A routine CT scan already contains a precise measurement of how much muscle a patient has, and that muscle is the body's reserve of protein and energy under surgery, cancer treatment or critical illness. Automated segmentation can now extract this measurement reliably. A simple whole-body ratio, muscle volume divided by bone volume, predicts cancer survival better than body mass index or the traditional single-slice measurement. Interpretable prognostic models for sarcopenia, meaning pathological muscle loss, have not kept pace with this. They either run explanation tools over lab values and tape-measure proxies, or they feed one image-derived number into a standard regression.
We propose a survival model that starts from that single muscle-to-bone ratio. It adds finer detail only when the detail measurably earns its place. That detail can be which body region the muscle sits in, how much fat has infiltrated it, or how fast it was lost between two scans. We pair this model with an audit. A deep image network is trained only on the risk the transparent model fails to explain, and its attention is mapped onto a shared anatomical template. The result is a readable model plus a map of where the missing information lives, or strong evidence that nothing important is missing.
Core Claim
In prognosis after major cancer treatment, the interpretability question clinicians care about is not "which pixels did the network look at." It is "does anything beyond overall muscle reserve matter, and if so, what." We claim this can be answered by pairing two components. The first is a reluctant additive survival model: a linear baseline on the muscle-to-bone ratio, plus penalized smooth terms for compartment-specific volumes, muscle density and change over time. The smooth terms are penalized more heavily than the linear ones, so nonlinearity must beat the simple story to be kept. The second is a residual audit. It asks whether an opaque image model can still find risk the glass-box model missed, and it localizes any such risk anatomically.
What Exists and Where This Departs
The prognostic signal is well established and motivates the work. Borys et al. introduced the Sarcopenia Index (SI), muscle volume over bone volume computed by the Body and Organ Analysis segmentation tool on 10,340 cancer patients. They reported mortality hazard ratios of 0.56 (abdomen) and 0.40 (thorax) for high SI. Their model uses one static scalar. It does not ask which compartments carry the signal, and it does not use muscle density. Li et al. showed in 198 gastric cancer patients that 30-day postoperative muscle loss is a stronger mortality predictor (HR 11.231) than baseline sarcopenia (HR 2.332), using 2D slice indices and fixed cut-offs. The ICU meta-analysis of Zhang et al. reports a pooled OR of 2.28 but different cut-offs across populations. That is a reason to learn smooth risk curves rather than impose thresholds.
Existing interpretable sarcopenia models set the bar for simplicity, and they also constrain the design. Liu et al. on hemodialysis found that logistic regression beat boosted trees (test AUC 0.873). Liu et al. on older inpatients found a decision tree (AUC 0.7072) far ahead of a neural network (0.5665). The NHANES study used Random Forest with SHAP, a post-hoc method that splits a prediction into per-feature contributions. All three use tabular or anthropometric inputs with no imaging. Together they warn that added capacity often fails on small clinical cohorts, which is exactly why our model is built to be reluctant.
The method is borrowed. Maronge et al. (RAITR) fit a sparse linear model first. They then fit penalized splines to what the linear model leaves unexplained, penalizing the nonlinear terms more heavily. We transfer this from treatment rules to a Cox survival likelihood, the standard regression for time-to-death. NAMformer motivates keeping marginal effects identifiable while allowing interactions. We use it as a higher-capacity comparator, not as our model.
The audit builds on Starck et al.. They registered Grad-CAM heatmaps (gradient-based maps of where a network attends) from a 3D ResNet age predictor on UK Biobank MRI into a common atlas. They found deep back muscles to be a dominant aging signal, but noted that the maps are too coarse to yield measurements. We change what the saliency is asked about. Instead of explaining the whole prediction, it explains only the residual the measured quantities cannot, so a hotspot points to an anatomical quantity we should measure next.
