Latent profile analysis of multidimensional muscle profiles versus traditional sarcopenia metrics for capturing sagittal malalignment in adult spinal deformity: a retrospective study

Ethics statement

This study adhered to the Declaration of Helsinki and was approved by the Hospital for Special Surgery institutional review board (approval number: 2018-1599). The requirement for informed consent was waived owing to the retrospective and observational nature of the study.

Study design

The study was designed as a retrospective cohort study.

Patient sample

This cohort included 145 patients with ASD who underwent fusion at ≥5 levels (upper instrumented vertebra [UIV]: T1–L1) with ≥1-year of follow-up. The exclusion criteria were as follows: (1) deformities secondary to other diseases, (2) missing data, and (3) a history of spinal surgery.

Data collection

We extracted patient demographics from our database and recorded the intraoperative and postoperative details, as described below.

Muscle assessment

Lumbar magnetic resonance imaging (MRI; supine T2-weighted axial images) was used to assess the psoas and paralumbar (multifidus and erector spinae) muscles at L4, as the L4 paralumbar level has been reported to be a relevant and reliable surrogate for evaluating muscle status in degenerative ASD with sagittal imbalance [16,17]. Muscle quality was graded according to the Goutallier classification. The bilateral psoas FI was averaged from the left and right sides, whereas the paralumbar FI was calculated as the mean FI of the bilateral multifidus and erector spinae muscles [18,19].

Muscle quantity was assessed using the psoas vertebral body index (PVBI) and paralumbar vertebral body index (PLVBI), calculated from axial T2-weighted MRI at L4 as the ratio of the bilateral muscle CSA to the vertebral body CSA. Most measurements were performed using the Sectra Workstation software by a single, experienced observer following a standardized protocol, as previously reported by our team [20,21].

Spinal alignment

Standing full-length spine radiographs were used to measure alignment at three time points: preoperative, immediate postoperative, and last follow-up (≥1 year). The following parameters were obtained from standing full-spine radiographs: coronal C7 plumb line, coronal maximum Cobb angle, thoracic kyphosis (T4–T12), lumbar lordosis (LL, L1–S1), sacral slope (SS), pelvic incidence (PI), PI–LL mismatch, pelvic tilt (PT), sagittal vertical axis (SVA), L1 pelvic angle (L1PA), L1PA offset (measured L1PA–the normal value, where normal=0.5×PI−21° [22]), T4 pelvic angle (T4PA), and T4–L1PA mismatch (T4PA–L1PA). Changes in alignment (▲) from the immediate postoperative period to the final follow-up were calculated.

Fixation strategy

The UIV was classified as upper thoracic (T1–T6) or lower thoracic (T7–L1) [23,24], and the lowest instrumented vertebra (LIV) was categorized based on whether pelvic fixation was performed. The UIV anchoring technique (hooks and/or screws) and length of fusion were also collected.

Outcomes

Functional outcomes were assessed at baseline and at the final postoperative follow-up (≥1 year) using the Oswestry Disability Index (ODI), and the ODI change from baseline to final follow-up was calculated. Similarly, back/leg pain (Visual Analog Scale [VAS] scores) was recorded but not included in the analysis because of incomplete data.

The intraoperative data included estimated blood loss, operative duration, intraoperative complications (as per the Adult Symptomatic Lumbar Scoliosis–International Spine Study Group classification [25]), and length of hospital stay. Postoperative outcomes included proximal junctional kyphosis (PJK), proximal junctional failure (PJF), reoperations, and any surgical or medical complications.

Statistical analysis

Continuous variables are presented as mean±standard deviation or median with interquartile range, depending on the data distribution. FI and VAS scores were treated as approximate continuous variables in the statistical analysis. Pearson’s or Spearman’s correlation coefficients (r) were calculated to assess the associations among muscle parameters before LPA.

