Motor-dominant symptoms predict persistent neurological impairment in mild-to-moderate degenerative cervical myelopathy: a multi-state modeling study

Abstract

To characterize symptom domain-defined trajectories in mild-to-moderate degenerative cervical myelopathy (DCM) during scheduled non-operative follow-up, quantify transition probabilities and sojourn durations using a continuous-time multi-state Markov (MSM) model, and explore latent clinical heterogeneity through latent class analysis (LCA). This study is a secondary analysis of an existing multicenter randomized, double-blind trial dataset and does not evaluate comparative treatment effects. We conducted a secondary analysis of longitudinal data from a multicenter randomized, double-blind trial dataset (ChiCTR-INR-16009723) involving adults with mild-to-moderate DCM in the archived analytic dataset and complete Japanese Orthopaedic Association (JOA) domain data. The present analysis was restricted to the scheduled non-operative follow-up period and was designed to describe symptom-domain trajectories rather than treatment efficacy. Neurological status was classified into six mutually exclusive states based on JOA domains: asymptomatic; sensory-only impairment; isolated upper-extremity motor deficit; isolated lower-extremity motor deficit; combined upper- and lower-extremity motor deficits with normal bladder function; and sphincter involvement. Visits used in the present analysis took place at baseline; weeks 1, 2, 4, 6, and 8; and months 3 and 6. Transition intensities, 1-year transition probabilities, and mean sojourn lengths were estimated using a continuous-time MSM. Covariate effects (sex, age group, BMI category, prior conservative treatment, comorbidities) were modeled only for transitions meeting the prespecified ≥ 5-event threshold. LCA was performed using repeated state indicators across the scheduled visits to identify data-driven trajectory classes. This secondary analysis included only participants who remained on non-operative management throughout the analyzed follow-up period; patients who underwent surgery during treatment or follow-up were excluded at the data-screening stage. Ninety patients (mean age 58.6 years; 61.1% male) exhibited 71 observable state shifts, with the majority of assessments indicating stability. Sensory-predominant states exhibited considerable dynamism, characterized by frequent bidirectional transitions between asymptomatic and sensory-only states, with very brief mean sojourn durations (~ 0.25–0.31 years). Conversely, motor-dominant phenotypes exhibited enhanced short-term persistence; isolated lower-extremity motor deficits revealed the longest mean sojourn time (~ 2.76 years) and the highest one-year persistence (~ 71.9%). The advancement to more severe conditions over one year was infrequent (< 11%), but a significant proportion exhibited improvement from combined motor impairment to less severe stages. Covariate effects were largely unclear owing to limited occurrences; however, prior medication history and elevated BMI exhibited the most robust relationships with certain transitions. LCA supported a three-class solution (entropy ~ 0.99) including diverse longitudinal severity patterns, roughly differentiating between milder/fluctuating and more motor-dominant, more persistently severe trajectories. In non-operatively managed mild-to-moderate DCM, the clinical course during scheduled follow-up appeared non-linear and phenotype-dependent: sensory-only impairment tended to fluctuate with brief sojourn times, whereas motor deficits, especially lower-extremity motor deficits, were more persistent and showed limited spontaneous return to asymptomatic status. These exploratory MSM and LCA findings may help identify patients who warrant closer monitoring, but they should not be interpreted as establishing surgical timing thresholds or proving benefit from earlier intervention.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-52792-1.

Introduction

Degenerative cervical myelopathy (DCM) is a chronic, non-traumatic spinal cord disorder caused by age-related degenerative changes of the cervical spine, including disc degeneration, osteophyte formation, ligamentous hypertrophy, and related canal narrowing¹. Common manifestations include sensory disturbance, hand clumsiness, gait dysfunction, and bladder symptoms². Current clinical practice guidelines generally recommend surgical intervention for patients with moderate-to-severe DCM, whereas management in mild DCM remains more heterogeneous³. Nonetheless, uncertainty remains greatest for patients with mild DCM and for selected patients managed initially without surgery⁴.

Numerous studies have demonstrated that conservative treatment exhibits some efficacy and may prevent the necessity for surgical intervention in patients with mild DCM^5,6. A systematic review indicated that 20% to 62% of patients with symptomatic myelopathy undergoing conservative management experienced a decline of at least one point on the Japanese Orthopaedic Association (JOA) score within three to six years⁷. Moreover, conservative treatment frequently does not meet the minimum clinically important difference (MCID) across multiple clinical parameters, and the incidence of hospitalisation due to spinal cord injury is reported to be greater in patients receiving conservative treatment than in those undergoing surgical intervention^5,8,9. Fehlings et al. reported substantial postoperative enhancement in patients with mild DCM relative to their preoperative condition; however, the degree of improvement in modified JOA (mJOA) scores was less significant than that seen in patients with moderate to severe DCM¹⁰. Furthermore, although the mean score alteration in patients with moderate to severe DCM met the MCID, numerous investigations have indicated that the mean score change in patients with mild DCM failed to meet the MCID^11–14. The results are ascribed to the ceiling effect of the mJOA score^10,15, and now, no standardised approach is available for assessing surgical outcomes in individuals with moderate DCM.

