AI Medicine Lab
AI Medicine Lab
Plenty of medical websites will tell you “we have AI”. This page sets out to do the opposite: to lay the work open — where AI is used, how it was built, how accurate it is, and how far it has actually got.
Because in medicine, whether or not AI was used has never been the point. The point is: what specific problem does this tool solve? What is its error rate? What happens when it is wrong? Who is answerable for the result?
So every item below is labelled with the stage it is at, and work that is still in progress is never described as though it had already been validated. What is public at the moment is the tools that are already live and that anyone can actually use; topics still in progress, and work not yet published, are not listed here until there are results.
Three things to be clear about first
- What “AI” means here
- It covers deep learning on images (nodule detection on chest CT, radiomics), machine learning on structured data (sleep staging, prognostic stratification), and clinical tools that turn the text of a guideline into rule-based decision logic. That last kind is not machine learning, but it is still clinical knowledge turned into a program that can be run the same way every time, and we will not dress it up as AI.
- What we do not do
- We do not use AI to replace diagnosis, we do not handle any identifiable personal health data on this site, and we do not put unvalidated models anywhere the public can use them directly. Every tool on this site does its computing on the user’s own device; nothing is uploaded and nothing is kept.
- Who is answerable
- The content of every clinical tool is compiled and reviewed by Dr. Hwa-Yen Chiu, and labelled with its evidence strength and the version of the guideline it rests on. A tool is an aid, not a diagnosis; responsibility still lies with the decision the doctor and the patient make together.
Clinical tools Live
Tools that are already on the site and that anyone can use right now. These are rule-based clinical decision support — guideline text, reimbursement rules and dose calculations written out as logic that runs the same way every time, taking out the human error of looking things up and converting units. They are not machine learning models, and we do not claim that they are.
Smart report system (VHCT) （中文）
Turns the test reports seen most often in chest medicine into tools you can work through item by item: lung function (spirometry, bronchodilator testing, IOS, DLCO, body plethysmography), sleep (PSG, CPAP/APAP), questionnaires (CAT, ACT, ESS), FeNO and chest ultrasound.
Go to the tools → Clinical decision supportClinical pathway toolkit （中文）
Thirteen disease topics, with international guidelines and National Health Insurance (NHI) reimbursement rules written out as interactive logic: enter the patient’s details, and it returns suggested management, doses, a reimbursement check and a positioning chart. Covers asthma, COPD, ILD, hyperlipidaemia, heart failure, diabetes, eCKD, multidrug-resistant organisms, tuberculosis, LTBI, LDCT and hyperbaric oxygen.
Go to the tools → Symptom triageSymptom self-assessment and triage tools
Triage logic for six common symptoms: red-flag warning signs are caught first to settle how urgently you need to be seen, then the pattern is used to suggest a likely direction and list the tests that should be done, and finally you are pointed to health information, report interpretation or the outpatient clinic.
Go to the tools → Indication matchingSelf-assessment of suitability for non-drug treatments
A tool for checking indications and contraindications across five treatments: hyperbaric oxygen, intravenous laser, external counterpulsation, transcranial magnetic stimulation and acupuncture. The emphasis is on catching safety problems — stopping absolute contraindications, and labelling honestly how strong or weak the evidence is for each use.
Go to the tools →Four cautions about AI in medicine
- High accuracy does not mean useful. A model that is “99% accurate” on a rare disease may be so only because it always answers “no”. What you need to look at is sensitivity, specificity, and the positive predictive value at the prevalence you actually see.
- Accurate at hospital A does not mean accurate at hospital B. A different machine, different acquisition settings and a different population all make performance fall. If a model has no external validation, discount its numbers.
- What the AI can see and what the doctor cares about are not necessarily the same thing. The model optimises some mathematical measure; the doctor optimises the patient’s outcome and quality of life. Lining the two up takes a person.
- In the end it is still a person who is answerable. A tool can cut down omissions and arithmetic mistakes, but deciding whether to do a particular test, or whether to carry a particular risk, is a judgement the doctor and the patient make together.
To find out more, or to discuss working together
If you are a researcher, an engineer or a healthcare institution with an interest in any of the directions above, do get in touch. If you are a patient, every tool on this page can be used directly — no registration, and nothing left behind.
This page sets out the nature and the stage of the tools and research on this site. Anything labelled “In progress” or “Future plans” has not been validated and must not be used for clinical decisions. Every tool that is live is an aid and a reference; it cannot replace a doctor’s diagnosis or prescribing judgement. All computing is done on the user’s own device, and no data is uploaded or kept.