Clustering of muscle profiles

LPA was performed to identify distinct muscle phenotypes. The variables included in the model were PVBI, PLVBI, and FI of the psoas and paralumbar muscles. Analyses were performed using Mplus ver. 8.3 (Muthén & Muthén, Los Angeles, CA, USA). Model fit was evaluated using standard information criteria and likelihood ratio tests, and classification certainty was assessed using entropy. A radar chart was created to illustrate each muscle profile.

Comparison of LPA-based classification and single-parameter sarcopenia classifications in predicting sagittal alignment

We evaluated the predictive performance of the LPA classification for key sagittal alignment parameters (SVA, L1PA offset, T4–L1PA mismatch, PI–LL, and PT) in comparison with the conventional binary definitions of sarcopenia derived from single muscle metrics using linear regression models.

Before this comparison, conventional binary definitions (non-sarcopenic vs. sarcopenic) were applied separately to each muscle parameter. Two classification strategies were used: (1) the lowest tertile, as reported in prior studies [8], and (2) an optimal cutoff value. To determine the optimal cutoff values, we followed a previously described approach [26], in which binary sarcopenia classification was used to predict the five sagittal parameters based on linear regression models. The average R² across the five models was calculated for each candidate cutoff value, and the cutoff yielding the highest mean R² was defined as the optimal threshold. This analysis was conducted in R ver. 4.4.1 (The R Foundation for Statistical Computing, Vienna, Austria) using the readxl package, and the relationship between the cutoff values and R² performance was visualized using line plots for each muscle metric.

Subsequently, we used univariable linear regression to compare the predictive performance (R² and p-values) of LPA-based classification with that of binary sarcopenia definitions based on both the lowest tertile and optimal cutoffs (defined as ≤cutoff for sarcopenia; >cutoff for non-sarcopenia). A higher R² with a p-value <0.05 was interpreted as better predictive accuracy.

Comparison across LPA-derived muscle subtypes

Continuous variables were compared using analysis of variance or Kruskal-Wallis tests, and categorical variables were compared using chi-square or Fisher’s exact tests with appropriate post hoc and multiple-comparison adjustments. For alignment parameters showing significant group differences at the last follow-up, additional multivariate linear regression analyses were performed, adjusting for age, fusion length, UIV category (upper thoracic [T1–T6] vs. lower thoracic [T7–L1]), and LIV category (with vs. without pelvic fixation). The muscle phenotype was entered as a categorical variable, with the Hete-Deg group serving as the reference category. Regression coefficients, standard errors (SEs), and 95% confidence intervals (CIs) were also calculated. Statistical significance was set at p<0.05. Analyses were performed in R ver. 4.4.1.

Baseline characteristics

The baseline characteristics of the 145 patients with ASD (Fig. 1) are summarized in Table 1. No strong collinearity was observed among the four muscle parameters (|r| <0.7), justifying their inclusion in the LPA model (Supplement 1).

Fig. 1

Flowchart of patient selection and analytical framework. UIV, upper instrumented vertebra; LPA, latent profile analysis; PVBI, psoas vertebral body index; PLVBI, paralumbar vertebral body index; FI, fat infiltration; Mild-Deg, mild degeneration group; Mod-Deg, moderate degeneration group; Seve-Deg, severe degeneration group; Hete-Deg, heterogeneous degeneration group.

Table 1

Clinical characteristics for the 145 patients with adult spinal deformity

LPA-based cluster analysis

LPA model fit indices (the lowest AIC, BIC, and aBIC values, significant LMR and BLRT tests, and high class-specific average posterior probabilities [AvePPs]) consistently supported four muscle phenotypes (Table 2): Mild-Deg (high muscle quantity, low FI; n=49; AvePP=0.935), Mod-Deg (moderate muscle quantity and quality; n=39; AvePP=1.000), Seve-Deg (low muscle quantity, high FI; n=34; AvePP=1.000), and Hete-Deg (low muscle quantity, high paralumbar FI but low psoas FI; n=23; AvePP=0.871) (Fig. 2).

