Title : Post-hysterectomy pelvic abscess mimic: An AI-assisted diagnostic stewardship workflow
Abstract:
Background: Postoperative pelvic fluid collections after hysterectomy commonly trigger empiric broad-spectrum antimicrobial therapy, yet sterile postoperative collections and topical hemostatic agents may mimic abscess on imaging. We used a solved postoperative gynecologic oncology case as a worked example within a mentored, auditable workflow designed to transform de-identified clinical cases into literature-informed educational tools for infectious diseases reasoning.
Methods: A 43-year-old woman with recurrent diabetic ketoacidosis (DKA), abnormal uterine bleeding, thickened endometrium, endometrioid carcinoma, total hysterectomy/bilateral salpingo-oophorectomy, and readmission for a hysterectomy-bed air/fluid collection was reconstructed into a time-stamped dataset of serial vitals, laboratory values, microbiology, imaging, interventions, and outcome. A human-directed literature review preceded case mapping to reduce anchoring bias. An ask-verify-revise process was then used to compare the case against evidence on postoperative pelvic abscess, sterile postoperative collections, hemostatic-agent imaging mimics, biomarkers, and competing oncologic explanations. The large language model functioned only as a synthesis aid; all retained variables, exclusions, and interpretations were checked against primary sources and documented in a prompt ledger linked to a variable dictionary.
Results: The case trajectory included recurrent DKA, thrombocytosis to 831 × 10^3/uL, leukocytosis to 33 × 10^3/uL, low procalcitonin (<0.25 ng/mL), negative cultures, innumerable pulmonary nodules on readmission imaging, subsequent peritoneal carcinomatosis, ICU metabolic decompensation, and transition to hospice. Operative review identified intraoperative Arista AH absorbable hemostatic powder as the most plausible explanation for the postoperative pelvic collection, redirecting interpretation from presumed pelvic abscess toward a sterile postoperative mimic in the setting of aggressive malignancy. The workflow yielded four auditable outputs: a structured case summary, time-stamped dataset, prompt ledger, and variable dictionary. Educationally, the case made visible three teachable ID tasks: resisting premature anchoring to imaging alone, reassessing empiric antimicrobial therapy when biomarkers and microbiology conflict with the initial diagnosis, and documenting how a diagnostic pivot is justified. A Table of Provisional variable dictionary for trajectory-based risk conceptualization in endometrioid carcinoma with a postoperative pelvic abscess mimic is provided below.
Conclusions: This project does not validate a bedside prediction model. Instead, it shows how a single solved case can be converted into an auditable, literature-grounded educational scaffold linking diagnostic stewardship, antimicrobial stewardship, and responsible AI use. The next step is prospective replication across additional ID and non-ID masquerader cases to test transferability, refine variables, and evaluate educational impact.
