The Challenge
Where Protocol Ends and Judgment Begins
Cardiothoracic surgery is a domain of acute stakes and compressed time. For the most common procedure types, roughly 80% of intraoperative decisions are protocol-driven: well-documented, teachable, and increasingly codified in clinical guidelines. That 80% is where existing clinical AI performs well.
The remaining 20% is different. These are the decisions that arise from unexpected hemodynamic instability mid-procedure, from anatomical variants that don't match the pre-operative imaging, from the moment when a surgeon knows — from experience, not from any published protocol — that the standard next step is the wrong one. These are the decisions that separate good outcomes from preventable complications, and they are precisely the decisions that cannot be learned from textbooks or extracted from clinical databases.
Junior practitioners don't have access to that 20%. They are trained on the documented case — the expected presentation, the textbook anatomy, the by-the-book haemodynamic response. The unexpected case is where they are most vulnerable, and it is where the guidance they need is least available. An experienced attending surgeon cannot be present for every procedure. And generic clinical AI, built on literature and aggregated case records, handles the 80% competently but fails precisely where it matters most.
The question Praxa set out to answer: could the specific judgment of a highly experienced surgeon — the tacit knowledge built across a 24-year career — be encoded into a system that could surface that judgment at the moment a junior practitioner needs it most?
The Approach
From Tacit to Explicit: The Knowledge Encoding Process
Finding the Right Expert
Dr Amara Osei was identified through a structured expert identification process — not as the most credentialled surgeon available, but as a practitioner who combined deep procedural experience with an unusual capacity for reflective articulation. The ability to perform at a high level and the ability to explain the reasoning behind that performance are distinct skills. Both were required. Dr Osei, a cardiothoracic surgeon with 24 years of practice across three major teaching hospitals, had spent years informally mentoring junior colleagues — which meant he had already developed a working vocabulary for the judgments that most experienced surgeons carry implicitly.
Structured Knowledge Sessions
Encoding began with a series of structured knowledge-elicitation sessions — not interviews, but collaborative mapping exercises. The goal was not to document what Dr Osei knew about cardiothoracic surgery in general, but to map the specific decision logic he applied in specific intraoperative scenarios. Sessions were structured around case types: coronary artery bypass grafting, valve replacement procedures, aortic root surgery, and three additional procedure categories where Dr Osei had identified the highest concentration of judgment-dependent moments.
Each session focused on a narrow decision domain: what conditions trigger a re-evaluation, what anatomical or haemodynamic signals change the operative plan, what the "when not to act" logic looks like when the standard response would be the wrong one. These are the judgments that do not appear in surgical textbooks because they are context-specific — they are the product of having seen a particular failure mode across enough cases to recognise its early signature.
Encoding Methodology
Tacit surgical knowledge resists direct transcription. The standard elicitation problem — asking an expert to explain what they know — produces the sanitised account: the retrospectively coherent narrative, the decision logic that sounds good in hindsight but doesn't match the actual cognitive process in the operating theatre. Praxa's encoding methodology addresses this through case-anchored reconstruction: rather than asking Dr Osei to describe his decision process abstractly, sessions were structured around specific cases from his practice, working forwards through the decision logic as it actually unfolded.
This produced something qualitatively different: 47 discrete decision nodes, each linked to a specific class of intraoperative conditions, each with explicit reasoning chains and documented edge cases. Crucially, it also produced the negative cases — the scenarios where the intuitive response is wrong, and where the experienced surgeon's judgment consists precisely in not acting when less experienced practitioners would.
Validation
The encoded knowledge base was validated through two stages. In the first stage, Dr Osei reviewed every decision node against his case records — identifying errors, adding specificity, and flagging the scenarios where the encoded logic required additional nuance. In the second stage, a structured review of 60 historical cases was conducted to test whether the decision-support logic would have produced appropriate recommendations, with Dr Osei's retrospective assessment as the benchmark. Cases where the encoded system diverged from his judgment were examined in detail, and the encoding was refined accordingly.
The Product
What Was Built, and What It Does Not Do
The resulting product is a real-time intraoperative decision-support tool — not an autonomous decision-maker, but a structured checklist that operates at the level of Dr Osei's experience. When specific haemodynamic conditions or anatomical signals are logged during a procedure, the system surfaces the relevant decision logic: the conditions that Dr Osei, across 1,400+ procedures, has identified as requiring a structured re-evaluation before proceeding.
The system is explicit about what it does not do. It does not replace the clinical judgment of the operating surgeon. It does not make recommendations that extend beyond the documented case base. It does not produce outputs that cannot be traced to a specific class of documented procedures from Dr Osei's practice. Every recommendation in the system links directly to a procedure category and a documented reasoning chain — not to a statistical model's opaque output, but to an identifiable piece of practitioner logic.
The format is designed for the operating theatre environment: non-disruptive, readable at a glance, integrated within the existing workflow rather than requiring a separate interaction. The system is auditable at every point — for clinical governance purposes, every recommendation can be traced back to its encoded source, reviewed, and if necessary, challenged. This auditability was a core design requirement from the outset, not a feature added at the end.
The primary intended user is not the experienced surgeon — Dr Osei does not need a system that reflects his own judgment back at him. The primary user is the junior practitioner encountering an unexpected intraoperative scenario, who now has access to the structured reasoning of a 24-year veteran at precisely the moment they need it.
24
Years of surgical experience systematically encoded for the first time
47
Distinct decision nodes mapped across 6 procedure types
1,400+
Procedures underpinning the encoded case base — each recommendation auditable to source
"The value isn't in what the AI knows. It's in what I know that no one else has ever written down — and now it's operational."— Dr Amara Osei, Cardiothoracic Surgeon
Build with domain expertise at the core.
Praxa works with practitioners who have knowledge that has never been systematically encoded. If that describes you, let's talk.
Work with Praxa