Table 2

Comparison of fit statistics for LPA models with varying class numbers

Fig. 2

Visualization of intergroup differences across latent profile analysis-defined muscle profiles using standardized means of four muscle metrics: psoas vertebral body index (PVBI), paralumbar vertebral body index (PLVBI), psoas fat infiltration (FI), and paralumbar FI. Mild-Deg, mild degeneration group; Mod-Deg, moderate degeneration group; Seve-Deg, severe degeneration group; Hete-Deg, heterogeneous degeneration group.

LPA vs single-parameter prediction

We determined cutoff values for traditional sarcopenia definitions. The lowest tertile cutoffs for the cohort were approximately PVBI ≤1.44, PLVBI ≤2.68, psoas FI ≤0.00, and paralumbar FI ≤2.00 (Table 3). Using the optimal R2 method, the best cutoffs were PVBI ≤2.05, PLVBI ≤2.61, psoas FI ≤0.50, and paralumbar FI ≤3.25 (Fig. 3, Table 3). Patients were classified as sarcopenic (≤cutoff) or non-sarcopenic (>cutoff), and the group sizes for both cutoff methods are summarized in Table 3.

Table 3

Optimal R²-based and conventional tertile cutoffs for muscle metrics with patient stratification

Fig. 3

Change in mean R² with different cutoff values of the psoas vertebral body index (PVBI) (A), the paralumbar vertebral body index (PLVBI) (B), psoas fat infiltration (FI) (C), and paralumbar FI (D): identifying optimal cutoffs for single muscle metric. At each cutoff of muscle metric, patients were dichotomized, and the resulting binary variable was used as the predictor in univariate linear regressions, with sagittal vertical axis, L1 pelvic angle (L1PA) offset, T4–L1PA mismatch, pelvic incidence–lumbar lordosis, and pelvic tilt used sequentially as the dependent variables. For each cutoff, the mean R² across the five models was calculated and plotted. The curve illustrates how the predictive power (mean R²) changes with different cutoffs, with the peak identifying the optimal cutoff for predicting sagittal malalignment.

Using these binary sarcopenia definitions, we performed linear regression for each key sagittal metric. The LPA-based classification achieved the highest R2 values for T4–L1PA mismatch (0.155), PI–LL (0.160), and PT (0.139), and the second highest for SVA (0.116), outperforming all single-metric definitions (Table 4).

Table 4

Linear regression models of muscle metrics predicting key sagittal spinal alignment parameters

Clinical and radiographic comparisons by LPA group

Table 5 summarizes the results of group comparisons. PVBI, PLVBI, and FI values differed significantly among the groups (p<0.001), further confirming the validity of the classification. Patient age increased significantly from Mild- through Seve-Deg (median, 51, 66, and 68 years; p<0.001), and the prevalence of hypertension, coronary artery disease, and higher Charlson comorbidity index (CCI) also increased with more severe muscle degeneration.

Table 5

Comparative analysis of clinical characteristics and surgical outcomes across LPA-derived muscle profiles

Preoperative sagittal alignment deteriorated across Mild to Seve-Deg; PI–LL, PT, SVA, L1PA offset, and T4–L1PA mismatch all worsened progressively (p<0.001). Notably, the Hete-Deg group had one of the worst alignments (mean PI–LL, 24.0°; median SVA, 88.1 mm; and median T4–L1PA mismatch, 12.0°), with the second-largest L1PA offset (4.8°).

In the fixation/intraoperative data, the Mild-Deg group had fewer patients with pelvic fixation (p=0.013). The Hete-Deg group tended to have longer fusions (median, 11; p=0.019) and operative times (median, 314 minutes; p=0.011).

