Spine Age Gap (SAG): Biomarker for Spine Health

Updated 3 July 2026

Spine Age Gap (SAG) is defined as the bias-corrected difference between MRI-derived spine age and chronological age, where a positive value indicates accelerated spine aging.
SAG is computed using a 3D CNN with precise spine segmentation and statistical bias correction, ensuring uniform error distribution across age groups.
Empirical evaluations reveal that SAG correlates with degenerative spine conditions and lifestyle factors, with MAE around 3.67 years and R² up to 0.85 demonstrating its clinical relevance.

Searching arXiv for recent and relevant papers on Spine Age Gap and related spine age estimation.

Spine Age Gap (SAG) denotes the discrepancy between MRI-derived spine age and chronological age, operationalized in current work as the bias-corrected predicted spine age minus actual age. In this formulation, a positive SAG indicates that the spine appears older than the person’s chronological age, a negative SAG indicates a younger-appearing spine, and a value near zero indicates alignment between MRI-estimated spine age and chronological age. The construct is used to convert spine age estimation into a candidate biomarker of overall spine health, while adjacent work in whole-body MRI, spine DXA, and sagittal morphology synthesis provides methodological support without itself defining SAG (Bazargani et al., 21 Nov 2025).

1. Definition and interpretive framework

In the explicit SAG formulation reported for whole-spine MRI, the operational quantity is

$\mathrm{SAG} = \hat{Y}_c - Y,$

where $Y$ is chronological age and $\hat{Y}_c$ is the bias-corrected predicted spine age. The underlying paper describes SAG as “the difference between actual spine age and model-predicted age” and repeatedly interprets it as “the discrepancy between chronological age and spine age,” with the sign convention clearly implying predicted-minus-chronological age (Bazargani et al., 21 Nov 2025).

The bias-correction step is central because the reported model, like many age regressors, exhibits regression toward the mean age. The correction is defined by fitting

$\hat{Y} = \alpha Y + \beta,$

then computing

$\hat{Y}_c = \frac{\hat{Y} - \beta}{\alpha}.$

The manuscript states that bias correction slightly increases MAE but improves error uniformity across age groups as captured by weighted MAE, and that for the rest of the study “spine age estimation refers to bias corrected values” (Bazargani et al., 21 Nov 2025).

The interpretive semantics of SAG are straightforward but clinically consequential. A large positive SAG is treated as accelerated spine aging or worse spine health, whereas a negative SAG is treated as a relatively younger or healthier spine appearance. The paper’s regression and odds-ratio analyses are built on this interpretation, using SAG as the dependent variable when testing associations with degenerative spine findings and lifestyle factors (Bazargani et al., 21 Nov 2025).

A recurrent misconception is that any age-prediction residual involving the torso or skeleton is equivalent to SAG. The available evidence does not support that equivalence. In particular, a whole-body age gap from neck-to-knee MRI is not a spine-specific biomarker, even when the spine is salient, because the model output can also be driven by the cardiac region, autochthonous back muscles, knees, thyroid, and abdominal fat. This distinction is explicit in the related whole-body MRI literature and is important for preserving the anatomical specificity implied by the term “Spine Age Gap” (Starck et al., 2023).

2. Reference cohort and age-model construction

The direct SAG implementation uses sagittal 3D T2-weighted whole-spine MRI and a large multi-site dataset comprising 18,070 MRI series from 17,394 individuals, acquired over 13 years (2011–2024) on 19 Philips and Siemens scanners across 10 clinics in North America. The reported age range is 25 to 84 years, with 10-year bins spanning 25–34, 35–44, 45–54, 55–64, 65–74, and 75–84 (Bazargani et al., 21 Nov 2025).