According to these findings, current clinical practice guidelines suggest either surgical decompression or a supervised trial of structured rehabilitation for selected patients with mild DCM³. However, most prior studies have summarized outcomes using global functional scales and have offered limited insight into domain-specific symptom evolution under non-operative care^10–12. Importantly, the present work is a secondary, post hoc analysis of a randomized trial that was originally designed to evaluate Yishenyangsui granule rather than to compare conservative versus surgical treatment strategies. Accordingly, our aim is to characterize the natural history of symptom-domain trajectories during the non-operative follow-up period, not to infer causal differences between treatment options.

In recent years, multi-state Markov (MSM) models have been used in chronic neurological and cardiovascular disorders to study movement between clinically meaningful states^16,17. This framework can estimate how quickly patients leave a given state, how long they tend to remain there, and the probability of being in another state after a specified interval. In the present exploratory secondary analysis of a multicenter randomized, double-blind trial cohort, we applied an MSM to symptom-domain-defined states in mild-to-moderate DCM during scheduled non-operative follow-up. We aimed to characterize disease fluctuation and persistence, estimate state-specific transition probabilities and sojourn times, and examine selected baseline covariates only where the observed number of transitions was sufficient to support stable estimation. We additionally used latent class analysis (LCA) based on repeated state indicators to summarize longitudinal heterogeneity into data-driven phenotypes. Because the cohort was entirely analyzed during non-operative follow-up, the study informs short-term natural history under conservative management and not the comparative effects of early versus delayed surgery.

Methods

Sources of data and criteria for participant eligibility

Patients with DCM who received non-surgical interventions from 2016 to 2017 and had complete Japanese Orthopaedic Association (JOA) score domain data were identified from the archived dataset of a multicenter randomized, double-blind trial conducted at three Chinese centers (Chinese ClinicalTrials.gov identifier: ChiCTR-INR-16009723)¹⁸. The parent trial was conducted in accordance with the Declaration of Helsinki and was approved by the ethics committee of Wangjing Hospital, China Academy of Chinese Medical Sciences (WJEC-KT-2016-004-P001). The trial was registered at the Chinese Clinical Trials Registry (ChiCTR-INR-16009723), and written informed consent was obtained from all participants in the parent trial before enrollment. The present study was a secondary post hoc analysis of de-identified trial data restricted to the non-operative follow-up period. No additional informed consent was required for this secondary analysis of de-identified trial data, in accordance with institutional requirements. Missing baseline values were resolved using multiple imputation via chained equations. Longitudinal outcome data were not imputed; models were generated under a missing-at-random assumption using maximum likelihood. The analytic panel was defined by the protocol-specified non-operative visits available in the final dataset, spanning baseline through 6 months. Patients who underwent surgical intervention during treatment or follow-up were excluded at the data-screening stage; therefore, the present analytic cohort comprised only individuals who remained under non-operative management throughout the analyzed follow-up period (Table 1).

Table 1.

Baseline characteristics of enrolled individuals (N = 90).

Characteristic (n, %)	All Patients
Age, years	58.6 ± 8.6
40–60 years	47 (52.2%)
61–80 years	43 (47.8%)
Sex
Male	55 (61.1%)
Female	35 (38.9%)
Height, cm	167.1 ± 6.9
Weight, kg	69.6 ± 12.3
BMI, kg/m2	24.9 ± 3.7
Normal weight (18.5–23.9)	41 (45.6%)
Overweight/Obese (≥ 24)	49 (54.4%)
Occupation type
High-frequency static flexion	34 (37.8%)
Moderate dynamic flexion	20 (22.2%)
Low-frequency/non-fixed posture	36 (40.0%)
Other/undetermined	0 (0%)
Any previous conservative treatment	60 (66.7%)
Pharmacotherapy (medications)	49 (54.4%)
Manual therapy	34 (37.8%)
Acupuncture	15 (16.7%)
Traction therapy	31 (34.4%)
Physical therapy	34 (37.8%)
Any comorbidity (HTN, diabetes, or urolithiasis)	6 (6.7%)

