Using decision aids and AI at the bedside
- Topic
A score or an AI suggestion is one more clue, not the answer. Used inside your own reasoning it catches misses; obeyed blindly it adds errors of its own.
What it is
This page is about the moment of use: what to do with a prediction rule's score, an automated risk alert or a chatbot answer while you reason about one patient. How these systems are built, validated and governed is covered in Clinical decision support and Artificial intelligence in medicine: uses and limits. At the bedside the output is evidence for your diagnostic reasoning, never a replacement for it.
Step by step
- Set your own pre-test probability first, from the history and exam, before you look at any score or suggestion.
- Check fit: was the tool validated in patients like this one (setting, age, how common the disease is)? If not, give its output little weight.
- Read the output as a test result: a high or low result moves your probability up or down; it does not replace your starting estimate.
- Ask what the tool could not see: a finding, a history or a result it had no input for.
- If it disagrees with you, find the finding that explains the gap and re-examine the patient; do not accept or dismiss it by reflex.
- Override it when the patient contradicts it: a sick patient outweighs a reassuring score, and a well patient outweighs an alarming one.
- For a chatbot answer, confirm each fact and reference in a guideline or primary source before acting, and never enter identifiable data.
- Record your reasoning and, when you overrode the tool, why.
What each output is worth at the bedside, and when to overrule it.
| Output | Treat it as | Check before acting | Overrule it when |
|---|---|---|---|
| Prediction rule score you compute (for example a Wells score) | A structured estimate of pre-test probability | Your patient matches the population it was validated in; every item was scored honestly | A finding the rule does not include points elsewhere, or the patient is outside its population |
| Automated alert or risk score (for example a sepsis alert) | A prompt to look again, with many false alarms and misses | What data fed it, and whether the patient really looks like the alert says | The clinical picture disagrees: a low score never clears a sick patient |
| No alert, or a low score | Absence of a warning, not evidence of health | That the tool had the data it needed, and your own assessment | Always, if the patient looks unwell: tools miss cases |
| Chatbot answer | An unchecked draft that may contain invented facts or references | Each fact and reference in a guideline, drug label or primary source | It cannot be traced to a source, or it conflicts with one |
Clinical notes
- Automation bias is over-reliance on the tool. It causes two kinds of error: omission (missing a problem because the tool stayed silent) and commission (following advice that is wrong).
- Automation bias grows with workload, time pressure and task complexity; training, stressing that the user stays accountable, and tools that give information rather than orders reduce it.
- A silent tool is not reassurance: a sepsis model used in hundreds of US hospitals missed two thirds of sepsis cases and still alerted on 18 % of all inpatients.
- A score's numbers hold only in patients like those it was validated on; in a different setting or case mix its estimate of probability can be badly off.
- Think first, then look: forming your impression before you see the output lets the tool check you instead of leading you.
- Skills fade with reliance: after AI-assisted colonoscopy was introduced, endoscopists' adenoma detection without AI fell from 28.4 % to 22.4 % in one multicentre study.
- Language models can write fluent, confident answers that are false, including references that do not exist; trace every claim you will act on to a source.
- Public chatbots are not a safe place for patient data: remove identifiers or use only tools your institution has approved, and keep the decision under human control.
Common traps
- A low score or no alert means the patient is fine: tools miss cases (errors of omission), so a sick patient outweighs a quiet screen.
- The tool's output replaces my clinical estimate: it is one more piece of evidence that moves your pre-test probability, which still starts from the patient.
- A validated score works in any patient: it holds only in patients like those it was validated on; outside them, give it little weight.
- A fluent chatbot answer with references has been checked: language models can invent facts and references; open the source before acting.
- Following the tool protects me: the clinician who acts on it decides and is accountable, whatever the tool suggested.
Test yourself
Think of the answer, then tap the question to see it.
A patient's automated sepsis score is low, but she is hypotensive, febrile and newly confused. What do you do?
Act on the patient, not the score: these tools miss many cases, so a low score does not exclude sepsis. Assess and treat as the clinical picture demands, and record why you overrode the tool.
A resident signs off an ECG on the machine's reading without looking at the tracing and misses a finding it did not flag. What error is this, and how is it prevented?
Automation bias, an error of omission: the silent tool was trusted. Read the tracing yourself first, then compare with the machine; training and clear personal accountability reduce it.
Your pre-test probability of pulmonary embolism is low, but a validated decision aid returns a high-risk result. How should it change your thinking?
Treat it as evidence: it raises your probability by as much as its accuracy deserves, if the patient fits its population. Look for the finding it saw and you missed, then choose testing on the new estimate.
A chatbot gives you a tidy differential and cites a guideline. How do you use it safely?
As a draft: test each diagnosis against the findings, open the guideline to confirm it exists and says that, and never enter identifiable patient data in a public tool. The decision stays yours.
Why form your own impression before you look at an AI suggestion?
So the tool checks you instead of leading you: seeing its answer first anchors you to it and feeds automation bias. It also keeps your skills, which fade when you rely on the tool.
Elsewhere in DurgiPedia
- Internal and external validity
- Spectrum bias
- Likelihood ratios
- Pre-test and post-test probability
- Evidence-based medicine
- Patient safety and adverse events
- The electronic health record: Ley 2015 de 2020
Connections
Educational content on the clinical exam; it is learned and practised under supervision. It does not replace assessment by a health professional and is not for self-diagnosis.
Sources: Framework for the impact analysis and implementation of Clinical Prediction Rules (Wallace and colleagues, BMC Medical Informatics and Decision Making 2011) (Europe PMC, 2011); External validation of a widely implemented proprietary sepsis prediction model in hospitalized patients (Wong and colleagues, JAMA Internal Medicine 2021) (Europe PMC, 2021); Automation bias: a systematic review of frequency, effect mediators, and mitigators (Goddard and colleagues, JAMIA 2012) (Europe PMC, 2012); Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study (Budzyń and colleagues, Lancet Gastroenterology & Hepatology 2025), abstract (Europe PMC, 2025); The future landscape of large language models in medicine (Clusmann and colleagues, Communications Medicine 2023) (Europe PMC, 2023); WHO issues first global report on Artificial Intelligence (AI) in health and six guiding principles for its design and use (World Health Organization, 2021).
Spotted a mistake?We are always improving for the community. Tell us what is wrong and our team will review it.Report a mistake
Representative model, for educational purposes only.