The training cohort was not defined as “no findings.” Instead, the study constructed a “normal for age” reference set from radiology reports using age-stratified clustering. Report-derived features initially included 7 structural/canal pathologies and 8 degenerative conditions tracked over cervical, thoracic, and lumbar levels, yielding a sparse vector of

$26 \times 8 + 7 = 215$

features in anomalous enumeration settings. Because radiologists often summarize findings by region, the representation was then aggregated by region, condition type, and severity, reducing the degenerative feature space to 60 features, plus structural/canal pathology features, for a final vector of 67 features (Bazargani et al., 21 Nov 2025).

Eligibility for model development was determined separately within each age bracket using UMAP followed by HDBSCAN. UMAP used the Canberra distance

$d(p,q)=\sum_{i=1}^{m}\frac{|p_i-q_i|}{|p_i|+|q_i|},$

with number of neighbors $=15$ and minimum distance $=0$ . HDBSCAN used minimum cluster size $=1\%$ of the population in each age bracket, minimum samples $Y$ 0, and merge thresholds of 0.3 for the 70s and 80s, 0.7 for the 40s and 60s, and 1.0 for the 30s and 50s. Clusters containing more than 15% of the population in an age bracket were treated as normal, and clusters containing less than 15% were treated as abnormal. This procedure yielded 10,611 normal series for age-model development and 7,459 abnormal series (Bazargani et al., 21 Nov 2025).

The resulting “normal” definition was explicitly age-conditional. In the 30-year bracket, a dominant normal cluster contained 1–3 mild lumbar disc bulges and the next cluster contained no clinically significant finding. In later decades, normal clusters could still include mild lumbar bulges and mild cervical osteophytes or bulges, with increasing counts as age increased. By the 80s, mild cervical uncovertebral osteophyte also appeared as part of normal aging. This age-conditional reference standard is foundational for SAG, because the predictor is intended to learn expected aging rather than simply detect any abnormality (Bazargani et al., 21 Nov 2025).

For the age-prediction network specifically, the normal series were split into 8,491 train, 1,051 validation, and 1,069 test. For downstream clinical association analyses, the study created a full-test set of 8,528 series by combining the 1,069 normal test cases with the 7,459 abnormal series (Bazargani et al., 21 Nov 2025).

3. Model architecture, calibration, and reliability

To constrain prediction to spinal anatomy, the study first applied a semantic segmentation model based on prior nnU-Net work. The segmentation mask included cervical, thoracic, lumbar, and sacral vertebrae, intervertebral discs, ribs, cerebrospinal fluid, and spinal cord. The mask was then dilated and used to remove non-spine regions from the MRI. All series were resampled to

$Y$ 1

and center cropped or padded to

$Y$ 2

The age regressor itself was a 3D CNN with 2,950,401 parameters, trained with Adam, learning rate 0.01, reduce-on-plateau scheduling with factor 0.3 and patience 5, batch size 2, and Mean Squared Error loss (Bazargani et al., 21 Nov 2025).

Model adequacy was examined through multiple ablations. With bias correction, the data-size ablation reported MAE 9.76, $Y$ 3, WMAE 10.55 for 85 samples; MAE 5.33, $Y$ 4, WMAE 5.50 for 850 samples; and MAE 3.67, $Y$ 5, WMAE 3.60 for the full proposed model. The loss ablation reported MAE 3.94, $Y$ 6, WMAE 4.00 for smooth-L1 and MAE 3.67, $Y$ 7, WMAE 3.60 for MSE, making MSE modestly superior in this setting (Bazargani et al., 21 Nov 2025).

Region ablation is especially relevant to the ontological scope of SAG. With bias correction, the cervical-only model achieved MAE 5.57, $Y$ 8, WMAE 5.61; thoracic-only, MAE 4.57, $Y$ 9, WMAE 4.58; lumbar-only, MAE 4.35, $\hat{Y}_c$ 0, WMAE 4.57; and whole-spine, MAE 3.67, $\hat{Y}_c$ 1, WMAE 3.60. The lumbar region was therefore the strongest single-region predictor, but the whole spine performed best, leading the authors to interpret the result as evidence that “all regions of the spine are important for the assessment of the biological spine age” (Bazargani et al., 21 Nov 2025).

