Characterization of Plan Complexity and Its Role in Quality Assurance for AI-Assisted CBCT-Based Online Adaptive Radiotherapy of Prostate Cancer

Background/Objectives: Online adaptive radiotherapy (oART) generates plans at each fraction by exploiting AI-assisted optimization engines without explicit user control over modulation. This process challenges quality assurance since measurement-based Patient Specific Quality Assurance (PSQA) cannot be performed daily. This study aimed: (i) to characterize plan complexity in IOE-generated plans for prostate cancer using a reproducible set of PCMs, including the decomposition of inter-patient and intra-patient variability sources; (ii) to evaluate the association between PCMs and delivery accuracy within a cohort-informed SPC framework validated through leave-one-patient-out cross-validation; (iii) to investigate whether inter-fraction anatomical variations explain the observed plan complexity patterns, or whether complexity is predominantly an intrinsic signature of the AI-assisted optimizer. Methods: Twenty-one prostate cancer patients treated on a CBCT-based oART platform were retrospectively analyzed across three anatomical targets: prostatic bed (PrB), prostate (Pr), and prostate with seminal vesicles (PrSV). Six PCMs, namely MU/cGy, Modulation Complexity Score (MCS), Aperture Area Variability (AAV), Leaf Sequence Variability (LSV), Average Leaf Gap (ALG) and Plan Irregularity, were extracted. Additionally, five anatomical metrics (AMs) were computed from daily contours. Linear mixed-effects models (LMEMs) compared reference/online plans, decomposed variance via intraclass correlation coefficients (ICCs), and assessed PCM–gamma passing rate (GPR) associations. Leave-one-patient-out cross-validation (LOPO-CV) evaluated SPC threshold stability. The relationships between PCMs and AMs were investigated using LMEMs. Results: The AI-assisted optimization engine generated plans characterized by elevated monitor unit demand (average MU/cGy ≥ 6.8 ± 0.9) and narrow MLC apertures (ALG ≤ 17.7 mm ± 1.9 mm). No complexity differences emerged between offline and online-adapted plans, nor between anatomical targets. All PCMs showed significant associations with global GPR (p ≤ 0.027), though marginal R² remained low (≤ 0.122). Notably, GPR dispersion increased systematically at higher complexity values, indicating that highly modulated plans exhibit reduced delivery predictability. LOPO-CV demonstrated stable tolerance/action limits. Anatomical variations explained less than 35% of the total variance in PCMs. Conclusions: Plan complexity in oART reflects the optimization paradigm and patient-specific anatomy rather than daily adaptation. PCMs can serve as surveillance indicators flagging high-risk fractions to support SPC-based monitoring.