty was defined using the modified frailty index (mFI-5) with ≥2 defined as frail.17 This study was conducted in accordance with the ethical principles for medical research involving human subjects set forth by the World Medical Association (WMA) Declaration of Helsinki. Written informed consent was obtained from all recruited patients prior to study inclusion, and the study protocol was approved by local ethics committees (REB# 05-0769, REB# 07-0763, REB# 11-0548) at each participating medical center.14, 15, 16 Given that the analysis used anonymized, previously collected data without new patient encounter, additional ethical approval was not required. Outcomes Three core outcome measures were assessed at baseline, 6 months and 1 year postoperatively:•Modified Japanese Orthopaedic Association (mJOA) score (range 0–18; higher scores indicate better neurological function)•Neck Disability Index (NDI) (range: 0–50; expressed as percentage, with higher values indicating greater disability)•SF-36 Physical Component Score (SF-36 PCS) (range 0–100; higher scores indicate better physical health) •Modified Japanese Orthopaedic Association (mJOA) score (range 0–18; higher scores indicate better neurological function) Modified Japanese Orthopaedic Association (mJOA) score (range 0–18; higher scores indicate better neurological function) •Neck Disability Index (NDI) (range: 0–50; expressed as percentage, with higher values indicating greater disability) Neck Disability Index (NDI) (range: 0–50; expressed as percentage, with higher values indicating greater disability) •SF-36 Physical Component Score (SF-36 PCS) (range 0–100; higher scores indicate better physical health) SF-36 Physical Component Score (SF-36 PCS) (range 0–100; higher scores indicate better physical health) These validated metrics served as class-defining variables for the trajectory analysis and were selected to represent the key domains outlined in the 2024 AO RECODE-DCM core measurement set.7 The Spinal Adverse Events System (SAVES) classification system, although part of the RECODE-DCM set, was excluded from the trajectory analysis due to its categorical structure and focus on treatment-related complications rather than disease-specific recovery. Bayesian Consensus Clustering (BCC) To characterize heterogeneity in recovery across mJOA, NDI, and SF-36 PCS, we applied Bayesian Consensus Clustering (BCC) using the BCClong package in R (version 4.3.1).18 Unlike traditional mixture models, BCC allows each longitudinal outcome to form local clusters while enforcing alignment with a global patient-level trajectory class.19 Thus, BCC allows patients to cluster differently across outcomes while still aligning into an overall recovery pattern. This dual mechanism enables the identification of discordant or mixed recovery patterns frequently observed in DCM. Prior to analysis, data was standardized to zero mean and unit variance to improve numerical stability during model estimation. Models were tested across a range of cluster numbers (K = 2 to 8) with random initialization. We selected the optimal number of clusters using the mean adjusted adherence (MAA) index, a BCC-specific goodness-of-fit criterion.18 Model parameter estimation was performed using Gibbs sampling with a burn-in of 2000 iterations (maximum iteration = 12,000) and thinning interval of 5. Posterior cluster assignments were based on the most frequently assigned class across Markov chain Monte Carlo (MCMC) draws, and posterior probabilities were extracted to evaluate assignment certainty. Final trajectory membership was assigned using the maximum posterior probability criterion, whereby each patient was classified into the trajectory group with the highest posterior probability. Thus, each individual belonged to a single trajectory group. Posterior probability distributions were further examined to assess classification uncertainty and cluster separation. LOESS-smoothed trajectory plots were generated for each outcome to visually assess temporal trends and inter-cluster differences. We pooled individual patient-level data from the three source studies to maximize statistical power, improve stability of