Measurement and categorization of symptom states

Neurological states were delineated using domain-specific elements of the Japanese Orthopaedic Association (JOA) score, encompassing upper- and lower-extremity motor function, sensory function, and bladder control. Rather than relying on the overall JOA score, individual domains were used to create hierarchical categories reflecting the anatomical extent of spinal cord involvement. Evaluations included baseline and follow-up assessments at 1, 2, 4, 6, and 8 weeks, as well as 3 and 6 months. Motor function in the upper and lower extremities was considered normal at 4 and abnormal at < 4. Bladder function was considered normal at 3 and abnormal at < 3. Sensory function (upper extremities, lower extremities, trunk) was considered normal at 2 and abnormal at < 2. Patients were categorized into six mutually exclusive states: (1) asymptomatic (all domains normal); (2) sensory-only impairment (≥ 1 sensory abnormality with normal motor and bladder function); (3) isolated upper-extremity motor deficit; (4) isolated lower-extremity motor deficit; (5) combined upper- and lower-extremity motor deficits with normal bladder function; and (6) sphincter involvement (abnormal bladder function irrespective of other domains). In the presence of multiple deficits, patients were assigned to the highest relevant category, with sphincter dysfunction representing the most severe level. State 6 was prespecified to be merged with State 5 only if it represented < 5% of the cohort to preserve model stability; this threshold was not crossed in the analytic sample, so all six states were retained (Table 2).

Table 2.

Symptom-domain state definitions used in the multi-state model.

State	JOA domain pattern	Clinical description
1	All domains normal	Asymptomatic
2	≥ 1 sensory abnormality; motor and bladder normal	Sensory-only impairment
3	Upper-extremity motor abnormality only	Isolated upper-extremity motor deficit
4	Lower-extremity motor abnormality only	Isolated lower-extremity motor deficit
5	Upper- and lower-extremity motor abnormalities; bladder normal	Combined motor impairment
6	Bladder abnormality irrespective of other domains	Sphincter involvement

Fundamental covariates

Baseline covariates were collected at study entry and included demographic parameters^19,20(age, sex, height, weight, body mass index [BMI]), occupational classification, disease-related factors (duration of symptoms), prior conservative treatment, and comorbid conditions. BMI was calculated as weight (kg)/height (m²) and classified according to Chinese standards as underweight (< 18.5 kg/m²), normal weight (18.5–23.9 kg/m²), and overweight/obese (≥ 24 kg/m²). Age was categorized into two groups: 40–60 years and 61–80 years. Occupations were categorized according to the frequency and duration of neck-flexion exposure: (1) high-frequency static flexion occupations (e.g., accountant, teacher, researcher, office staff), marked by prolonged forward head posture; (2) moderate-frequency dynamic flexion occupations (e.g., worker, farmer, driver), involving intermittent or repetitive neck flexion or sustained axial loading; (3) low-frequency or non-fixed posture occupations; and (4) other/undetermined occupations. Previous conservative treatment was documented as binary variables, including medication, manual therapy, acupuncture, traction therapy, and physical therapy. Comorbidities included hypertension, diabetes mellitus, and urolithiasis. Owing to sparse transitions and the risk of quasi-complete separation, covariate effects were limited to transition pathways meeting the prespecified ≥ 5-event rule; all remaining pathways were modeled using baseline transition intensities only.

MSM models and statistical analysis

A continuous-time multi-state Markov (MSM) model was used to evaluate the course of DCM during the protocol-defined follow-up period²¹. We defined six discrete neurological states based on JOA domains, reflecting different combinations of sensory, motor, and bladder impairment. Follow-up time was measured in years, with observation points at 0, 1/52, 2/52, 4/52, 6/52, 8/52, 3/12, and 6/12 years. Under the Markov assumption, the future transition rate depends only on the current state and not on the pathway taken to reach that state. For clinical interpretation, an instantaneous transition intensity represents the model-estimated short-interval tendency to move from one state to another; a sojourn time represents the expected average duration spent in a state before transitioning elsewhere; and a 1-year transition probability represents the model-based probability of occupying each state after one year, derived from the fitted continuous-time process.

The transition intensity matrix (Q-matrix) was prespecified from the transitions observed in the study cohort and corresponded to a six-state model. The allowed direct transitions were: State 1 to 2; State 2 to 1, 3, 4, and 5; State 3 to 1 and 2; State 4 to 1; State 5 to 2, 3, and 4; and State 6 to 5 (Table 5). All other off-diagonal elements were fixed to zero. Off-diagonal entries represent transition intensities, and each diagonal term is the negative sum of the allowed outgoing intensities so that each row sums to zero.

Table 5.

Transition pathways eligible for covariate adjustment under the prespecified ≥ 5-event rule.

Transition pathway	Covariate adjustment	Reason
1 to 2	Yes	Met ≥ 5 observed-transition threshold
2 to 1	Yes	Met ≥ 5 observed-transition threshold
5 to 2	Yes	Met ≥ 5 observed-transition threshold
5 to 3	Yes	Met ≥ 5 observed-transition threshold
5 to 4	Yes	Met ≥ 5 observed-transition threshold
6 to 5	Yes	Met ≥ 5 observed-transition threshold
All remaining allowed pathways	No	Modeled with baseline intensities only because fewer than 5 observed events were available

No direct transitions from States 1–5 to State 6 were observed during the analyzed non-operative follow-up. These pathways were therefore fixed to zero as a modeling constraint to improve identifiability in this small sample; this should be interpreted as an analytic necessity rather than as proof that such transitions cannot occur in DCM. Remaining transition intensities were estimated by maximum likelihood using the Broyden-Fletcher-Goldfarb-Shanno optimization algorithm. Because participants who converted to surgery were excluded during data screening, surgical conversion was not modeled as a censoring event or as an absorbing state in the MSM framework.

