Title : Building a clinical reasoning tool from post-transplant MRSA sepsis: A mentored ai workflow
Abstract:
Background: Solved clinical cases can be converted into worked examples—cases with a known diagnosis and outcome that are used to teach clinical reasoning—but building these from real records is time-intensive. We developed a mentored AI workflow to transform a de-identified case of post-transplant MRSA sepsis, initially treated as an upper respiratory illness, into a reusable clinical reasoning tool. In this study, diagnostic pivot refers to the point at which new data changed the leading explanation for the patient’s illness.
Methods: A medical student and faculty mentor applied an ask-verify-revise workflow to a de-identified case of a 62-year-old liver-transplant recipient with fever to 103°F, recent abdominal wall mesh repair, nasal congestion and dry cough despite doxycycline, negative blood cultures, CT-defined peri-mesh fluid collection, and MRSA recovered by image-guided aspiration. First, the model was asked case-agnostic questions across clinical history, postoperative risk, vital signs, laboratory data, imaging, microbiology, and source control. Second, candidate features were checked against primary literature and the source case record; weak or unsupported suggestions were discarded. Third, retained features were rewritten into plain-language definitions and mapped to the case timeline. This process generated a nine-variable dictionary: transplant immunosuppression, recent mesh surgery, antibiotic non-response with upper-respiratory-like presentation, fever burden, tachycardia, white-blood-cell dysfunction pattern, CRP elevation, creatinine/renal stress, and CT evidence of peri-mesh collection.
Results: The workflow made the diagnostic pivot explicit: the working diagnosis shifted from presumed respiratory infection to occult postoperative MRSA source infection after CT imaging and image-guided aspiration. It also produced a reusable teaching package consisting of a time-ordered case timeline, a variable dictionary, and an auditable record of prompts and mentor decisions. For educators, these materials show how a complex real case can be converted into a structured reasoning exercise. For learners, they support three high-yield discussions: infection may be masked in immunocompromised hosts, negative blood cultures do not exclude deep infection, and failure of outpatient antibiotics should trigger reassessment of source. Because the reasoning trail is visible, the workflow can be reviewed, adapted, and replicated by other educators. A table of provisional variable dictionary for trajectory-based risk conceptualization of Post-Transplant MRSA Sepsis Masquerading as Upper Respiratory Illness is provided.
Conclusions: A mentored AI workflow can convert a solved infectious diseases case into a clinical reasoning tool for teaching diagnostic revision and responsible AI use. Its main value is transparency and reproducibility, not prediction from a single case. This tool has not yet been evaluated with learners, so claims about educational effectiveness remain preliminary. Future work should test whether it improves reasoning, auditability, and case-based infectious diseases teaching.