trajectory estimation, and reduce the risk of identifying spurious, dataset-specific recovery patterns from smaller individual cohorts. Comparison with other trajectory models We compared the performance of BCC against three commonly used longitudinal trajectory modelling approaches: (1) Group-based multi-trajectory modelling (GBMTM), (2) Latent class mixed model (LCMM), and (3) K-means longitudinal clustering. Each method was implemented using its respective R package (gbmt, lcmm, and kml3d). The optimal number of clusters for each was determined using the Bayesian information criterion (BIC), which provides better penalization for model complexity than AIC (Akaike Information Criterion).20 Posterior cluster probabilities were extracted to evaluate classification uncertainty. Classification tree analysis We employed a supervised machine learning approach using Classification and Regression Tree (CART) analysis to predict trajectory group membership from baseline variables. For the CART analysis, trajectory membership derived from the BCC model served as the outcome variable. A full decision tree was initially generated using recursive partitioning, followed by pruning based on the complexity parameter (α) that minimized cross-validated (CV) misclassification error (five-fold CV). Statistical analysis Baseline clinical and demographic variables not used to define trajectory classes were compared across clusters using ANOVA for continuous variables and chi-square tests for categorical variables. We used logistic regression to examine the association between trajectory membership and achievement of the minimal clinically important difference (MCID) in mJOA, SF-36 PCS, and NDI at one year. Thresholds for MCID were based on established severity-specific cutoffs for mild, moderate and severe DCM, respectively: ≥1, 2, and 3 for mJOA; ≥4.8, 8.4, and 14.8 points for SF-36 PCS; and ≥10.5, 15, and 17.5 for NDI.21,22 Two multivariable logit models were developed to evaluate the incremental value of trajectory phenotypes in predicting MCID: i) a full model that included BCC-derived trajectory class along with established prognostic factors (age, sex, smoking, symptom duration, frailty, and baseline function) and ii) a reduced model excluding the BCC class. Model performance was internally validated using bootstrap resampling (100 iterations) through the validate function in the rms package, estimating the optimism-adjusted discrimination index (Dxy), explained variance (R2), Harrell’s g-index, and Brier score. Multinomial logistic regression was also used to identify predictors of trajectory group membership, modelling the probability of belonging to each trajectory class relative to a reference group, while adjusting for clinically relevant covariates. Missing outcome data at follow-up were addressed using multiple imputation under the assumption of missing at random (MAR) (See Supplementary Fig. S1 and Table S2). All analyses were conducted using R version 4.3.1 (R Foundation for Statistical Computing, Vienna, Austria). A two-sided p-value < 0.05 was considered statistically significant for all post hoc comparisons. Role of the funding source The present work did not receive any funding. The original studies were funded by AO Spine North America.14, 15, 16 For the original studies, the funder did not have any input regarding study design, data collection, data analyses, data interpretation, or writing of the report. Results Population characteristics The final analytical cohort consisted of 1047 patients diagnosed with DCM (Fig. 1). The mean age was 56.8 ± 11.4 years, and 60.7% were male. A quarter of patients reported a history of smoking (25%). The mean symptom duration was 29.8 months, and most patients exhibited involvement of two spinal levels (Supplementary Table S3). Baseline scores indicated moderate neurological impairment and considerable functional limitation, with mean mJOA, NDI and SF-36 PCS scores of 12.4, 39.1, and 35.6, respectively.Fig. 1Flowchart of study participants. Flowchart of study participants. BCC-derived trajectory patterns The BCC algorithm identified three distinct postoperative recovery phenotypes