To reduce numerical instability caused by sparse transitions and the modest sample size, covariate effects were estimated only for transition pathways with at least five observed events. In the final model, this criterion was met for transitions 1 to 2, 2 to 1, 5 to 2, 5 to 3, 5 to 4, and 6 to 5; all other pathways were modeled without covariates. Covariates included sex, age group (40–60 versus 61–80 years), occupational classification, prior conservative interventions (medication, manual therapy, acupuncture, traction, and physical therapy), and selected comorbidities. Hazard ratios (HRs) with 95% confidence intervals (CIs) are reported as exploratory estimates and should be interpreted cautiously given the limited event counts.

Transition probabilities were derived from the matrix exponential of Q. The mean sojourn time was calculated as the reciprocal of the absolute value of the relevant diagonal element and represents the expected mean time spent in that state before any transition. We also examined state occupancy probabilities across time to describe the modeled disease course at the cohort level, but because the analysis was conducted on an archived panel dataset with sparse transitions, we emphasize the tabulated transition summaries rather than over-interpreting fine-grained trajectory plots.

All analyses were conducted utilizing the msm package in R (version 4.0). A two-tailed P value of less than 0.05 was deemed statistically significant.

Results

Baseline characteristics of enrolled individuals

Ninety patients with mild to moderate DCM were enrolled (Table 1, Fig. 1). The mean age was 58.6 years, with 52.2% of the cohort aged between 40 and 60 years, and 47.8% aged between 61 and 80 years. 61.1% of the cohort consisted of males. The mean height and weight were 167.1 ± 6.9 cm and 69.6 ± 12.3 kg, respectively. The mean body mass index (BMI) was 24.9 ± 3.7 kg/m². According to Chinese BMI criteria, 45.6% of the population was classified as having a healthy weight, while 54.4% were categorized as overweight or obese.

Fig. 1.

Frequency distribution of DCM states over time. Patients were distributed among six clinical states at baseline, one to eight weeks, and three and six months. During follow-up, motor-dominant states (States 4–5) continued to be common, although sphincter dysfunction (State 6) was continuously uncommon.

Regarding occupational exposure, 37.8% were employed in positions necessitating significant neck bending, 22.2% in roles requiring minimal neck bending, and 40.0% in occupations that did not necessitate any neck bending. 66.7% of patients had previously undergone conservative treatment, with medication (54.4%) being the predominant method, followed by manual therapy (37.8%), physical therapy (37.8%), traction therapy (34.4%), and acupuncture (16.7%). Comorbidities were seen in 6.7% of the cohort.

Transitions among DCM states

Across all scheduled follow-up visits, 71 state transitions were observed, whereas most assessments reflected stability rather than change (Table 3, Supplementary Table S1). At baseline, all patients were symptomatic because the parent trial enrolled only mild-to-moderate DCM. The most common baseline state was combined motor impairment (State 5, 34.4%), followed by isolated lower-extremity motor deficit (State 4, 21.1%) and sensory-only impairment (State 2, 21.1%). States characterized by focal motor deficits (State 4) and mixed motor involvement (State 5) showed notable persistence, with lower outward transition intensity than early sensory states. State 4 had only one modeled outward pathway (to State 1), highlighting the relative stability of isolated lower-extremity motor dysfunction over short intervals. State 5 had several permitted outward pathways, but the predominance of self-persistence indicated that combined motor deficits frequently remained stable during follow-up. State 6 (sphincter involvement) most often transitioned back to State 5, suggesting fluctuation within advanced motor impairment rather than rapid return to mild states. In contrast, sensory-dominant states showed marked bidirectional activity. The transition intensity from State 1 (asymptomatic) to State 2 (sensory-only impairment) was 4.05 (95% CI 2.13–7.69), exceeding the reverse transition intensity of 2.34 (95% CI 1.29–4.25), indicating frequent oscillation between asymptomatic and sensory-only states. No surgery event was represented in the MSM because the model was restricted to scheduled non-operative visits; surgical treatment data were collected separately during long-term follow-up in the parent trial.

Table 3.

Transition intensity matrix (Q) for the multi-state model in patients with degenerative cervical myelopathy (per year, 95% CI)¹.