Postoperative alignment was largely comparable between the groups. Only PI–LL (p=0.005) and T4–L1PA mismatch (p=0.042) were slightly higher in the Mod-Deg/Seve-Deg groups than in the Mild-Deg/Hete-Deg groups. At the last follow-up, patients in the Mod-Deg and Seve-Deg groups had significantly greater malalignment; PI–LL (p=0.002), T4–L1PA mismatch (p=0.005), T4PA (p=0.017), SS (p=0.006), and LL (p<0.001) were all larger than in the Mild-Deg/Hete-Deg group. In additional multivariable linear regression adjusting for age, fixation strategy (UIV/LIV), and fusion length, the association of muscle phenotype with PI–LL (overall p=0.027) and LL (overall p=0.003) remained significant. Using the Hete-Deg group as the reference category, the Mod-Deg (β=8.90, SE=3.96; 95% CI, 1.06–16.75; p=0.026) and Seve-Deg (β=9.04, SE=4.02; 95% CI, 1.08–17.00; p=0.026) groups demonstrated significantly greater PI–LL values. Similarly, the Mod-Deg (β=−10.32, SE=3.83; 95% CI, −17.90 to −2.74; p=0.008) and Seve-Deg (β= −9.71, SE=3.88; 95% CI, −17.39 to −2.02; p=0.014) groups exhibited significantly lower LL values than the Hete-Deg group. In contrast, the associations for T4–L1PA mismatch (overall p=0.399), T4PA (overall p=0.423), and SS (overall p=0.497) were no longer significant, suggesting that some radiographic differences may be partially confounded by surgical factors. There was a trend toward larger changes in PI–LL and T4PA (postoperative → follow-up) in the Seve-Deg group (p=0.065, and 0.086, respectively).

Finally, there were no significant differences in functional recovery (ODI scores) or complication rates (PJK, PJF, etc.) among the four groups.

Summary of findings

Paraspinal sarcopenia is closely associated with sagittal malalignment in ASD; however, traditional definitions based on a single muscle or metric fail to capture the complex and heterogeneous nature of trunk muscle degeneration in patients with ASD. By applying multidimensional muscle profiling with LPA, we identified four distinct muscle phenotypes with varying combinations of quantity and quality in the psoas and paralumbar muscles. Compared with conventional single-metric sarcopenia definitions, LPA-based classification more accurately predicted key sagittal alignment parameters, highlighting its superiority in characterizing sagittal alignment. Importantly, the LPA method effectively identified the distinct Hete-Deg subtype, which is characterized by functional compensation failure rather than structural degeneration. Despite marked preoperative sagittal deformity, patients with Hete-Deg were able to maintain long-term alignment. In contrast, the Seve-Deg subtypes exhibited structural muscle damage associated with both severe preoperative deformity and progressive postoperative imbalance. These findings highlight the value of LPA-based classification not only in phenotypic resolution but also in capturing clinical differences in postoperative alignment stability and progression.

Multidimensional muscle profiling enhances malalignment prediction

Compared to traditional sarcopenia definitions based on single muscle measures, our LPA-based classification better predicted sagittal imbalance by capturing muscle health multidimensionally. This is particularly important given the notable heterogeneity in degeneration between the paralumbar and psoas muscles [14,15] and between muscle quality and quantity within the same muscle group [12,13]. Consistently, our data showed only weak-to-moderate correlations among the four muscle parameters, indicating the need for multidimensional profiling. Conventional criteria often focus on a single metric at a single site, resulting in inconsistent definitions and limited sensitivity to localized changes. However, our LPA model integrates both quantitative (CSA) and qualitative (FI) features across multiple muscle groups, offering a more comprehensive assessment aligned with recent expert recommendations favoring machine learning–based imaging analysis [27]. Such an approach may better reflect the complex, multisite nature of sarcopenia, provide a person-centered framework to identify clinically meaningful phenotypes beyond isolated muscle parameters, and offer a novel framework for future investigations of muscle degeneration and its clinical relevance. Although the sample size was modest and one latent profile was relatively small, methodological studies suggest that reliable latent profile solutions may still be obtained when supported by adequate indicator quality and class separation [28,29]. Consistent with this, the high entropy and favorable fit indices in our model support the stability of the identified profiles, although validation in larger cohorts is warranted (Table 2).