based on longitudinal trends across mJOA, NDI, and SF-36 PCS scores (Table 1). Trajectories were modelled using nonlinear functions to reflect the observed patterns over limited time points (Supplementary Fig. S2). Model performance was optimized at K = 3 clusters, which achieved the highest mean adjusted adherence score (MAA = 0.70; Supplementary Fig. S3). These trajectories were characterized as follows (Fig. 2):Trajectory 1: Minimal Recovery/Non-Responder (n = 177, 17%)This group has the lowest baseline neurological function and quality-of-life scores and demonstrated only limited postoperative improvement. Mean mJOA increased marginally from 11.03 (±2.34) to 12.12 (±2.62) at one year. NDI remained elevated, and SF-36 PCS improved only slightly (28.29–28.33).Trajectory 2: Neck Pain Responder (n = 524, 50%)The largest subgroup experienced significant improvement in neck-specific disability. NDI decreased by 18 points [44.37 (±17.76) to 25.93 (±17.39)], accompanied by moderate improvements in mJOA and SF-36 PCS scores.Trajectory 3: Global Responder (n = 346, 33%)This phenotype demonstrated the most robust gains across all domains. Mean mJOA increased from 14.03 to 16.58, NDI decreased from 25.35 to 15.26, and SF-36 PCS increased from 43.50 to 48.87 at one year.Table 1Comparison of baseline, six-month, and one-year scores for mJOA, NDI, and SF-36 PCS across DCM trajectories.aTrajectory 1 (n = 177)Trajectory 2 (n = 524)Trajectory 3 (n = 346)Total (n = 1047)Baseline scores mJOA11.03 (2.33)11.64 (2.34)14.02 (2.076)12.39 (2.56) NDI53.88 (19.91)44.37 (17.76)25.35 (16.734)39.08 (20.61) SF-36 PCS28.29 (6.33)32.20 (6.65)43.50 (7.988)35.59 (9.25)Six-month scores mJOA12.22 (2.50)14.33 (2.50)16.23 (1.87)14.70 (2.65) NDI48.25 (18.45)30.07 (17.65)18.03 (15.86)28.43 (19.71) SF-36 PCS29.90 (8.15)39.42 (8.04)48.20 (7.13)41.14 (9.83)One-year scores mJOA12.12 (2.62)15.03 (2.26)16.58 (1.86)15.16 (2.60) NDI50.22 (18.67)25.93 (17.39)15.26 (14.35)25.62 (19.93) SF-36 PCS28.33 (5.98)42.17 (8.33)48.87 (7.36)42.56 (10.08)Change from baseline to 6 months (Δ)b mJOA+1.19 (2.53)+2.69 (2.85)+2.21 (2.48)+2.31 (2.73) NDI−5.63 (17.13)−14.30 (21.09)−7.32 (18.99)−10.65 (20.19) SF-36 PCS+1.61 (8.46)+7.21 (10.0)+4.70 (9.91)+5.55 (9.96)Change from baseline to 1 year (Δ)b mJOA+1.09 (2.66)+3.39 (3.01)+2.55 (2.46)+2.77 (2.89) NDI−3.66 (17.45)−18.44 (22.14)−10.09 (19.39)−13.46 (21.30) SF-36 PCS+0.03 (7.05)+9.96 (10.73)+5.37 (10.72)+6.97 (10.86)mJOA, modified Japanese Orthopaedic Association score; NDI, Neck Disability Index; PCS, Physical Component Summary; SD, standard deviation.aMean (SD).bPositive changes in mJOA and SF-36 PCS indicate improvement, whereas negative changes in NDI indicate improvement.Fig. 2Bayesian consensus clustering trajectories of degenerative cervical myelopathy (DCM) Recovery Based on Joint Modelling of modified Japanese Orthopaedic Association (mJOA) score, Neck Disability Index (NDI), and SF-36 physical component score (SF-36 PCS). Trajectory 1: Minimal Recovery/Non-Responder (n = 177, 17%)This group has the lowest baseline neurological function and quality-of-life scores and demonstrated only limited postoperative improvement. Mean mJOA increased marginally from 11.03 (±2.34) to 12.12 (±2.62) at one year. NDI remained elevated, and SF-36 PCS improved only slightly (28.29–28.33). Trajectory 1: Minimal Recovery/Non-Responder (n = 177, 17%)This group has the lowest baseline neurological function and quality-of-life scores and demonstrated only limited postoperative improvement. Mean mJOA increased marginally from 11.03 (±2.34) to 12.12 (±2.62) at one year. NDI remained elevated, and SF-36 PCS improved only slightly (28.29–28.33). This group has the lowest baseline neurological function and quality-of-life scores and demonstrated only limited postoperative improvement. Mean mJOA increased marginally from 11.03 (±2.34) to 12.12 (±2.62) at one year. NDI remained elevated, and SF-36 PCS improved only slightly (28.29–28.33). This group has the lowest baseline neurological function and quality-of-life scores and demonstrated only limited postoperative improvement. Mean mJOA increased marginally from 11.03 (±2.34) to 12.12 (±2.62) at one year. NDI