Many older surgical patients already have a CT scan in their record before the operation. It might be a cancer staging scan, or a scan taken to plan spine hardware. These images contain measurements nobody reads: how dense the vertebral bone is, how much fat has infiltrated the back muscles, how large and dense the abdominal fat is, how much calcium sits in the aorta. Software can now extract these numbers automatically. Separately, health economists have found that when a complication occurs, the extra hospital spending does not come mainly from the operation itself. It comes from supportive care: intensive monitoring, oxygen, repeated laboratory tests and extra bed-days. This proposal connects the two observations. We test whether body-composition numbers taken automatically from the preoperative CT predict the supportive-care portion of a patient's hospital bill, after the procedure itself is accounted for.
The choice of target matters. Predicting total hospital cost mostly rewards knowing which operation was performed, and that information swamps any signal about the patient's frailty. Our core claim is that tissue measurements from routine CT carry information about physiological reserve, meaning how well the body tolerates stress. That reserve should act on the cost that remains after removing the procedure-determined part, and especially on the upper tail of that remaining cost. It should not act on the fixed procedural cost at all. The claim is narrow enough to be wrong in informative ways: CT may add nothing beyond serum albumin and comorbidity scores, or an uninterpretable whole-image network may capture the same signal.
What exists, and what is different
The motivating evidence comes from the economic side. In a propensity-matched spinal tuberculosis cohort, Liu, Chen et al. found that complicated cases cost about 51,700 versus 42,800 RMB. The difference came from monitoring, oxygen and laboratory charges, and postoperative albumin supplementation was the only independent predictor. They used no imaging. Gao and Liu modelled colorectal cancer hospitalization cost with random forests and reached test R² 0.833. The dominant features were major procedure and length of stay, so a large part of that accuracy comes from information known only at discharge. The authors also note that their data were limited to front-page discharge summaries with no physiologic or imaging features. Our design takes both lessons into account. Length of stay is an outcome, not an input, and the procedure is conditioned out of the target rather than used to inflate fit.
The method comes from the screening side. Comp2Comp is an open-source pipeline that segments muscle, visceral fat and subcutaneous fat with Dice scores around 0.96–0.97 internally. Its tissue-attenuation errors are below 1% and its trabecular-region errors are below 7.1 HU. We reuse it rather than building new segmentation. Pickhardt et al. showed that the same kind of automated biomarkers beat demographics for 10-year survival (AUC 0.880 vs 0.779). Muscle attenuation, aortic plaque, visceral fat density and bone density carried most of the signal. That study targets long-horizon mortality, not near-term inpatient resource use. Kong et al. showed that combining muscle with bone beats bone alone for fracture prediction. We test whether the same holds for perioperative cost.
Three papers come closest to this proposal, and each differs in a specific way. Löffler et al. used opportunistic vBMD from planning CT to predict pedicle screw loosening (AUC 0.769 in 46 patients). That work covers one mechanical complication, bone only, with no costing. The UCSF chest-radiograph study regressed 1-, 3- and 5-year aggregate spending from emergency-department radiographs with an end-to-end network. It is uninterpretable, is not surgical, and does not distinguish acute supportive care from everything else a patient spends. The oQCT Markov model showed that opportunistic bone screening is cost-saving for outpatient fracture prevention. It explicitly excludes inpatient surgical care. None of the three uses itemized inpatient categories, procedure-adjusted cost as the target, or soft-tissue biomarkers in surgical patients.
Core Claim and Target
For patient $i$ we define excess cost as
$$E_i = C_i - \tilde{C}_{s(i)},$$
where $C_i$ is total inpatient cost and $\tilde{C}_{s(i)}$ is the median cost of uncomplicated patients in the same procedure-code, hospital and year stratum $s(i)$. We also break $E_i$ into its itemized supportive categories: monitoring, oxygen, laboratory, bed-days and medications. The claim has three parts. First, a small set of CT tissue biomarkers improves prediction of $E_i$, and of membership in its top decile, beyond a strong preoperative tabular model that includes albumin. Second, the improvement is concentrated in the supportive categories. Third, the biomarkers carry no information about the fixed procedural component, meaning operating-room and implant charges. That third part is a built-in falsification test: if "frailty" features predicted implant cost, they would be acting as a proxy for case mix rather than measuring reserve.
Research Program
We will assemble a retrospective cohort at two or more hospitals that issue itemized inpatient bills. Itemized billing of this kind is standard in the Chinese tertiary centres behind the spinal tuberculosis and colorectal studies. The cohort covers two procedure families where a preoperative CT routinely includes the L1–L3 levels: colorectal cancer resection, with its staging CT, and lumbar instrumentation, with its planning CT. Inclusion requires age at least 50 and a CT within 90 days before surgery.
Comp2Comp extracts about eight prespecified biomarkers. These are L3 skeletal muscle index and muscle attenuation (the attenuation is the myosteatosis signal), visceral fat area and density, subcutaneous fat area, L1–L2 trabecular HU converted to vBMD, and an aortic calcification score. Contrast phase, kVp and slice thickness are recorded as harmonization covariates. Spine-hardware scans use the metal-artifact masking described by Löffler et al.
The models are gradient-boosted and quantile regressions on the tabular features, the CT features, and both together. Outcomes are $E_i$, its supportive sub-components, top-decile membership, and coded complications. A second analysis restricts to patients with no coded complication. This tests whether low reserve also raises supportive spending through slower recovery that never receives a complication code, which is the part of the cost that complication-based models cannot see.
