Observable Classification Patterns and Diagnostic Uncertainty in Parotid Ultrasound: A Multimethod Secondary Analysis.

Background/Objectives: Parotid ultrasound requires integration of sonographic morphology with clinical context under uncertainty, but conventional agreement and performance metrics provide limited insight into examiner-dependent classification patterns. Building on previously derived examiner-specific surrogate models, we applied a behavioral decision-science framework to examine observable classification patterns and case-level ambiguity. Methods: In this exploratory retrospective single-center secondary analysis, six examiners rated predefined ultrasound image sets from 149 histopathologically verified lesions. Surrogate-tree structure, feature-reference-class associations, signal detection metrics, and case-level ambiguity were analyzed. Internal robustness was assessed using cross-validation and sensitivity analyses. Results: Pruned surrogate trees identified a context-based classification pattern, a morphology-based classification pattern, and a hierarchical morphology-context pattern. Tumor history and lesion boundary were the only retained upper-level classification cues and showed the strongest overall categorical and sonographic associations with the histopathological reference class, respectively. Across 30 cross-validation trees, boundary was the root split in 22 and tumor history in eight; no other root feature occurred. Examiners differed in diagnostic discrimination and estimated decision criterion, with Examiner 3 showing the highest discrimination. Boundary remained the sonographic feature most strongly associated with the histopathological reference class across sensitivity analyses. Descriptor disagreement was associated with decision disagreement and error burden, and four cases showed unanimous but incorrect classifications. Conclusions: In this exploratory retrospective single-center cohort of histopathologically verified lesions, examiner-dependent classification was characterized by observable, model-derived cue structure, feature-use/reference-association alignment, estimated decision-criterion patterns, and case-level ambiguity. These cohort-specific findings do not directly measure examiner cognition and should not be generalized to unselected parotid-ultrasound populations; however, the complementary analyses provide an empirically grounded framework for characterizing observable examiner-dependent classification and identifying testable targets for prospective studies of examiner feedback, descriptor standardization, and decision support.