The AI Learning Hub Journal

Decision Support, Triage, and Patient-Facing Tools

How directly does the tool touch the clinical decision?the further down this list a tool sits, the more oversight it needs and the higher the barTIERWHAT THE TOOL DOESWHO DECIDESOVERSIGHT / EVIDENCEInformation and navigationfurthest from the decisionSurfaces material and answers general questions, sosomeone can find their way to the right service.The reader, with noclinical claim attached.OVERSIGHTEVIDENCEDocumentation supportclose by, but one step backDrafts, summarises or structures what a clinician hasalready said or written down.The clinician, who signswhat the record says.OVERSIGHTEVIDENCEAlerting and flagginginterrupts the workflowDraws attention to something in the data and asks forit to be looked at again.The clinician — but theagenda has been set.OVERSIGHTEVIDENCERecommendationproposes a specific courseNames a preferred option, so declining it becomes anactive choice somebody has to make.Still the clinician,now against a default.OVERSIGHTEVIDENCEAutonomous actionacts without a person in the loopTakes the action directly, with review after the factif it happens at all.The system, untilsomebody intervenes.OVERSIGHTEVIDENCEThe bar rises with proximity to the decision — one tier down is a different evidence caseWHAT QUIETLY MOVES A TOOL DOWN THE SPECTRUMFraming does not set the tierA tool described as informationalbut read as an instruction isoperating a tier further down.Defaults are decisionsWhen declining a suggestion takeseffort, the suggestion is nearerthe decision than it looks.Oversight has to be realSomeone who cannot realisticallyreview every output is not theoversight the tier assumes.Where a tool sits is set by what it does to the decision, not by how it is describedTwo tools with identical outputs sit on different tiers if one of them is harder to ignore
Educational orientation only — no tool, tier or threshold here is a recommendation for practice

Decision Support: An Old Field, A New Engine

Clinical decision support long predates modern AI — rule-based alerts and risk scores have been embedded in record systems for decades, with a well-documented failure mode: alert fatigue. Clinicians override alerts that fire too often or too imprecisely, and overriding becomes reflexive. Machine learning changes how the prediction is produced, not the human factors problem it lands in. A statistically better model that still fires constantly, or that fires without an actionable next step, will be dismissed exactly like its predecessors. The literature on deterioration and sepsis prediction in particular includes prominent examples of systems that were widely deployed and then found, on independent evaluation, to perform substantially worse in practice than their original claims implied.

  • Alert fatigue is the dominant failure mode and it is a human factors problem, not a modelling one
  • Widely deployed prediction systems have been found, on independent evaluation, to underperform their original claims
  • A prediction without a defined, feasible action attached is an interruption, not decision support

Prediction Is Not the Same as Benefit

This is the single most misunderstood point in clinical AI. A model that identifies at-risk patients accurately still has to show that identifying them earlier leads to a different action, that the action is available and taken, and that taking it improves outcomes. Each link can break. The action may already have been happening. The staffing to respond may not exist. The intervention itself may lack strong evidence. And there is a structural irony: a successful intervention changes the outcome the model was trained to predict, so the model appears to degrade precisely when it is working. Asking "what changed for patients?" rather than "how accurate is it?" separates useful decision support from expensive noise.

  • Accurate prediction, available action, action actually taken, action that helps — every link must hold
  • Successful intervention degrades apparent model performance, which complicates monitoring
  • The question is what changed for patients, not how accurate the score was

Operational Triage and Patient-Facing Tools

Two zones with opposite risk profiles. Operational triage — reordering a worklist, flagging studies for earlier review, forecasting demand — changes the sequence of work rather than clinical conclusions, which makes it comparatively tractable, though a prioritisation system that systematically deprioritises a group is a fairness problem even when every case is eventually seen. Patient-facing tools are the opposite: symptom guidance, chat interfaces, and messaging assistants reach people with no clinician between the output and the action. Design has to assume the user cannot evaluate the answer, may be in distress, and may not recognise an emergency. Escalation paths and hard boundaries matter more than conversational quality.

  • Triage reorders work rather than changing conclusions — but systematic deprioritisation of a group is still a harm
  • Patient-facing tools have no clinician between the output and the action
  • Escalation to a human and explicit refusal boundaries matter more than fluency
  • Assume users cannot judge correctness and may not recognise an emergency

Prefer slides, quizzes, and saved progress? Read this lesson in the library — free, no sign-up.