On the held-out normal test set, the best model reported MAE 3.47 years, $\hat{Y}_c$ 2, and WMAE 3.60 without bias correction, and MAE 3.67, $\hat{Y}_c$ 3, and WMAE 3.60 with bias correction. The paper prefers the corrected version because it removes bias toward the mean age and makes the error more uniform across age brackets, which is crucial when the residual itself is the biomarker of interest (Bazargani et al., 21 Nov 2025).

Repeat-scan reliability was evaluated on 303 individuals with two scans in the full test set and an average interscan interval of 1.59 years. The reported SAG intraclass correlation coefficient was 0.73, with 95% bootstrap CI 0.68 to 0.78; by sex, the ICC was 0.72 (0.64, 0.79) for male participants and 0.74 (0.63, 0.81) for female participants. This is not perfect repeatability, but it indicates moderate-to-strong stability for a residualized imaging biomarker measured over a nonzero time interval (Bazargani et al., 21 Nov 2025).

4. Clinical and epidemiological correlates

The reported inferential framework uses separate linear regression models with SAG as the outcome and either degenerative findings, structural/canal pathologies, or lifestyle factors as predictors, while controlling for biological sex. The paper also compares cases with large positive SAG $\hat{Y}_c$ 4 to cases with large negative SAG $\hat{Y}_c$ 5 using odds ratios for degenerative and structural conditions (Bazargani et al., 21 Nov 2025).

Significant positive associations were observed for several lumbar and cervical degenerative findings. In the lumbar spine, disc bulge showed increasing SAG with severity, and disc osteophyte showed some of the strongest positive effects. In the cervical spine, mild and moderate disc bulges and mild and moderate disc osteophytes were also positively associated with SAG. By contrast, lumbar desiccation, annular fissure, vertebral endplate change, protrusion, and extrusion were reported as non-significant in this dataset. Most thoracic associations were likewise non-significant, with one notable exception: severe thoracic disc desiccation showed a negative association of $\hat{Y}_c$ 6 years with 95% CI $\hat{Y}_c$ 7, a result the paper does not discuss in detail and that therefore warrants cautious interpretation (Bazargani et al., 21 Nov 2025).

Several structural and canal pathologies were also associated with higher SAG. Positive effects were reported for spondylolisthesis, scoliosis, kyphosis or lordosis, fracture, spinal stenosis, and congenital spinal canal narrowing, while cord abnormalities, transitional vertebra, and Tarlov perineural cyst were non-significant. The odds-ratio analysis reinforced these regressions: moderate disc bulges had about 4-fold higher odds in positive versus negative SAG, severe disc bulges about 8-fold, and fractures, stenosis, spondylolisthesis, and canal narrowing about 2- to 4-fold higher odds (Bazargani et al., 21 Nov 2025).

Variable	SAG effect	Notes
Severe lumbar disc bulge	$\hat{Y}_c$ 8 years $\hat{Y}_c$ 9	Positive
Mild lumbar disc osteophyte	$\hat{Y} = \alpha Y + \beta,$ 0 years $\hat{Y} = \alpha Y + \beta,$ 1	Positive
Moderate lumbar disc osteophyte	$\hat{Y} = \alpha Y + \beta,$ 2 years $\hat{Y} = \alpha Y + \beta,$ 3	Positive
Spinal stenosis	$\hat{Y} = \alpha Y + \beta,$ 4 years $\hat{Y} = \alpha Y + \beta,$ 5	Positive
Fracture	$\hat{Y} = \alpha Y + \beta,$ 6 years $\hat{Y} = \alpha Y + \beta,$ 7	Positive
Spondylolisthesis	$\hat{Y} = \alpha Y + \beta,$ 8 years $\hat{Y} = \alpha Y + \beta,$ 9	Positive