Hete-Deg: a distinct phenotype reflecting functional compensation

A particularly novel finding was the Hete-Deg phenotype, characterized by a low overall muscle quantity and markedly high paralumbar FI but near-normal psoas FI. Previous studies have noted that psoas FI can remain low even in patients with ASD [30], and is sometimes comparable to levels observed in healthy individuals [31]. However, the absence of concurrent assessment of paralumbar muscle status in these studies precluded the identification of this distinct “Hete-Deg” pattern in their results. Our Hete-Deg group had some of the most severe preoperative deformities; interestingly, they maintained sagittal alignment after >1 year of follow-up after surgery. We speculate that this pattern reflects functional compensation rather than irreversible atrophy, as their preserved psoas function may have compensated for weak paraspinal muscles to sustain alignment after correction [14].

In contrast, the Mod-Deg and Seve-Deg subtypes (moderate-to-low muscle quantity with high FI) had progressively worse baseline malalignment and poorer maintenance of correction over time. Given that FI in the trunk muscles is largely irreversible [32], these groups may represent irreversible structural muscle damage and substantial loss of functional capacity, which may contribute to the progressive deterioration of postoperative alignment over time.

Paraspinal sarcopenia corresponds with greater comorbidity burden

We also observed that more severe muscle degeneration paralleled poor systemic health. Hypertension, coronary artery disease, and higher CCI were more common in the Mod-Deg and Seve-Deg groups, mirroring the pattern in systemic sarcopenia, in which frailty and comorbidities increase with muscle loss. Population data show that individuals with sarcopenia often have more cardiovascular/metabolic diseases, and sarcopenia can accelerate cardiovascular disease progression [33]. In addition, a higher CCI is negatively correlated with muscle mass in older adults [34].

Radiographic-functional discordance

Although LPA-derived phenotypes differed in long-term alignment behavior, no significant differences in ODI improvement or complication rates were observed. This may be related to the relatively short follow-up period, which may have been insufficient for radiographic differences to translate into measurable functional divergence. Moreover, the modest sample size and low event rates may have limited the statistical power to detect between-group differences, and ODI may have been insufficiently sensitive to capture subtle phenotype-related differences in postoperative function. Prior studies have also suggested that functional outcomes may not fully parallel radiographic differences and that ODI may not capture all dimensions of ASD-related disability [35].

Limitations

First, although we found significant associations between LPA-based phenotypes and sagittal alignment, the retrospective design of this study precludes causal interpretation. Several studies have demonstrated that greater spinal malalignment is associated with poor muscle quality and reduced muscle quantity [30]. Most likely, there is a reciprocal relationship [36]; as muscles weaken or atrophy, mild deformities can progress unchecked, highlighting the importance of an accurate muscle profile in patients with ASD. In addition, the single-center design and highly selected surgical cohort may limit the generalizability of these findings to broader ASD populations, particularly patients with less severe deformities or those treated at other institutions. Moreover, the relatively short follow-up period may not have fully captured longer-term progression or functional divergence across phenotypes. Second, the optimal cutoffs were derived and evaluated within the same dataset, raising the possibility of overfitting. External validation is required to confirm the robustness of the identified phenotype. Third, muscle quality was assessed using the semiquantitative Goutallier grading system, which has not been validated for evaluating lumbar muscles. Similar to prior studies treating ordinal variables (e.g., pain scores) as continuous for LPA [37], we approximated FI grades as continuous, although this remains an approximation. Future studies should explore advanced modalities, such as quantitative MRI [38], to better characterize muscle quality and further strengthen LPA-based profiling. Moreover, muscle parameters were assessed using supine MRI, whereas alignment was measured using standing radiographs, and this positional discrepancy may have influenced the observed associations.