L1-S1 Posterior Paraspinal Muscle Fatty Infiltration and a Radiomics Signature Predict Prolonged Length of Stay After Lumbar Spine Surgery in Older Adults

Retrospective cohort study. To determine whether multilevel L1 to S1 paraspinal muscle fatty infiltration (FI) and a posterior paraspinal radiomics signature improve prediction of prolonged length of stay (LOS) after lumbar spine surgery in geriatric patients. We retrospectively included 248 patients aged older than or equal to 75 years undergoing open posterior transforaminal lumbar interbody fusion (TLIF) with preoperative axial T2-weighted MRI covering L1 to S1. Paraspinal muscles were segmented (MuscleMap) to derive global L1 to S1 fatty infiltration (all muscle FI) and CSA/BMI. A posterior paraspinal radiomics score (RadScore) was developed from multifidus+erector spinae radiomics features using L1-penalized logistic regression within leakage-free nested cross-validation (outer five-fold; inner five-fold). Prolonged LOS was defined as LOS ≥16 days (75th percentile). Discrimination (AUC), calibration, and clinical utility (DCA) were assessed using out-of-fold predictions; bootstrap 95% CIs were reported. Prolonged LOS occurred in 62/248 (25.0%). Patients with prolonged LOS had lower BMI and a markedly higher prevalence of frailty (Fried ≥3: 87.1% vs. 22.6%). All muscle FI was strongly associated with prolonged LOS after adjustment for clinical and operative factors, and RadScore remained independently associated in radiomics-augmented models; in the combined model, the association for All muscle FI was attenuated, suggesting shared prognostic information between conventional FI and radiomics-derived muscle heterogeneity. In leakage-free nested cross-validation, the clinical model achieved AUC 0.848, which improved to 0.922 after adding All muscle FI, and to 0.933 with RadScore; the combined model yielded the highest AUC (0.936). In older adults undergoing lumbar fusion, global multilevel paraspinal degeneration measured by conventional FI provides major incremental value for predicting prolonged LOS beyond clinical and geriatric factors, whereas posterior paraspinal radiomics offers an additional but more modest improvement. Leakage-free validation supports the robustness and clinical relevance of integrating automated muscle quantification with imaging-based risk stratification.