From state2	To state2	Transition intensity (95% CI)
1 (Asymptomatic)	2 (Sensory-only)	4.05 (2.13–7.69)
2 (Sensory-only)	1 (Asymptomatic)	2.34 (1.29–4.25)
2 (Sensory-only)	3 (Upper-extremity motor)	0.20 (0.04–1.01)
2 (Sensory-only)	4 (Lower-extremity motor)	0.20 (0.05–0.85)
2 (Sensory-only)	5 (Combined motor)	0.50 (0.19–1.33)
3 (Upper-ext motor)	1 (Asymptomatic)	0.73 (0.27–2.02)
3 (Upper-ext motor)	2 (Sensory-only)	0.60 (0.19–1.89)
4 (Lower-ext motor)	1 (Asymptomatic)	0.36 (0.14–0.97)
5 (Combined motor)	2 (Sensory-only)	0.97 (0.49–1.92)
5 (Combined motor)	3 (Upper-extremity motor)	1.00 (0.52–1.92)
5 (Combined motor)	4 (Lower-extremity motor)	0.63 (0.28–1.42)
6 (Sphincter dysf.)	5 (Combined motor)	2.76 (1.24–6.14)

¹Each off-diagonal entry indicates the predicted instantaneous transition rate from the “From” state to the “To” state, accompanied by a 95% confidence interval in parentheses. The diagonal items are omitted as they represent the negative sum of each row. No transitions to State 6 occurred throughout the follow-up, hence the intensities were assigned a value of 0.

²Definition of States: State 1: Asymptomatic; State 2: Sensory-only impairment; State 3: Upper-extremity motor deficit; State 4: Lower-extremity motor deficit; State 5: Combined motor impairment; State 6: Sphincter dysfunction.

One-year transition probability matrix

One-year transition probabilities were calculated from the continuous-time MSM using the matrix exponential of the fitted Q-matrix (Fig. 2). For clinical readers, these probabilities summarize where a patient starting in a given state is expected to be one year later if the fitted transition process continues to apply. Because the analyzed visit panel extended through 6 months, the 1-year probabilities should be interpreted as model-based projections from the fitted continuous-time MSM rather than as directly observed 12-month proportions. Most patients remained in their baseline states at one year, particularly those with motor-dominant impairment. Progression to more severe states was generally uncommon (< 11%). Patients starting in State 1 had a 33.6% probability of remaining asymptomatic, a 31.2% probability of moving to sensory-only impairment, and a 25.1% probability of moving to isolated upper-extremity motor deficit. Patients in State 2 showed bidirectional change, with 45.4% reverting to State 1 and 42.8% remaining sensory-only. Once motor-dominant states were established, persistence increased markedly: State 4 showed the highest 1-year persistence (71.9%), whereas State 5 remained unchanged in 54.0% of patients and improved to less severe states in 37.0%. Patients in State 6 most often moved to State 5 (64.0%), with rare direct movement to milder states.

Fig. 2.

Model-based 1-year transition probability matrix among six DCM states. Rows indicate the starting state and columns indicate the projected state after 1 year under the fitted continuous-time MSM. Darker shading indicates higher model-estimated probability. Because the analyzed visit panel extended through 6 months, these values should be interpreted as model-based projections rather than directly observed 12-month proportions.

Overall, the 1-year probability matrix supports a non-monotonic illness course with dynamic variation in early sensory states and more structural stability in established motor impairments (Fig. 3).

Fig. 3.

LCA classes based on the Conditional Probability Plot of enrolled individuals. Three latent classes were identified from longitudinal state patterns observed between baseline and 6 months (36.8%, 28.9%, and 34.3%). The vertical bars show the conditional probability of each observed state indicator within each class. For interpretive purposes, Class 1 corresponded to a milder/fluctuating pattern, Class 2 to an intermediate motor-dominant pattern, and Class 3 to a more severe/persistent impairment pattern.

Influence of covariates on state transitions

When baseline factors were incorporated, most covariate effects were highly imprecise because only a small number of transition events were available for each modeled pathway. Prior medication history and BMI showed the largest point estimates, but these results should be interpreted as exploratory rather than confirmatory. For example, prior pharmacologic treatment was associated with a higher estimated hazard of transition from State 5 to State 2 (HR approximately 22.7, 95% CI 1.7–300), but the confidence interval was extremely wide. Likewise, overweight/obesity was associated with a numerically higher estimated hazard for some transitions, such as State 5 to 2 (HR approximately 8.66, 95% CI 0.62–120), yet these estimates were not statistically precise. Age, sex, and comorbidities showed no clear association with transition hazards after adjustment. Overall, the covariate analyses should be viewed as hypothesis-generating.

Variations in transition patterns among subgroups

Stratified summaries of transition probabilities across demographic and clinical subgroups are shown in Table 4. These analyses were exploratory and were not powered for definitive subgroup inference. Younger patients appeared somewhat more likely than older patients to improve from combined motor impairment to less severe states within 6 months, while short-term worsening remained uncommon in both age groups. Sex-related differences were minimal. Patients with BMI ≥ 24 showed more state changes overall than normal-weight patients, but the direction and clinical meaning of this pattern remain uncertain because the corresponding HRs were imprecise and based on few events. Likewise, participants with prior conservative treatment appeared to have slightly fewer transitions, which may reflect baseline selection rather than a true stabilizing effect. Baseline severity remained the clearest descriptive factor: patients beginning in moderate motor-dominant states often showed only partial improvement, and none regained a sustained asymptomatic status during short-term follow-up. These subgroup observations should therefore be interpreted as descriptive patterns rather than robust prognostic findings (Table 5).