Lifestyle and occupational analyses further positioned SAG as a candidate biomarker of overall spine health. Smoking showed a positive association of 0.93 years of SAG per packs-per-day smoked $\hat{Y}_c = \frac{\hat{Y} - \beta}{\alpha}.$ 0. Days per week consuming alcohol showed a statistically significant but small effect of 0.08 $\hat{Y}_c = \frac{\hat{Y} - \beta}{\alpha}.$ 1. Time sedentary was non-significant. Physically heavy work was associated with higher SAG at 0.67 $\hat{Y}_c = \frac{\hat{Y} - \beta}{\alpha}.$ 2, whereas physically moderate work was not significant. Moderate exercise and vigorous exercise were associated with lower SAG at $\hat{Y}_c = \frac{\hat{Y} - \beta}{\alpha}.$ 3 $\hat{Y}_c = \frac{\hat{Y} - \beta}{\alpha}.$ 4 and $\hat{Y}_c = \frac{\hat{Y} - \beta}{\alpha}.$ 5 $\hat{Y}_c = \frac{\hat{Y} - \beta}{\alpha}.$ 6, respectively (Bazargani et al., 21 Nov 2025).

The study also examined extreme discordance between chronological age and predicted spine age. In the full test set, 30 cases had an absolute discrepancy exceeding 15 years; expert review judged 23/30 clinically plausible, with the remainder comprising 2 MRI artifacts, 1 segmentation failure, and 4 prediction failures. Grad-CAM visualizations at block 5 suggested that the model often focused on disc bulges as major indicators of aging, but also revealed failure modes such as underweighting vertebral fracture and occasional attention partly outside the spine (Bazargani et al., 21 Nov 2025).

The direct SAG literature is currently narrow, and much of the broader methodological landscape consists of studies that do not define SAG explicitly but nonetheless constrain what a spine-specific age-gap biomarker can plausibly mean.

Paper	Relation to SAG	Reported contribution
"Atlas-Based Interpretable Age Prediction In Whole-Body MR Images" (Starck et al., 2023)	Indirect	Whole-body MRI age prediction; spine is one of three primary areas of interest
"From Baseline to Follow-Up: Counterfactual Spine DXA Image Synthesis in UK Biobank Using a Causal Hierarchical Variational Autoencoder" (Zhang et al., 21 May 2026)	Indirect	Causal age intervention in AP spine DXA through L1–L4 width, height, and area
"LatXGen: Towards Radiation-Free and Accurate Quantitative Analysis of Sagittal Spinal Alignment Via Cross-Modal Radiographic View Synthesis" (Zhao et al., 29 Sep 2025)	Indirect	Radiation-free recovery of sagittal spinal morphology and alignment from posterior RGBD

The whole-body MRI study is the most immediate precursor. It predicts chronological age from 3D neck-to-knee whole-body Dixon MRI using the water contrast, trains a 3D ResNet-18 with hidden layer size 256, and reports Test MAE: 2.57 years after bias correction. Its principal interpretability result is that the spine, the autochthonous back muscles, and the cardiac region are the three dominant areas of importance, with the cardiac region highest overall. The authors further state that “the cervico-thoracic spine region shows strong activations,” that “the focus on the spine increases with age,” and that the accelerated age group shows stronger activations in the spine and autochthonous muscles than the decelerated group. This does not yield SAG, because the model is whole-body rather than spine-only, but it supports the biological plausibility of spine-centered age inference and demonstrates atlas-based population saliency mapping as a transferable methodology (Starck et al., 2023).

The AP spine DXA study introduces a different kind of foundation. It does not predict age from DXA and does not define SAG, but it models the effect of an age intervention $\hat{Y}_c = \frac{\hat{Y} - \beta}{\alpha}.$ 7 in a metadata-conditioned CHVAE with an explicit SCM and AAP counterfactual inference. The image decoder is conditioned on transformed L1–L4 width, L1–L4 height, and L1–L4 area, and longitudinal evaluation on 319 repeat-imaged participants reports strong absolute-level agreement between counterfactual and observed follow-up morphometry: $\hat{Y}_c = \frac{\hat{Y} - \beta}{\alpha}.$ 8 for width, 0.854 for height, and 0.931 for area, with sign agreement = 0.615 at the change level. This suggests that a future causal SAG could be defined through inverse age inference or through supervised regression on an age-responsive morphometric state rather than through a purely black-box residual (Zhang et al., 21 May 2026).