Research Program
Measurement comes first. We train a 3D nnU-Net on SAROS, 900 whole-body CTs whose baseline muscle Dice is 0.971. We use Comp2Comp for vertebral levels T12–L5, muscle density in Hounsfield units, and visceral and subcutaneous fat. The SAROS muscle class is a single label, so we define compartments by vertebral and body-region boundaries: thoracic, abdominal-wall-plus-psoas at L1–L5, and paraspinal. Each compartment volume is normalized by bone volume, following Borys. For patients with two scans, change terms are computed per day of interval.
The model is written as follows:
$$\log h(t\mid x) = \log h_0(t) + \beta\,\mathrm{SI} + \gamma^\top z + \sum_k g_k(x_k) + \sum_k d_k(\Delta x_k).$$
Here $h_0$ is the baseline hazard, $z$ holds clinical covariates (age, sex, stage, treatment), and $x_k$ are the compartment, density and fat features. The terms $g_k$ and $d_k$ are penalized splines on static features and on changes $\Delta x_k$ respectively.
Fitting follows RAITR's two stages. First, a LASSO Cox model on SI and $z$. Second, spline terms fit against the stage-one residual risk, with smoothness and inclusion penalties scaled by a reluctance factor $\rho>1$ relative to the linear penalty. We choose $\rho$ on a development site only, never on the test site. Every retained term can then be drawn as a one-dimensional risk curve.
For the audit, we freeze the glass-box model and train a 3D CNN survival head whose output is added as an offset to the glass-box log-hazard. By construction, the CNN can only contribute risk the glass-box lacks. Its Grad-CAM maps are registered to an atlas as in Starck et al. and averaged over patients where the residual contribution is large. We then test whether a candidate measurement read off the hotspot, for example paraspinal fat infiltration, closes the gap when added back to the glass-box model.
The target cohort is retrospective oncologic surgery patients with pre-treatment and roughly one-month post-treatment CT and overall survival, drawn from at least two institutions: one for development, one held out. No such cohort is public in our file, so data access is the main practical risk. A second setting, using the same code, is UK Biobank MRI age prediction. There, the 3D ResNet-18 benchmark of 2.57 years mean absolute error from Starck et al. gives a public test of whether compartment measurements plus an audit can nearly match an opaque model.
Baselines
The naive reference is a Cox model on clinical covariates and body mass index alone. It represents what a clinic can do without any image processing. Beating it is necessary but not interesting.
The first imaging baseline is the same Cox model plus the Comp2Comp L3 skeletal muscle index, the consensus single-slice standard. Laousy et al. (MIDL, DQN localization) report that automated L3 area differs from reference by 5.06 cm² against human interobserver variability of 6.12 cm². Measurement noise therefore cannot excuse a loss to this baseline.
The strongest simple competitor is Cox plus SI, reproducing Borys et al. Their reported numbers are hazard ratios (0.56 abdomen, 0.40 thorax). We found no reported concordance index on Lacuna for SI-based survival models, so we will compute Harrell's C, time-dependent AUC and integrated Brier score for SI-Cox ourselves on identical splits.
For the dynamic setting, the comparator is Li et al.'s rule of thresholded 30-day muscle loss (≥7.23% men, ≥6.57% women; reported HR 11.231), added to SI-Cox.
The strongest flexible competitor is gradient-boosted survival trees with SHAP on exactly the same feature set. The second flexible competitor is NAMformer adapted to survival. Both get the same folds and a tuning budget matched in number of configurations to the reluctant model's search over $\rho$ and spline penalties.
The end-to-end competitor is the 3D CNN survival model trained from the scan alone, using the same architecture as the audit network.
Two ablations isolate our ingredients. The first sets $\rho=1$, giving an ordinary penalized GAM (generalized additive model), which tests whether reluctance matters. The second removes the change terms, which tests whether dynamics matter.
Beating SI-Cox means a statistically supported gain in held-out-site concordance, by bootstrap interval excluding zero, with calibration no worse. Matching boosted trees within their bootstrap interval while producing more stable explanations counts as a success on the interpretability axis.
