Artificial intelligence in healthcare: what it does today

Few areas of healthcare attract more exaggeration.

The reality is narrower and more useful than the marketing suggests.

Where it is genuinely used

• Flagging possible abnormalities on images for the radiologist to review.
• Prioritising urgent studies in a reporting queue.
• Detecting patterns in monitoring data.
• Reducing administrative work such as documentation.
• Supporting laboratory and pathology workflows.
• In all of these a clinician remains responsible.

What it is not doing

• Making diagnoses independently.
• Replacing clinical examination.
• Deciding treatment without human review.
• Understanding your circumstances and preferences.
• These require judgement, not pattern matching.

Why narrow tasks work best

• Systems are trained on specific data for specific questions.
• Performance drops outside those conditions.
• Populations differ between hospitals and countries.
• Equipment and protocols differ too.
• A tool validated elsewhere may perform differently locally.

The evidence question

• Ask whether a tool has been evaluated in real clinical use.
• Ask whether it was tested on patients like you.
• Ask what outcome improved.
• Accuracy on a dataset is not the same as benefit to patients.
• Ask who funded the evaluation.

Bias

• Systems reflect the data they were trained on.
• Under-represented groups may be served less well.
• This is a recognised and active problem.
• It is reasonable to ask how it has been addressed.
• Human review is part of the mitigation.

Symptom checkers

• Useful for general orientation.
• They do not know your history.
• They cannot examine you.
• They tend towards either alarm or false reassurance.
• Never use them to decide against seeking care for serious symptoms.
• Use them to prepare questions, not to reach conclusions.

General purpose chat tools

• They can explain terminology in plain language.
• They can help you prepare questions.
• They can produce confident errors.
• They do not have your records.
• Verify anything important with your clinician.
• Never adjust treatment based on such output.

Data and privacy

• Ask whether your data is used to develop these tools.
• Ask whether it is identifiable.
• Ask whether it leaves the institution or the country.
• Ask whether you can decline without affecting your care.
• Practice varies widely between institutions.

Questions you may reasonably ask

• Was an algorithm involved in my assessment?
• Who reviewed its output?
• Who is responsible for the final decision?
• What happens if the tool and the clinician disagree?
• These are legitimate questions, not obstruction.

Where it may help you directly

• Shorter waits for imaging reports in some services.
• Earlier flagging of urgent findings.
• More clinician time through reduced paperwork.
• Better monitoring in some chronic conditions.
• The benefits are practical rather than dramatic.

Where scepticism is warranted

• Products claiming to detect many diseases at once.
• Consumer tools promising diagnosis.
• Anything sold directly to patients bypassing clinicians.
• Claims without published evaluation.
• Ask for the evidence and read who produced it.

Administrative uses you may notice

• Automatic transcription of consultations is increasingly common.
• Ask whether the conversation is being recorded.
• Ask what happens to the recording afterwards.
• Ask whether you may decline.
• Ask whether the summary is checked before filing.
• Errors in an automated summary can persist in your record for years.

Checking your own record

• Ask to see what was written after an appointment.
• Correct factual errors promptly.
• Check that your medicines and allergies are accurate.
• Check that resolved conditions are not listed as active.
• This matters more as automation spreads.

Triage systems

• Some services use algorithms to sort appointment requests.
• Describe symptoms plainly and completely.
• Do not minimise to seem reasonable.
• Say clearly if you are worried or deteriorating.
• Ask to speak to a person if the outcome seems wrong.

Regulation

• Medical software is regulated in many jurisdictions.
• Wellness products often are not.
• The distinction determines what claims may be made.
• Ask whether a tool is a regulated medical device.
• Unregulated does not mean useless, but claims deserve more scrutiny.

Consumer devices that claim clinical detection

• Some wearables claim to detect rhythm problems or other conditions.
• A detection is a prompt to seek assessment, not a diagnosis.
• False positives are common in low-risk populations.
• Take the recording to a clinician rather than acting on it.
• Ask whether the feature is cleared as a medical device where you live.

Keeping perspective

• The most valuable improvements are often unglamorous.
• Reliable records and good communication matter more.
• Continuity with a clinician who knows you matters more.
• Technology supports those things; it does not replace them.

Who is accountable when something goes wrong

• Responsibility rests with clinicians and institutions, not software.
• Ask what oversight exists for tools in use.
• Ask how errors are reported and reviewed.
• Ask whether performance is monitored locally over time.
• A tool that is never re-evaluated after purchase is a risk.

How to talk to a clinician about it

• Ask neutrally rather than accusingly.
• Say what you read and where.
• Ask whether it applies to your case.
• Most clinicians welcome an informed question.
• Few welcome a printout presented as a diagnosis.
• The difference in framing changes the whole conversation.

What to do with an alarming output

• Do not act on it alone.
• Write down what it said.
• Take it to your clinician as a question.
• Ask whether it applies to your situation.
• Many such outputs are simply wrong.

The point to remember

Three things:

1. A clinician remains responsible for every decision.
2. Symptom checkers and chat tools do not have your records — verify everything.
3. Ask how your data is used and whether you can decline.

Câu hỏi thường gặp