remained elevated, and SF-36 PCS improved only slightly (28.29–28.33). Trajectory 2: Neck Pain Responder (n = 524, 50%)The largest subgroup experienced significant improvement in neck-specific disability. NDI decreased by 18 points [44.37 (±17.76) to 25.93 (±17.39)], accompanied by moderate improvements in mJOA and SF-36 PCS scores. Trajectory 2: Neck Pain Responder (n = 524, 50%)The largest subgroup experienced significant improvement in neck-specific disability. NDI decreased by 18 points [44.37 (±17.76) to 25.93 (±17.39)], accompanied by moderate improvements in mJOA and SF-36 PCS scores. The largest subgroup experienced significant improvement in neck-specific disability. NDI decreased by 18 points [44.37 (±17.76) to 25.93 (±17.39)], accompanied by moderate improvements in mJOA and SF-36 PCS scores. The largest subgroup experienced significant improvement in neck-specific disability. NDI decreased by 18 points [44.37 (±17.76) to 25.93 (±17.39)], accompanied by moderate improvements in mJOA and SF-36 PCS scores. Trajectory 3: Global Responder (n = 346, 33%)This phenotype demonstrated the most robust gains across all domains. Mean mJOA increased from 14.03 to 16.58, NDI decreased from 25.35 to 15.26, and SF-36 PCS increased from 43.50 to 48.87 at one year. Trajectory 3: Global Responder (n = 346, 33%)This phenotype demonstrated the most robust gains across all domains. Mean mJOA increased from 14.03 to 16.58, NDI decreased from 25.35 to 15.26, and SF-36 PCS increased from 43.50 to 48.87 at one year. This phenotype demonstrated the most robust gains across all domains. Mean mJOA increased from 14.03 to 16.58, NDI decreased from 25.35 to 15.26, and SF-36 PCS increased from 43.50 to 48.87 at one year. This phenotype demonstrated the most robust gains across all domains. Mean mJOA increased from 14.03 to 16.58, NDI decreased from 25.35 to 15.26, and SF-36 PCS increased from 43.50 to 48.87 at one year. Table 1Comparison of baseline, six-month, and one-year scores for mJOA, NDI, and SF-36 PCS across DCM trajectories.aTrajectory 1 (n = 177)Trajectory 2 (n = 524)Trajectory 3 (n = 346)Total (n = 1047)Baseline scores mJOA11.03 (2.33)11.64 (2.34)14.02 (2.076)12.39 (2.56) NDI53.88 (19.91)44.37 (17.76)25.35 (16.734)39.08 (20.61) SF-36 PCS28.29 (6.33)32.20 (6.65)43.50 (7.988)35.59 (9.25)Six-month scores mJOA12.22 (2.50)14.33 (2.50)16.23 (1.87)14.70 (2.65) NDI48.25 (18.45)30.07 (17.65)18.03 (15.86)28.43 (19.71) SF-36 PCS29.90 (8.15)39.42 (8.04)48.20 (7.13)41.14 (9.83)One-year scores mJOA12.12 (2.62)15.03 (2.26)16.58 (1.86)15.16 (2.60) NDI50.22 (18.67)25.93 (17.39)15.26 (14.35)25.62 (19.93) SF-36 PCS28.33 (5.98)42.17 (8.33)48.87 (7.36)42.56 (10.08)Change from baseline to 6 months (Δ)b mJOA+1.19 (2.53)+2.69 (2.85)+2.21 (2.48)+2.31 (2.73) NDI−5.63 (17.13)−14.30 (21.09)−7.32 (18.99)−10.65 (20.19) SF-36 PCS+1.61 (8.46)+7.21 (10.0)+4.70 (9.91)+5.55 (9.96)Change from baseline to 1 year (Δ)b mJOA+1.09 (2.66)+3.39 (3.01)+2.55 (2.46)+2.77 (2.89) NDI−3.66 (17.45)−18.44 (22.14)−10.09 (19.39)−13.46 (21.30) SF-36 PCS+0.03 (7.05)+9.96 (10.73)+5.37 (10.72)+6.97 (10.86)mJOA, modified Japanese Orthopaedic Association score; NDI, Neck Disability Index; PCS, Physical Component Summary; SD, standard deviation.aMean (SD).bPositive changes in mJOA and SF-36 PCS indicate improvement, whereas negative changes in NDI indicate improvement. Comparison of baseline, six-month, and one-year scores for mJOA, NDI, and SF-36 PCS across DCM trajectories.a mJOA, modified Japanese Orthopaedic Association score; NDI, Neck Disability Index; PCS, Physical Component Summary; SD, standard deviation. Mean (SD). Positive changes in mJOA and SF-36 PCS indicate improvement, whereas negative changes in NDI indicate improvement. Bayesian consensus clustering trajectories of degenerative cervical myelopathy (DCM) Recovery Based on Joint Modelling of modified Japanese Orthopaedic Association (mJOA) score, Neck Disability Index (NDI), and SF-36 physical component score (SF-36 PCS). Clinical characterization of BCC-derived recovery phenotypes In post hoc analysis, the three trajectories demonstrated distinct baseline profiles (Supplementary Table S3). Patients in Trajectory 1 (Minimal Recovery) exhibited the greatest disease burden, characterized by the highest prevalence of severe DCM (60.5%), comorbidities (78.4%), smoking history (32.2%), and subluxation (9.6%). In contrast, Trajectory 3 (Global Responder) represented a clinically distinct subgroup comprising younger patients with lower comorbidity burden, and higher education. Notably, almost 40% of patients in Trajectory 3 presented with mild DCM, compared with only 11.5% and 4.0% in Trajectories 2 and 1, respectively. Although both Trajectories 2 and 3 demonstrated postoperative improvement, Trajectory 2 was characterized predominantly by moderate-to-severe DCM (88.4%) (Supplementary Fig. S10). Comparison with alternative trajectory models Model comparison demonstrated differences in the number and consistency of identified clusters. Both BCC and LCMM (Supplementary Fig. S4) detected three trajectory groups, whereas GBMTM yielded eight (Supplementary Fig. S5) and k-means suggested two groups (Supplementary Fig. S6). BCC exhibited the lowest clustering uncertainty, with most patients assigned to trajectories with posterior probabilities close to 1. In contrast, LCMM and GBMTM showed greater uncertainty, particularly GBMTM due to its higher cluster count. Clinical relevance of BCC-derived trajectories Multivariable logistic regression confirmed that trajectory membership was a strong predictor of achieving the MCID for each outcome measure at one year. Likelihood ratio (LR) tests demonstrated that inclusion of BCC classes significantly improved model fit for MCID prediction in mJOA (LR χ2 = 75.68, df = 2, p < 0.0001), SF-36 PCS (LR χ2 = 80.38, df = 2, p < 0.0001) and NDI (LR χ2 = 57.69, df = 2, p < 0.0001). Patients in Trajectory 2 had nearly 3-fold greater odds of achieving MCID for mJOA (OR = 3.10, 95% CI: 1.97–4.87, p < 0.0001), while those in Trajectory 3 had even higher odds (OR:10.90, 95% CI: 6.15–19.32, p < 0.0001) compared to patients in Trajectory 1 (Fig. 3A). Similarly, for SF-36 PCS, odds of MCID achievement were significantly higher in Trajectories 2 (OR = 11.41, 95% CI 5.65–23.04) and 3 (OR = 7.32, 95% CI 3.53–15.22) compared to patients in the first group (Fig. 3C). For NDI, patients in Trajectory 2 had the highest likelihood of improvement (OR = 4.71, 95% CI 3.03–7.31), followed by Trajectory 3 (OR = 3.02, 95% CI 1.85–4.93), both with reference to Trajectory 1 (Fig. 3B).Fig. 3Predicted probabilities derived from logistic regression models stratified by recovery trajectory class. Probability of achieving minimal clinically important difference (MCID) in (A) modified Japanese Orthopaedic Association (mJOA) score; (B) neck disability index (NDI); (C) SF-36 physical component score (SF-36 PCS). Predicted probabilities derived from logistic regression models stratified by recovery trajectory class. Probability of achieving minimal clinically important difference (MCID) in (A) modified Japanese Orthopaedic Association (mJOA) score; (B) neck disability index (NDI); (C) SF-36 physical component score (SF-36 PCS). Predictive performance and internal validation Inclusion of trajectory class improved the predictive performance of all models. For mJOA MCID, the full model with BCC classes achieved higher discrimination [corrected Dxy 0.52 (0.44–0.59) vs. 0.36 (0.29–0.43)] and explained variance [corrected R2 0.21 (0.14–0.27) vs. 0.11 (0.06–0.15)] with lower prediction error [Brier score 0.14 (0.13–0.15) vs. 0.15 (0.13–0.16)] compared to the reduced model (Supplementary Materials Tables S4 and S5). Similarly, for SF-36 PCS MCID, the model including trajectory classes yielded better discrimination [corrected Dxy 0.35 (0.29–0.41) vs. 0.21 (0.13–0.27)], greater explained variance [corrected R2 0.14 (0.10–0.18) vs. 0.04 (−0.001 to 0.07)], and slightly lower prediction error [Brier score 0.21 (0.20–0.22) vs. 0.22 (0.21–0.24)] (Supplementary Materials Tables S6 and S7). For NDI MCID, Dxy increased from 0.12 (0.05–0.19) to 0.27 (0.20–0.34), R2 from 0.01 (−0.02 to 0.04) to 0.08 (0.03–0.12) and Brier score improved from 0.25 to 0.23, with better discrimination [g = 0.60 (0.42–0.76) vs. 0.25 (0.11–0.38)] (Supplementary Materials Tables S8 and