Table 4.

Covariate effects on selected state transitions (Hazard Ratio, 95% CI) ¹.

Transition from	Transition to	Sex (Male vs. Female)	Age (61–80 vs. 40–60)	BMI (≥ 24 vs. < 24)	Prior Cons. treatment
1	2	0.35 (0.04–3.2)	0.48 (0.05–4.5)	2.02 (0.18–22.5)	1.11 (0.12–9.9)
2	1	1.08 (0.30–3.85)	0.88 (0.24–3.18)	1.26 (0.34–4.68)	1.32 (0.38–4.57)
5	2	6.32 (0.65–61.2)	4.09 (0.42–40.1)	8.66 (0.62–120)	1.48 (0.31–7.10)
5	3	1.54 (0.36–6.65)	1.16 (0.28–4.86)	0.54 (0.10–2.87)	0.96 (0.22–4.12)
5	4	1.01 (0.24–4.32)	1.32 (0.30–5.76)	1.09 (0.22–5.35)	0.89 (0.20–3.98)
6	5	HR not estimable 2	1.04 (0.10–10.9)	0.78 (0.07–8.57)	0.74 (0.07–7.94)

¹The displayed transitions had five or more occurrences and were adjusted for variables, including baseline sex, age group, BMI category, and preceding conservative therapy. Hazard ratios exceeding 1 signify an elevated transition hazard in the first group (e.g., male) compared to the reference group. Estimates for occupational exposure and specific treatment modalities are excluded for conciseness; they were not significant in adjusted models.

²Minimal or no events transpired in the comparison group for this covariate, resulting in an unreliable estimate (the hazard ratio could not be accurately determined).

Mean sojourn time and total stay of each DCM state

We calculated both total time spent in each state and the mean sojourn time to assess state stability and fluctuation in DCM (Table 6, B, and 5). The mean sojourn time reflects the average time a patient is expected to remain in a state before moving elsewhere. State 4 (isolated lower-extremity motor impairment) had the longest mean sojourn time at 2.76 years (95% CI 1.04–7.35), indicating pronounced persistence. By contrast, State 1 (asymptomatic) had a mean sojourn time of 0.25 years (95% CI 0.13–0.47), and State 2 (sensory-only impairment) had a mean sojourn time of 0.31 years (95% CI 0.19–0.50), underscoring the lability of early sensory-dominant patterns. State 5 (combined upper- and lower-extremity motor deficits with normal bladder function) showed an intermediate mean sojourn time of 0.39 years (95% CI 0.26–0.58). For State 6 (sphincter involvement), the estimated mean sojourn time was 0.36 years (95% CI 0.16–0.81); because this state was relatively uncommon, its precision was limited and the estimate should be interpreted cautiously (Tables 7, 8).

Table 6.

Mean sojourn time by JOA domain-based symptom state.

State	Definition (JOA domain-based)	qii (from Q)	Mean sojourn time, years	95% CI, years	Mean sojourn time (weeks)
1	Asymptomatic (all domains normal)	− 4.045	0.247	0.130–0.470	12.900
2	Sensory-only impairment	− 3.242	0.308	0.192–0.496	16.000
3	Isolated upper-extremity motor deficit	− 1.334	0.750	0.375–1.499	39.000
4	Isolated lower-extremity motor deficit	− 0.362	2.760	1.036–7.353	143.500
5	Combined upper + lower motor deficits (bladder normal)	− 2.596	0.385	0.258–0.576	20.000
6	Sphincter involvement (abnormal bladder function)	− 2.757	0.363	0.163–0.807	18.900

The longer the average sojourn time, the more stable the state; the shorter it is, the more likely state transitions are to occur (indicating more pronounced fluctuation).

Table 7.

Clinically oriented summary of state persistence and dominant 1-year transition patterns.

State	Mean sojourn time (years)	Dominant 1-year pattern
1	0.25	Dynamic; 33.6% remained asymptomatic and 31.2% moved to sensory-only impairment
2	0.31	Labile; 42.8% remained sensory-only and 45.4% improved to asymptomatic status
3	0.75	Intermediate persistence with bidirectional movement mainly involving States 1 and 2
4	2.76	Most persistent motor-dominant state; 71.9% remained in State 4 at 1 year
5	0.39	Moderately persistent; 54.0% remained in State 5 and 37.0% improved to less severe states
6	0.36	Usually returned to advanced motor impairment; 64.0% moved to State 5 and direct transitions to mild states were uncommon

Table 8.

Total stay time by JOA domain-based symptom state.

