# What is perioperative AI? > Perioperative AI is the use of artificial intelligence to support the care of surgical patients before, during and after an operation. Most of it today sits before surgery, where the work is gathering a patient's history, identifying risk and getting them ready. It isn't a robot anaesthetist; it's closer to infrastructure that gets the right information to the right clinician in time. *Source: https://getpathways.ai/articles/what-is-perioperative-ai* Perioperative AI is software that uses artificial intelligence to gather, interpret and act on patient information across the surgical journey, from the decision to operate to recovery at home. Here's what it covers, how it works, and what to ask before trusting it. ## The short answer Perioperative AI is the use of artificial intelligence to support the care of surgical patients before, during and after an operation. Most of it today sits before surgery, where the work is gathering a patient's history, identifying risk, and getting them ready. It isn't a robot anaesthetist. It's closer to infrastructure: a way of getting the right information, about the right patient, to the right clinician in time to act on it. ## Why the perioperative period? "Perioperative" covers three phases. Preoperative runs from the decision to operate until anaesthesia begins. Intraoperative is the time in theatre. Postoperative covers recovery, the ward and the return home. The preoperative phase is where AI has the most to offer, because its core problem is information logistics. A patient's story is scattered across a GP summary, specialist letters, a medication list, old discharge summaries and the patient's own memory. Pre-admission teams spend much of their time collecting and reconciling it before anyone can assess risk. The work is also high-volume and time-bound. Thousands of patients pass through a busy hospital's pathway each year, each with a fixed surgery date. A recent audit at Royal North Shore Hospital found that 30% of patients who missed their digital pre-op questionnaire were invited two days or less before surgery, on the day, or afterwards. The bottleneck was getting the right information at the right time. ## Where perioperative AI shows up | Phase | Task | What AI does | Example | | --- | --- | --- | --- | | Before surgery | Health history | Runs the questionnaire as a conversation, by text or voice | A voice agent phones the patient and works through their history | | Before surgery | Documents | Extracts key facts from referrals, letters and results | Pulling cardiac function and implanted devices from a cardiology letter | | Before surgery | Medications | Reconciles the medication list and infers likely conditions | Anticoagulants, SGLT2 inhibitors and GLP-1 receptor agonists flagged for perioperative review | | Before surgery | Risk and triage | Applies screening tools and rules to route patients | A positive sleep apnoea or frailty screen sends a patient to an anaesthetist | | Before surgery | Preparation | Reminds patients of instructions and checks they've followed them | Day-prior calls that flag a fasting breach or a medication not stopped | | During surgery | Monitoring | Predicts events from physiological data | Early warning of intraoperative hypotension | | After surgery | Deterioration | Watches observations and results for worrying trends | Flagging a ward patient whose observations are heading the wrong way | | After surgery | Follow-up | Checks in after discharge and collects outcome measures | A post-discharge call about pain, wound healing and nausea | ## Not all "AI" is the same Three quite different technologies get called AI in perioperative care. **Language models** understand and produce text and speech. They are good at conversation, summarising and pulling facts out of unstructured documents. Their weakness is that they can misread or invent details, so their outputs need checking and a clear link back to the source. **Predictive models** learn from historical data to estimate the probability of something, such as a complication or deterioration. Their performance depends on the data they were trained on, so they need validating on local patients and monitoring over time. **Rules and clinical logic** encode guidelines and screening tools directly. Strictly speaking this isn't AI, but it is often the backbone of a safe system: transparent, auditable and predictable. The strongest perioperative systems combine all three. Language models collect and structure information, rules make decisions explainable, and clinicians make the judgement calls. ## What perioperative AI isn't - **It isn't a replacement for the anaesthetist.** The anaesthetist remains responsible for the assessment and the plan. - **It isn't a chatbot bolted onto a web form.** A conversational front end on top of a 150-question form is still a 150-question form. - **It isn't a fix for a broken workflow.** If invitations go out at the wrong time, or to patients who have been admitted, transferred or have died, a smarter questionnaire won't help. Integration with the medical record matters as much as intelligence. ## The risks, and how good systems manage them **Errors that travel.** A detail misread early can flow through the record into the anaesthetic plan. Good systems link every extracted fact to its source, weight sources by reliability (a specialist letter outranks patient recall), and put a clinician between the AI's output and any clinical decision. **Inequity.** A pathway designed for confident, English-speaking smartphone users leaves others behind. Voice channels, multilingual support, carer involvement and a human fallback help close that gap. **Stale data.** AI acting on an out-of-date list will contact the wrong patients at the wrong time. It needs live events from the medical record: admissions, transfers, cancellations and deaths. **Unproven performance.** Systems should be tested against clinician judgement on local patients before go-live, then monitored for as long as they run. ## Regulation in Australia - **TGA.** Software intended to diagnose, monitor, predict or guide treatment can be regulated as a medical device. Some clinical decision support software is exempt, broadly where it supports a health professional's decision rather than replacing it. The TGA continues to review how AI is regulated. - **State frameworks.** In NSW, AI projects in government agencies, including NSW Health, are assessed under the NSW AI Assurance Framework. Other jurisdictions run their own processes. - **Privacy.** Health information is covered by the Privacy Act 1988 and state health records laws, such as the NSW Health Records and Information Privacy Act 2002. ## Six questions to ask before adopting perioperative AI 1. What exactly does the AI do, and what does a clinician still do? 2. Can every output be traced back to its source document or conversation? 3. How was it evaluated, on which patients, and against what standard? 4. How does it connect to our medical record, and what happens when the record changes? 5. Who does it work less well for, and what's the fallback for them? 6. What is its regulatory status, and where is our patients' data stored? ## The bottom line Perioperative AI is most useful where care is held back by the effort of collecting and making sense of information. Done well, it gives clinicians a fuller picture earlier, reaches patients a web form can't, and frees skilled staff for the complex cases. Done badly, it automates the old problems faster. The difference lies in integration, evidence and keeping clinicians in charge.