S9) Calibration plots demonstrated strong agreement between predicted probabilities and observed outcomes (Supplementary Materials Figs. S7–S9). Predictors of trajectory membership Multinomial logistic regression identified several baseline predictors of class membership (Supplementary Table S10). Compared to Trajectory 1, posterior surgical approach was associated with higher odds of belonging to Trajectory 2 (OR = 2.40; 95% CI, 1.03–5.56; p = 0.041). In contrast, higher frailty scores (OR = 1.42; 95% CI 1.14–1.76; p = 0.002) and anterior surgery (OR = 3.55; 95% CI 1.21–10.37; p = 0.021) were associated with increased odds of belonging to Trajectory 3 relative to Trajectory 1. CART-based stratification The pruned CART model identified early clinical thresholds predictive of trajectory class (Fig. 4). Only baseline SF36, NDI and mJOA scores improved node purity and were retained in the final tree. The first split occurred at SF-PCS = 41, with 26% of patients, who are above this threshold, assigned to Trajectory 3. Below 41, NDI ≥65 was associated with Trajectory 1 membership, while NDI<65 and mJOA <15 were indicative of Trajectory 2.Fig. 4Pruned classification and regression tree (CART) predicting trajectory membership using baseline symptom scores. The proportions displayed in each node represent the model-estimated class probabilities for each class and the percentage value represent the sample share at each node. Modified Japanese Orthopaedic Association (mJOA) score; Neck Disability Index (NDI); and SF-36 physical component score (SF36 PCS). Pruned classification and regression tree (CART) predicting trajectory membership using baseline symptom scores. The proportions displayed in each node represent the model-estimated class probabilities for each class and the percentage value represent the sample share at each node. Modified Japanese Orthopaedic Association (mJOA) score; Neck Disability Index (NDI); and SF-36 physical component score (SF36 PCS). Discussion In this study, Bayesian Consensus Clustering (BCC) was applied to jointly model longitudinal changes in neurological function (mJOA), neck-related disability (NDI), and physical quality of life (SF-36 PCS) among patients undergoing surgical treatment for DCM. This multidimensional approach identified three distinct recovery trajectories: (i) minimal responders (17%) with persistently poor outcomes; (ii) neck pain responders (50%) with substantial NDI improvements but modest gains in neurological function; and (iii) global responders (33%) with consistent improvement across all endpoints. These phenotypes differed in clinical features and in the proportion of patients achieving MCID at one year, underscoring the added value of multidimensional trajectory-based phenotyping over single-domain outcome assessment. Our findings reflect actual clinical experience and existing literature indicating that recovery in DCM is often discordant across neurological, functional, and quality-of-life domains.23, 24, 25 For example, patients in Trajectory 2 had significant NDI improvement despite only moderate mJOA gains, a pattern that would have remained undetected if the analysis focused solely on mJOA. The divergent predicted probability patterns for NDI and SF-36 PCS reflect differences in the constructs measured by these instruments. The probability of achieving NDI MCID decreased with increasing baseline mJOA, potentially because patients with better neurological function had lower baseline neck disability and therefore less capacity to achieve large absolute reductions in NDI. Conversely, the probability of achieving SF-36 PCS MCID increased with higher baseline mJOA, suggesting that patients with better preserved neurological function may have greater physiological reserve and functional capacity to realize broader physical health gains after surgery. Prior studies, such as Karim et al. (CSORN),24 and Swespine registry analyses, similarly reported discrepancies between neurological improvement and patient-reported outcomes.25 These findings emphasize that mJOA, NDI, and SF-36 PCS capture complementary but distinct domains of recovery and support the use of multidimensional