State	Definition (JOA domain-based)	N (patients)	Total stay (years)
1	Asymptomatic (all domains normal)	19	4.69
2	Sensory-only impairment	19	5.87
3	Isolated upper-extremity motor deficit	11	8.25
4	Isolated lower-extremity motor deficit	19	53.44
5	Combined upper + lower motor deficits (bladder normal)	31	11.94
6	Sphincter involvement (abnormal bladder function)	10	3.63

State-based latent classes analysis of DCM

Latent class analysis (LCA) was performed using eight repeated categorical state indicators derived from the scheduled visits (S0, S1, S2, S4, S6, S8, S3mon, and S6mon), representing the observed state sequence across early and intermediate follow-up (Table 9, Supplementary Table S2, Fig. 1). Models with one to six classes were examined. AIC and BIC decreased and then increased as the number of classes rose, and the three-class model showed the lowest BIC (1997.15), favorable adjusted BIC, and excellent entropy (0.99), supporting selection of the three-class solution over nearby alternatives. The BLRT was also consistent with retention of three classes, whereas the LMRT did not provide additional discriminatory support and was therefore interpreted cautiously. Clinically, the three classes may be interpreted as an exploratory milder/fluctuating class (Class 1; 37%), an intermediate motor-dominant class (Class 2; 29%), and a more severe/persistent impairment class (Class 3; 34%). These labels are intended only to aid clinical interpretation; the classes remain exploratory, data-driven patterns that should not be regarded as fixed biological categories and require external validation before being used for individual prognostic decisions.

Table 9.

Goodness-of-fit of the LCA classes of enrolled individuals.

Classes	LogLik	AIC	BIC	aBIC	Entropy	Class_Probs	p_LMRT	p_BLRT
1	− 1154.97	2383.93	2476.42	2359.65
2	− 853.14	1890.27	2120.25	1829.89	1.00	0.67/0.33	0.00	0.00
3	− 667.84	1629.68	1997.15	1533.21	0.99	0.37/0.29/0.34	1.00	0.00
4	− 594.97	1593.95	2098.91	1461.38	1.00	0.41/0.07/0.27/0.26	1.00	0.00
5	− 1042.92	2599.85	3242.30	2431.19	1.00	0.41/0.26/0.26/0.01/0.07	0.00	1.00
6	− 1006.08	2636.15	3416.09	2431.40	1.00	0.23/0.32/0.09/0.04/0.17/0.14	1.00	0.00

Discussion

This study is, to our knowledge, the first in DCM to combine a symptom-domain-defined multi-state Markov framework with latent class analysis in a secondary analysis of a multicenter randomized trial dataset. By focusing specifically on scheduled non-operative follow-up rather than comparative treatment effects, the analysis provides a more granular view of short-term symptom dynamics than is possible from simple pre/post score comparisons alone. Several findings are clinically notable. Sensory-dominant or early mild symptom states appeared relatively labile, consistent with prior literature suggesting that the natural history of mild DCM/CSM under non-operative management may be variable or intermittently progressive^22–24. In contrast, once objective motor involvement was present, especially lower-extremity motor deficit, patients were much more likely to remain in a motor-impaired state over time^25–28. These results suggest that under non-operative management, motor deficits may reflect a more persistent neurological phenotype than early sensory complaints.

These state transition patterns further suggest that ‘mild’ DCM is not a homogeneous entity but instead includes subgroups with distinct clinical trajectories^29,30. The LCA complemented the MSM findings by identifying three data-driven longitudinal patterns that broadly separated a severe motor-dominant/persistent impairment phenotype from intermediate and milder/fluctuating phenotypes. Importantly, we do not interpret these latent classes as fixed biological categories or as ready-to-use prognostic tools. Rather, they provide an exploratory summary of heterogeneity within this cohort and require validation in independent datasets. Our findings are consistent with recent unsupervised phenotyping studies in DCM that have similarly identified motor-dominant and other clinically distinct symptom profiles with different levels of baseline impairment and recovery potential^31,32. Taken together, the MSM and LCA results support closer monitoring of patients with motor-dominant patterns under conservative care, but they do not directly demonstrate that early surgery would alter these trajectories in this cohort.

Our data also suggested that transition dynamics may vary across patient subgroups, but these findings should be interpreted with great caution. Apparent differences by age, BMI, and prior treatment history were descriptive, based on sparse events, and vulnerable to overfitting and multiple-comparison concerns. For example, younger patients appeared somewhat more likely to improve from combined motor impairment than older patients, which is directionally consistent with prior evidence that older age may be associated with lower recovery potential^33,34. However, the present subgroup analyses were underpowered and should be regarded as hypothesis-generating. The same caution applies to the BMI-related findings: although overweight/obesity was associated with numerically higher transition rates in some pathways, the confidence intervals were very wide and do not support a firm conclusion that higher BMI predicts either improvement or worsening in non-operatively managed mild-to-moderate DCM^35,36.

