Purpose: The purpose of the study was to optimize the Patient-specific Quality Assurance (PSQA) process using artificial neural networks (ANNs), log file data and complexity indices (CIs) to identify treatment plans susceptible to deliverability issues. Methods: Log file accuracy was validated. CIs and Gamma analysis results from log files were evaluated for their ability to discriminate treatment plans through comparison with experimental gamma results. ANNs were trained to classify plans using log file gamma data and CIs, with experimental gamma results as the reference standard. Results: Log files data demonstrated high accuracy and reproducibility. Gamma passing rates from log files exceeded experimental measurements (99.0% ± 0.8% vs. 95.9% ± 2.8%). However, the correlation between experimental and log file-based gamma results was insufficient for standalone plan classification. Among the evaluated CIs, maximum leaf travel per MU, maximum gantry rotation, and beam delivered energy demonstrated strong predictive performance. In contrast, field irregularity and leaf-gantry synchronization showed limited predictive value, and small field contribution was the least informative. Multi-index ANN analysis achieved an average sensitivity of 0.70 ± 0.10, specificity of 0.68 ± 0.17, and an F1-score of 0.68 ± 0.08 across 20 training sessions. Discussion and Conclusions: The reliability of log file parameters depends on the accuracy of linac calibration and its thorough quality assurance procedures. Although gamma analysis alone has limited capacity to detect delivery issues, integrating CIs and log file data within a multifactorial machine-learning classification significantly enhances predictive accuracy. This approach provides a rational alternative to experimental PSQA verification, improving workflow efficiency.
Banos-Capilla et al. (Thu,) studied this question.