Baselines
The naive reference is the stratum median itself. It predicts $E_i = 0$ for everyone, and its top-decile ranking is chance. Any model that cannot beat it on Spearman correlation and on 0.9-quantile loss has learned nothing beyond the procedure.
The strongest existing competitor is a locally trained preoperative tabular model. Its inputs are age, sex, Charlson index, ASA class, BMI, preoperative albumin and hemoglobin, procedure level and payment type. It is fit with both Lasso and gradient boosting, and the better of the two is kept. This is the right opponent for two reasons. Albumin was the only independent predictor in the spinal tuberculosis cohort. And in Transfer Learning via Latent Factor Modeling, local Lasso on Duke data reached AUROC 0.76 for surgical complications, beating a transfer-learned latent model at 0.73. That result tells us not to assume added complexity helps. The best tabular cost result in the file is RF R² 0.833 (Gao and Liu). It is not comparable to our setting, because it uses length of stay as an input and targets total cost. We found no reported result on Lacuna for preoperative prediction of procedure-adjusted or itemized supportive-care cost from imaging. Tabular and CT-augmented models get identical patients, folds, preprocessing and hyperparameter search budgets, with 100 random-search trials each under nested cross-validation. Beating this baseline means a gain of at least 0.03 in top-decile AUROC, with a bootstrap 95% CI excluding zero, under both temporal and cross-hospital validation, plus a significant reduction in 0.9-quantile loss.
The image-side competitor is an end-to-end 3D network trained on the same preoperative CT volumes, with and without tabular fusion. It mirrors the TX1 design of the UCSF study, which reported 1-year top-50%-spender AUROC 0.806 and Spearman 0.561 on a different population and target. It gets the same folds and a matched compute budget.
Two ablations isolate the key ingredients. A bone-only CT model, with vBMD added to the tabular features in the style of Löffler et al., tests whether soft tissue matters. Refitting every model on total cost instead of $E_i$ tests whether the procedure-adjusted target is what makes the CT signal visible.
Why not just…?
Why not just use albumin, Charlson index and ASA class? This is the most serious threat. Albumin already summarizes nutritional and inflammatory reserve, and the spinal tuberculosis study found nothing else independent. The biological-age study shows CT biomarkers add to demographics, but its comparison model included no albumin and its outcome was mortality, so it does not settle the question. Our head-to-head comparison against the fully specified tabular baseline is the test. If the gain is below 0.02 AUROC, we will report that tabular data suffice for cost forecasting. The remaining CT contribution would then be the subgroup of patients with missing preoperative labs.
Why not just regress cost from the raw scan with a deep network? It may match or beat the handcrafted biomarkers on accuracy. The UCSF result shows whole-image models do find spending signal. If it wins clearly, we keep it as the predictor. The biomarkers would then serve to explain what the network has learned, and we would check this by measuring how much of the network's predictive signal the eight biomarkers account for. The argument for explicit biomarkers is not purely accuracy. They transfer across scanners through HU calibration, and they correspond to interventions such as nutrition and prehabilitation. They also make the procedural-cost falsification test interpretable. With a few thousand patients, a 3D network is also likely to overfit to the site, which the cross-hospital split will reveal.
Why not just predict complications and multiply by average complication cost? That approach ignores two things. Supportive spending varies widely within complicated patients, and reserve may raise cost among patients with no coded complication. The complication-free subgroup analysis measures this directly. If CT predicts nothing there, the simpler two-stage approach is adequate and we will say so.
Why not just flag patients below the ACR cutoff of 80 mg/cm³? A single bone threshold suits mechanical failure, as Löffler et al. found. Systemic supportive-care demand plausibly depends more on muscle. The bone-only ablation and a threshold rule will both be evaluated against the multi-tissue model.
Decisive Checks
The proposal succeeds only if four results hold together. First, the CT-augmented model beats the tabular model on top-decile excess cost under temporal and cross-hospital validation. Second, the gain shrinks or disappears when total cost is the target, which shows the adjusted target is what exposes the signal. Third, the biomarkers have near-zero importance for operating-room and implant charges. Fourth, a meaningful share of the gain persists among patients without coded complications.
We will also report a cost-capture curve: the fraction of total supportive spending that falls in the top 10% of patients ranked by each model. This translates prediction quality into the quantity a hospital budget actually experiences. Calibration of the 0.9-quantile predictions across contrast phases tests whether HU drift, the heterogeneity problem flagged for opportunistic CT, undermines the approach.
References
- Comparison of Clinical Data Between Patients With Complications and Without Complications After Spinal Tuberculosis Surgery: A Propensity Score Matching Analysis
- Prediction and the influencing factor study of colorectal cancer hospitalization costs in China based on machine learning-random forest and support vector regression: a retrospective study
- Comp2Comp: Open-Source Body Composition Assessment on Computed Tomography
- Biological age model using explainable automated CT-based cardiometabolic biomarkers for phenotypic prediction of longevity
- A Computed Tomography–Based Fracture Prediction Model With Images of Vertebral Bones and Muscles by Employing Deep Learning: Development and Validation Study
- Opportunistic Osteoporosis Screening Reveals Low Bone Density in Patients With Screw Loosening After Lumbar Semi-Rigid Instrumentation: A Case-Control Study
- Prediction of future healthcare expenses of patients from chest radiographs using deep learning: a pilot study
- Cost-effectiveness of opportunistic QCT-based osteoporosis screening for the prediction of incident vertebral fractures
- Transfer Learning via Latent Factor Modeling to Improve Prediction of Surgical Complications