AI-derived CT morphometric phenotypes predict survival, functional decline, and surgical morbidity following curative-intent surgical sarcoma resection

Abstract Background Outcomes after curative-intent sarcoma surgery vary substantially and are incompletely explained by tumor-centered factors alone. Although CT-based body composition metrics provide objective host-related information, most sarcoma studies rely on isolated parameters or binary sarcopenia definitions. AI-driven analytical approaches offer the opportunity to integrate multidimensional morphometric data into data-driven phenotypes that may better capture clinically relevant heterogeneity. Methods In this retrospective cohort study, our institutional sarcoma database (n = 2667) was screened to identify patients with osteosarcoma, myxofibrosarcoma, liposarcoma, or chondrosarcoma who underwent curative-intent surgical resection and had a preoperative CT including mid-L3 vertebral level for morphometric analysis (final cohort n = 234). Skeletal muscle index (SMI), skeletal muscle density (SMD), and visceral adipose tissue area (VAT) were quantified from a single axial mid-L3 slice. Unsupervised k-means clustering of standardized SMI, SMD, and VAT identified AI-derived morphometric phenotypes. Outcomes included overall survival (OS), ECOG performance status at follow-up, surgical site infection (SSI requiring surgical revision), and length of hospital stay (LOS). Multivariable regression models evaluated independent associations between phenotypes and outcomes, adjusting for relevant clinical covariates. Results Clustering identified four phenotypes: muscle-preserved (n = 88), myosteatotic (n = 62), sarcopenic (n = 56), and cachexia-like (n = 28). Morphometric profiles differed markedly: muscle-preserved (SMI 47.8 ± 6.3 cm 2 /m 2 ; SMD 41.6 ± 6.4 HU; VAT 118 ± 56 cm 2 ), myosteatotic (SMI 42.1 ± 5.9; SMD 28.9 ± 5.3; VAT 142 ± 62), sarcopenic (SMI 35.6 ± 4.8; SMD 34.0 ± 5.6; VAT 96 ± 48), and cachexia-like (SMI 31.2 ± 4.4; SMD 26.4 ± 4.9; VAT 64 ± 35). Median OS differed significantly across phenotypes (155 vs. 32 vs. 64 vs. 19 months; p < 0.0001). Postoperative functional status also worsened stepwise (median ECOG at follow-up: 1 ± 0.5 vs. 2.5 ± 1 vs. 3 ± 1 vs. 3 ± 0.5; p < 0.0001). In multivariable Cox regression, cachexia-like (HR 3.28, 95% CI 2.01–5.36; p < 0.001) and sarcopenic phenotypes (HR 1.89, 95% CI 1.26–2.83; p = 0.002) independently predicted mortality, whereas conventional sarcopenia did not. SSI rates increased across phenotypes (6.8% to 21.4%; p = 0.042), cachexia-like (HR 3.21, 95% CI 1.69–6.10; p < 0.001) and sarcopenic phenotypes (HR 2.08, 95% CI 1.17–3.70; p = 0.012) were independently associated with SSI. LOS was independently prolonged in sarcopenic (+ 3.4 days, p = 0.002) and cachexia-like patients (+ 6.2 days, p < 0.001). Conclusions AI-derived CT morphometric phenotypes obtained from routine preoperative imaging identify distinct host profiles in sarcoma patients and independently predict survival, postoperative functional decline and postoperative morbidity beyond conventional CT-based sarcopenia assessments. Integrating morphometric phenotyping into preoperative assessment may support risk stratification, counseling, and targeted perioperative optimization in curative-intent sarcoma surgery.