Uncertainty quantification-guided patient-specific quality assurance using Bayesian neural networks based on field complexity features and fluence maps

<i>Objective.</i>Recent advancements in artificial intelligence (AI)-driven prediction models for measurement-based patient-specific quality assurance (PSQA) necessitate uncertainty quantification (UQ) to ensure clinical safety.<i>Approach.</i>An uncertainty-guided framework was proposed for PSQA prediction. A classification model utilizing field complexity features and fluence maps was trained to categorize PSQA outcomes. Monte Carlo approximate Bayesian inference-based UQ method was employed to approximate the distribution of classification results. Pre-defined clinical uncertainty thresholds were defined via Correct-Certain (CC) and Incorrect-Uncertain (IU) curves to trigger manual intervention for high-uncertainty cases. A Multilayer Perceptron then predicted<i>γ</i>passing rates (GPR) by integrating classification model embeddings with uncertainty metrics. Prospective test was conducted on clinical test sets to systematically assess clinical reliability of the proposed framework.<i>Main results.</i>The classification model achieved sensitivities of 83.33% at 3%/3 mm, 93.33% at 3%/2 mm, and 94.74% at 2%/2 mm. The pre-defined clinical uncertainty thresholds were determined from CC and IU curves, ranged from 0.057-0.240 at 3%/3 mm, 0.014-0.207 at 3%/2 mm, and 0.080-0.244 at 2%/2 mm. Only 57.27% of manual intervention was required to ensure clinical safety, achieving 100% clinical sensitivity at 3%/3 mm. Compared to conventional measurement-based PSQA, this reduces the workload by 42.73%. The GPR prediction model yielded mean absolute errors of 1.64%, 1.88%, and 2.57% at 3%/3 mm, 3%/2 mm, and 2%/2 mm<i>γ</i>criteria, respectively, with uncertainty integration leading to a relative MAE reduction of 21.03% for the 'failed' cases. For clinical prospective test, the clinical sensitivity, specificity, and accuracy reached 100% under the minimum pre-defined uncertainty threshold at three<i>γ</i>criteria after manual intervention.<i>Significance.</i>This framework enhances AI-based PSQA reliability by systematically addressing prediction uncertainty, reducing manual workload, and improving GPR accuracy for 'failed' fields, thereby facilitating safe clinical adoption of automated PSQA systems.