Physician-as-architect: clinician-led LLM-assisted development and deterministic verification of an ESC-aligned peri-operative decision support system
M T A Egaimi, M H A M Serour, L M E Ibrahim, A M E Elnour, S S B Omer, M L M Al-Natour, A Katbeh, H Seo, A Almorraweh, W Choi, F Clementi, P CorvoAbstract
Introduction
Translating guideline flowcharts into bedside clinical decision support is frequently delayed by dependence on specialist software engineering, sustaining the evidence–practice gap. Large language model (LLM) coding assistants may enable clinicians to act as software architects; however, bedside deployment demands deterministic behaviour and verifiable pathway completeness.
Purpose
To demonstrate the feasibility of clinician-led, LLM-assisted development of a deterministic clinical decision support system (CDSS) implementing the ESC 2022 pre-operative cardiovascular assessment workflow for non-cardiac surgery (NCS), and to quantify algorithmic completeness by exhaustive state-space enumeration.
Methods
A cardiology team used an LLM coding assistant to implement the ESC 2022 NCS workflow as a standalone, client-side HTML/JavaScript hard-coded rules engine with no runtime AI. The CDSS encodes 12 conditionally presented decision nodes — surgical urgency, surgical risk category, symptoms/signs, established CVD, patient age, cardiovascular risk factors, family history of genetic cardiomyopathy, functional capacity, frailty, cardiac murmur, prior PCI/CABG, and ischaemia concern — with each node presented only when logically triggered by earlier answers. Completeness testing enumerated all tool-reachable input states (n = 3,313), requiring a structured, non-null recommendation for each. The large state count, despite only 12 nodes, reflects the multiplicative effect of conditional branching: the theoretical Cartesian product of all node values exceeds 31,000, but the conditional logic constrains reachable paths to 3,313. Outputs were then clustered into unique actionable recommendation bundles.
Results
The CDSS generated deterministic recommendations for all 3,313 reachable input states (100% coverage), collapsing to 67 unique actionable recommendation bundles. This reachable state space arises from the conditional branching structure of the 12 decision nodes, where later nodes are presented only when logically triggered by earlier answers, constraining enumerable paths well below the theoretical Cartesian product of all node values. All outputs were non-null, guideline-concordant, and reproducible, confirming complete pathway coverage with no dead-end or undefined states.
Conclusion
Physician-as-architect development, accelerated by LLM coding assistance and safeguarded by exhaustive state-space verification, can operationalise ESC guideline workflows into transparent, deterministic, institution-independent CDSS. Prospective evaluation should address usability, clinical governance, version control, and impact on peri-operative testing rates, appropriate escalation, and patient outcomes.Central Illustartion