Artificial intelligence and clinical decision support in paediatrics — viva

Opening (must-hit)

"Artificial intelligence in paediatrics is a computer system that performs a task a clinician would otherwise do by judgement. The four classes are prediction and early warning (the deterioration, sepsis and PICU-transfer score), diagnosis and image interpretation (the radiograph, retinopathy-of-prematurity and skull-fracture reader), rule-based decision support (the drug-interaction, dosing and allergy alert) and generative or triage tools. The candidate must distinguish a trained prediction model from a rule-based alert, because the two are validated and governed differently. The validation ladder is internal cross-validation, external validation in a different population, and a prospective silent trial — external validation is the gate. Sensitivity and specificity are not fixed but shift with prevalence and threshold, and the clinician remains accountable for every output. The four safeguards are clinician over-read, a written escalation plan, alert stewardship and ongoing monitoring." [2] [10]

Branch A — The deterioration score that reads low in a sick child

Examiner: It is two in the morning. The machine-learning deterioration score for a six-year-old reads low, but the child has increased work of breathing, cool peripheries and is less responsive than an hour ago. What do you do? [2]

Candidate: I read the child before the model. The bedside assessment overrides the score, so I escalate to the higher acuity now — I call my senior and the rapid-response team, and I begin the structured ABCDE assessment and the supportive care the child already needs. The most dangerous AI failure in paediatrics is the reassuring number in a sick child, and automation bias has caused missed deterioration when a normal-looking number silenced a worried clinician. I document the discrepancy and the over-ride so the event is auditable and the tool's performance is reviewed. After the child is safe, I would ask whether the model was externally validated in a population like ours, whether the inputs were noisy, and whether the threshold needs retuning — but none of that delays the escalation. [2]

Branch B — The drug-interaction alert the team dismisses

Examiner: Your team overrides more than nine in ten of the drug-interaction alerts they receive. Is that a problem, and what do you do about it? [7]

Candidate: Yes — that is alert fatigue, and it is a patient-safety failure, not an inconvenience. An alert overridden more than nine times in ten has lost its value, and the unmanageable burden drives an override culture in which the one actionable alert is lost in the noise. The stewardship response is to measure the override rate, retire or convert the lowest-value interruptive alerts to a passive display — the Fallon study replaced a burdensome interruptive alert with passive clinical decision support — and tune the thresholds with the prescribers, as the Simpao paediatric drug-interaction study did with a visual analytics dashboard. The Chaparro framework sets the best-practice cycle of monitoring and improving interruptive alerts, and I would treat the rising override rate as a signal to convene the stewardship team rather than endure it. [7] [8] [9]

Branch C — The retinal image flagged in the nursery

Examiner: A deep-learning tool flags a retinal image of a premature baby for retinopathy of prematurity. There is no ophthalmologist on site tonight. Walk me through the evidence and your action. [5] [6]

Candidate: The deep-learning evidence for retinopathy of prematurity is anchored by the Taylor study, which built a quantitative severity scale using deep learning, and the Young study, which tested smartphone-based telescreening with and without artificial intelligence in India, where the shortage of ophthalmologists makes AI-assisted telescreening an equity tool as much as a diagnostic one. My action is to treat the tool's flag as a prompt to escalate, not a diagnosis. I assess the baby's gestational and postnatal risk factors, arrange the formal ophthalmology review — by telehealth tonight if the ophthalmologist is off site — and ensure the image and the algorithm's output are over-read by the clinician responsible for the baby. The tool extends screening into a setting where the ophthalmologist is the limiting factor; it does not remove the over-read obligation, and the treating team remains accountable for the follow-up. [5] [6]

Branch D — Deploying a sepsis-prediction model

Examiner: Your hospital wants to deploy a machine-learning sepsis-prediction model. What evidence do you require, and what are the safeguards? [1] [2]

Candidate: I require the model to have climbed the validation ladder — internal cross-validation is not enough, so I need external validation in a population like ours, and ideally a prospective silent trial that proves it works in live practice before it drives an action. The Le machine-learning sepsis-prediction study is a canonical example of a trained model in a high-stakes, low-prevalence setting, and the Mayampurath PICU-transfer study is the one to cite for the external-validation gate. I would ask whether the model was calibrated for our prevalence, because sensitivity and specificity shift with prevalence and threshold, and a tool tuned in a tertiary unit may flood a district ward with false alarms. The four safeguards are clinician over-read of every output, a written escalation plan with explicit thresholds and a twenty-four-hour contact, alert stewardship that measures the override rate, and ongoing monitoring of performance drift, equity and alarm burden. Deployed together they hold the clinician accountable for every output. [1] [2]

Branch E — The generative output that is fluent but wrong

Examiner: A large-language-model scribe drafts a discharge summary that is fluent and confident but states the wrong diagnosis. Who is responsible, and what is the obligation? [10]

Candidate: The clinician who signs the summary is responsible, because the accountability does not transfer to the tool on deployment. The generative output is fluent and may be wrong — the failure modes are hallucination, outdated training data, misinterpretation of the prompt, and a confidently stated error — and the verification obligation is to check the draft against the source before it enters the record. This is the augmented-intelligence framing: the tool assists, the clinician decides, and the decision is auditable. I would correct the error, verify the rest of the summary, and sign only what I have confirmed, and I would feed the error back to the stewardship or governance team so the deployment is reviewed. [10]

Branch F — The regulatory and equity picture

Examiner: Name the regulators and the principle that frames AI as an aid rather than a replacement, and tell me the paediatric-specific risk. [10]

Candidate: The regulators are the Therapeutic Goods Administration in Australia, the Medicines and Healthcare products Regulatory Agency in the United Kingdom, the Food and Drug Administration in the United States, and Health Canada in Canada, and the World Health Organization ethics-and-governance guidance frames equity, transparency and accountability as the global floor. The unifying principle is augmented intelligence — AI assists rather than replaces the clinician, and the clinician remains accountable for every output. The paediatric-specific risk is that children are under-represented in the data that train most models, and a child changes physiologically across age bands, so a model calibrated to the older child will mis-score the neonate. That is why age-stratified external validation is required, and why the equity monitoring across subgroups — Indigenous, migrant, rural, disabled — is a deployment safeguard rather than an afterthought. [10]