Likewise, the suggestion that prior conservative treatment was associated with a more stable trajectory may simply reflect selection effects, because patients who had already received treatment before enrollment may have represented a subgroup with more indolent or better-controlled disease. Baseline severity remained the most clinically intuitive descriptive factor: patients beginning in moderate motor-dominant states often showed only partial improvement and rarely returned to asymptomatic status during follow-up^13,37,38. These observations align with the broader clinical impression that symptom-domain patterns carry prognostic information^25,31. Nevertheless, because this analysis was conducted in a modest non-operative cohort and did not randomize early versus delayed surgery, the findings should be used to motivate phenotype-aware surveillance and future comparative studies rather than to prescribe prompt surgical intervention.

Strengths and limitations

This study has several strengths, including the use of a well-characterized multicenter randomized trial dataset, explicit symptom-domain state definitions based on JOA subscores, and a continuous-time MSM framework that allowed estimation of transition intensities, 1-year transition probabilities, and mean sojourn times. The complementary LCA also provided an intuitive, data-driven summary of longitudinal heterogeneity. At the same time, several limitations warrant explicit emphasis. First, the sample size was small (90 patients), with only 71 observed state shifts, limiting the stability of a multi-parameter six-state MSM and increasing the risk of overfitting or imprecise covariate estimates. Second, JOA domain ratings are visit-based clinical assessments and may be affected by measurement variability or state misclassification around category thresholds. Third, the Markov assumption implies that future transition risk depends only on the current state and may not fully capture the influence of prior symptom history. Fourth, some model constraints, particularly structural zeros for unobserved transitions to State 6, were imposed for identifiability in a sparse dataset and should not be interpreted as biological impossibilities. Fifth, formal sensitivity analyses using alternative state aggregations or exclusion of rare states were considered but were not pursued because they would have further reduced the number of transition events per pathway and likely produced even less stable estimates. Finally, because this was a secondary analysis of a trial originally designed for another primary question, residual selection bias and unmeasured confounding cannot be excluded. In addition, because patients who underwent surgery during treatment were excluded during data screening, the present findings specifically reflect trajectories among patients maintained on non-operative management and should not be interpreted as capturing outcomes after treatment escalation.

Conclusions

In conclusion, this exploratory secondary analysis suggests that symptom-domain trajectories in mild-to-moderate DCM under non-operative care are heterogeneous and phenotype-dependent. Sensory-predominant symptoms are relatively labile, whereas motor-dominant states, particularly isolated lower-extremity motor deficits, appear more persistent with limited spontaneous resolution. These findings may improve phenotype-specific counseling and help clinicians identify patients who warrant closer follow-up and timely surgical evaluation when appropriate. However, the results should not be interpreted as direct evidence for surgical timing thresholds or as a substitute for prospective comparative studies.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Bo Xu: conceived the study, re-examined the data, and wrote the manuscript. Guo-Liang Ma: collected the data, wrote the manuscript, and reviewed and revised the manuscript. Zhi-Zhuang Wang: collected the data. Xin Chen: analyzed and re-examined the data. Bo-Wen Yang: analyzed and re-examined the data. Xiao-Kuan Qin: analyzed and re-examined the data. Dong-Sheng Yuan: collected the data. Peng-Hui Li: collected the data. Ren-Kun Zhao: re-examined the data. Jia-Cheng Zheng: re-examined the data. Min-Shan Feng: analyzed the data. Jie Yu: analyzed the data. Li-Guo Zhu and He Yin: reviewed and revised the manuscript. All authors contributed to the article and approved the submitted version.

Funding

This study was supported by the China Academy of Chinese Medical Sciences Basic Scientific Research Business Expenses Outstanding Young Scientific and Technological Personnel Training Project (grant no. ZZ14-YQ-022). This study was also supported by the National Key R&D Program of China, Systematic Study on TCM Diagnosis and Treatment of Bone and Joint Degenerative Diseases and Adolescent Scoliosis: Research on TCM Diagnosis and Treatment of Cervical Spondylotic Myelopathy and New Drug Development (grant no.2024YFC3507401), and the Science and Technology Innovation Project of the China Academy of Chinese Medical Sciences: Risk Stratification and Early Warning Model for Disease Progression in Mild-to-Moderate Degenerative Cervical Myelopathy Based on Multimodal Longitudinal Trajectories.

Data availability

Data will be made available on request.

Declarations

Competing interests

The authors declare no competing interests.

Ethical approval

The parent trial was conducted in accordance with the Declaration of Helsinki and was approved by the ethics committee of Wangjing Hospital, China Academy of Chinese Medical Sciences (WJEC-KT-2016-004-P001). The trial was registered at the Chinese Clinical Trials Registry (ChiCTR-INR-16009723), and written informed consent was obtained from all participants in the parent trial before enrollment. The present study was a secondary post hoc analysis of de-identified trial data restricted to the non-operative follow-up period. No additional informed consent was required for this secondary analysis of de-identified trial data, in accordance with institutional requirements.