outcome assessment in DCM. By integrating local and global clustering structures, BCC effectively uncovered such latent, discordant recovery profiles, providing a more nuanced understanding of recovery heterogeneity in DCM. The distinct recovery pattern in Trajectory 2 underscores the central role of neck pain in shaping perceived recovery. Although both Trajectories 2 and 3 demonstrated gains in PCS and mJOA, BCC identified them as distinct, emphasizing that pain relief alone may drive meaningful HRQoL improvement. Incorporating BCC-derived phenotypes into prognostic models significantly improved performance across multiple metrics, including optimism-corrected Dxy, Harrell’s g-index, and Brier scores, confirming the added prognostic value of phenotyping based on joint outcome profiles. Our work operationalizes the 2024 RECODE-DCM recommendation for multidimensional outcome assessment.7 Previous studies have primarily relied on single-outcome modelling. For example, Jaja et al. and Bak et al. used group-based trajectory modelling (GBTM) and latent growth curve analysis (LGCA), respectively, to delineate recovery phenotypes based on single outcomes such as mJOA or SF-36 PCS.9,10 While informative, these approaches could not identify convergent or discordant recovery patterns across multiple domains. A recent LGCA-based analysis11 incorporated NDI, but was limited by single-centre design and smaller sample size. To our knowledge, this is the first study to apply BCC for joint longitudinal trajectory modelling in DCM, building on our prior work on symptom clustering26 and offering a robust data-driven strategy for prognostic phenotype discovery. Our analysis also show that mJOA-defined “mild” DCM is not a homogeneous entity. Patients with the same baseline mJOA score followed markedly different recovery courses depending on concurrent changes in NDI and PCS. For example, two individuals with baseline mJOA of 15 (mild DCM) may end up in either minimal or global responder groups. This observation underscores the fact that DCM severity defined solely using mJOA is insufficient to characterize postoperative recovery because patients with similar baseline neurological impairment may follow divergent multidimensional recovery trajectories. To facilitate translation, we developed a decision-tree model using baseline predictors to classify patients into recovery trajectories. Frailty and surgical approach emerged as significant predictors of recovery phenotype. Frailty was associated with the global responder group, possibly reflecting greater symptom burden and room for improvement, consistent with prior studies showing that frail patients can achieve substantial postoperative gains.17 Anterior surgical approach was also predictive of better outcomes, potentially due to improved pain relief and reduced soft tissue disruption compared to posterior techniques.27,28 This observation should be interpreted cautiously, however, given the retrospective nature of the study. The contribution of sagittal alignment parameters and the extent to which they may represent a source of residual confounding could not be determined from the present dataset, as these variables were not systematically collected. Given the established relationship between sagittal balance and postoperative outcomes in DCM, incorporation of radiographic and biomechanical parameters into future phenotyping studies may further refine recovery trajectories. Nevertheless, these findings suggest that surgical strategy can influence not just neurological recovery, but the overall multidimensional recovery phenotype. Compared to alternative methods, BCC offered the best balance of interpretability and parsimony. GBTM generated eight overlapping clusters with limited clinical utility. LCMM and BCC both produced three-class solutions, but only BCC integrated local and global clustering to capture discordant trajectories.18 K-means clustering performed poorly, consistent with its inability to model temporal structure or account for measurement uncertainty.19 These results have implications for both clinical practice and future research. Phenotype-informed classification may