Why not just…?
Why not just put SI and clinical covariates into a Cox model? This is the real threat. It is already transparent, and Borys et al. show SI carries a large effect. If the reluctant model shrinks back to SI-Cox, we will report that. A clean negative is itself useful to clinicians: one number suffices. The change-term ablation and the held-out-site comparison settle whether compartment and trajectory detail add anything. Li et al.'s tenfold-larger hazard for muscle loss than for static sarcopenia is the main reason to expect they will.
Why not just run XGBoost with SHAP? On small clinical cohorts, the two Liu et al. studies found flexible learners did not beat simple ones. SHAP rankings from correlated body-composition features, such as compartment volumes that move together, are prone to instability. We measure this directly as the rank agreement of top features across bootstrap resamples and across sites, for SHAP and for our shape functions. If boosted trees are both more accurate and equally stable, the glass-box model loses on the merits.
Why not just train an end-to-end image network and show Grad-CAM? The Freeman study found direct classification reached 92% accuracy with false negatives, where segmentation-then-threshold was fully reliable. Starck et al. themselves describe Grad-CAM as too coarse for measurement. Our audit still gives this approach a fair chance: if the end-to-end network greatly outperforms the glass-box model, the residual stage will show a large incremental concordance, and the atlas map will say where.
Why not just use bedside grip strength or calf circumference? These are unavailable for sedated or ventilated patients (Segers et al.), and body mass index misses sarcopenic obesity (Zhang et al.). More practically, retrospective CT cohorts rarely record them. Where a site does record grip strength, we add it to the naive and SI baselines.
Decisive Checks
The first check is whether the reluctant model beats SI-Cox in concordance at the held-out institution. The second is whether the retained terms are mainly change terms, which would replicate Li et al. volumetrically without cut-offs. The third is whether the $\rho=1$ ablation matches accuracy but yields less stable curves across bootstrap samples, which is what the reluctance hypothesis predicts.
For the audit, the key quantity is the incremental concordance of the residual CNN over the frozen glass-box model, tested against a permutation null. A near-zero value certifies that the transparent model is not leaving prognostic signal on the table. A clearly positive value must be followed by the confirmatory step: the hotspot-derived measurement, added to the glass-box model, should recover most of that increment.
On UK Biobank, the parallel check is how close a compartment-plus-density additive model comes to 2.57 years mean absolute error. We also check whether the residual attention there concentrates on the back muscles Starck et al. identified, or elsewhere.
References
- Leveraging Sarcopenia index by automated CT body composition analysis for pan cancer prognostic stratification
- Preoperative sarcopenia and postoperative accelerated muscle loss negatively impact survival after resection of locally advanced gastric cancer
- Sarcopenia as a predictor of mortality among the critically ill in an intensive care unit: a systematic review and meta-analysis
- Development of an interpretable machine learning model for predicting sarcopenia in patients undergoing maintenance hemodialysis
- Machine learning based on nutritional assessment to predict adverse events in older inpatients with possible sarcopenia
- Exploring the link between the ZJU index and sarcopenia in adults aged 20–59 using NHANES and machine learning
- A reluctant additive model framework for interpretable nonlinear individualized treatment rules
- Beyond Black-Box Predictions: Identifying Marginal Feature Effects in Tabular Transformer Networks
- Atlas-Based Interpretable Age Prediction In Whole-Body MR Images
- SAROS: A dataset for whole-body region and organ segmentation in CT imaging
- Comp2Comp: Open-Source Body Composition Assessment on Computed Tomography
- Automated L3-based sarcopenia quantification in CT scans
- Deep Reinforcement Learning for L3 Slice Localization in Sarcopenia Assessment
- Deep Learning-Assisted Detection of Sarcopenia in Cross-Sectional Computed Tomography Imaging
- Feasibility of neuromuscular electrical stimulation in critically ill patients