The sagittal-alignment synthesis study is further removed from direct SAG, because it focuses on adolescents aged 10–18 with AIS, not age regression. Its significance lies in the extraction of sagittal structure from non-radiographic input. The model synthesizes a lateral radiograph from posterior RGBD through a dual-stage pipeline comprising SME and LRS, with attention-based FFC, SDN, and SLS. Downstream measurements derived from synthesized radiographs achieve $\hat{Y}_c = \frac{\hat{Y} - \beta}{\alpha}.$ 9 for TKA, 0.896 for LLA, and 0.870 for SSA. A plausible implication is that sagittal curve maps, SME latent features, and radiograph-derived sagittal parameters could serve as age-sensitive features in a future morphology-driven SAG system, but the paper itself does not validate any age-gap construct (Zhao et al., 29 Sep 2025).

6. Limitations, misconceptions, and open problems

The current evidence base supports SAG as a candidate biomarker, not a definitive one. The direct MRI study provides only internal validation within its own dataset; no external cohort is reported. Its association analyses are observational, not causal, and the reported models control for biological sex but not for a broader confounder set such as body habitus, pain status, osteoporosis, prior injury, genetics, socioeconomic variables, or scanner/protocol variation. Because SAG is a residual-like quantity, segmentation failures, age-prediction errors, and calibration artifacts directly propagate into the biomarker itself (Bazargani et al., 21 Nov 2025).

The definition of “normal for age” is also a methodological strength and a limitation simultaneously. It is sophisticated relative to a naïve “no findings” rule, but it is still based on report-derived clustering with parameter choices such as the 15% cluster threshold and age-specific HDBSCAN merge distances. The resulting normality reference is therefore not an external gold standard. In addition, the source dataset was drawn primarily from preventive health screening, which limited the frequency of rare or severe conditions and weakened some analyses involving uncommon findings (Bazargani et al., 21 Nov 2025).

A second misconception is that spine saliency in a non-spine model is sufficient to establish SAG. The whole-body MRI study explicitly shows why that inference is too strong: the cardiac region exhibits the highest importance, there is no spine-only performance evaluation, and the residual remains a whole-body age gap rather than a spine-specific one. Its contribution is foundational evidence for spine informativeness, not a validated SAG implementation (Starck et al., 2023).

The adjacent DXA and sagittal-synthesis literatures solve different problems. The CHVAE study offers age-causal morphometric modeling but reports no age MAE, age correlation, or age-gap distribution, so it cannot establish a reliable spine age estimator. The LatXGen study offers anatomically meaningful sagittal representations, but its cohort is AIS-only, 10–18 years, and predominantly female, with no direct age-regression experiment. These constraints mean that both studies are better understood as scaffolds for future SAG systems than as SAG evidence themselves (Zhang et al., 21 May 2026, Zhao et al., 29 Sep 2025).

Open problems follow directly from these limitations. The direct MRI paper proposes more data, especially for rare and severe spine conditions; newer architectures such as vision transformers; possible replacement of UMAP/HDBSCAN with encoder-decoder dimensionality reduction; and extension of the “normal aging + age gap + clinical relevance” framework to other organs. A plausible implication from the DXA literature is that inverse age intervention in a causal generative model could eventually yield a more explicitly biological, longitudinally grounded form of SAG, but this remains untested. At present, the most defensible formulation is the narrow one: SAG is the bias-corrected discrepancy between MRI-estimated whole-spine age and chronological age, validated internally and associated with multiple degenerative, structural, and lifestyle variables, yet still awaiting broader external, causal, and longitudinal confirmation (Bazargani et al., 21 Nov 2025).