enable more tailored postoperative care. For example, patients in the “Neck Pain Responder” group may benefit from targeted pain management or psychosocial interventions, even if neurological recovery is favourable. In clinical trials, stratifying participants by recovery trajectory could improve sensitivity to detect treatment effects by ensuring adequate representation of recovery subtypes, especially among global responders. Similar approaches have been used in spinal cord injury research, where tools like URP-CTREE (unbiased recursive partitioning) have helped exclude patients with low recovery potential.29 Several limitations warrant consideration. The identified recovery trajectories were derived exclusively from surgically treated patients, a majority of which were enrolled in North American prospective studies and may not be directly generalizable to non-operative populations or healthcare systems with different patient characteristics and treatment practices. Although neck pain emerged as a major recovery determinant, our dataset lacked detailed information regarding its etiology (e.g., axial, radicular, or central sensitization), postoperative analgesic use, rehabilitation protocol and its association with recovery of physical function. These factors could have contributed to residual confounding, particularly in Trajectory 2 patients. Incorporating these variables in future studies may further improve characterization of this phenotype. While longer-term data (e.g., 2 years) might capture different recovery patterns, we limited follow-up to one year to minimize missing data, which can introduce instability into trajectory modelling. Postoperative complications were not explicitly incorporated, though they likely impact recovery. Additionally, the anatomical, radiographic, and biomechanical factors influencing postoperative outcomes such as sagittal malalignment, paraspinal muscle atrophy (myofascial component), or incomplete decompression were not fully captured by our dataset. Future studies incorporating these factors may provide a more holistic picture of recovery phenotypes and improve prognostic precision. Nonetheless, this study leveraged the largest available multicenter dataset of prospectively enrolled DCM patients, providing a rigorous dataset for phenotypic exploration. Future studies should externally validate these findings in independent multicentre cohorts to determine the robustness and generalizability of the identified multidimensional recovery trajectories. In conclusion, recovery following surgery for DCM is heterogeneous and better captured through multidimensional profiling. Using BCC, we identified three recovery trajectories based on joint changes in neurological, functional, and quality-of-life measures. Compared to models using only baseline predictors, integration of BCC-derived phenotypes improved outcome prediction. These findings support the move beyond unidimensional scales such as mJOA and advocate for routine multidimensional assessment in both clinical care and research. Joint trajectory modelling provides a robust framework for refining prognostication and enabling personalized treatment strategies in DCM. Contributors KMP, MAA, MGF designed the study and wrote the manuscript. KMP, MAA, MGF accessed and verified the underlying data. KMP, MAA extracted and collected the data; KMP MAA, CL, VK, AB analysed data. KMP and MAA prepared the tables and figures. MGF, BMD, JSH, JHB, JRW helped with acquisition, funding and supervision of the study. WL, ZL, and KET provided statistical consult. All authors reviewed the results and interpreted data. All authors have made an intellectual contribution to the manuscript and approved the submission. Data sharing statement The datasets generated and analysed during the current study are available from the corresponding author on reasonable request. The code used in this article can be obtained by contacting the corresponding author, Michael G. Fehlings. Declaration of interests We declare no relevant competing interests. JSH discloses that he is on the Advisory Board for Medtronic, Depuy Spine and Ethicon (nothing specific to this topic).