Lesson 16 · Technology and data in healthcare
New technologies and treatments: apps, artificial intelligence, assistive technology
Investigating a new technology means more than describing what it does. It means asking what has to be true for it to help, what happens when that condition fails, and whether the people who would benefit most are the people who will actually get it. Those three questions work on any technology, including ones that do not exist yet.
An app claims it can tell you whether a mole needs checking, from a photograph.
Write down what would have to be true for that app to be genuinely useful. Then write what happens when it is wrong — in each of the two directions it can be wrong.
- Investigate health apps: what they do well, what they do badly, and how they are regulated.
- Investigate artificial intelligence in healthcare: where it is genuinely strong and where it fails.
- Investigate assistive technology: the least discussed and arguably the most directly beneficial of the three.
- Apply a repeatable three-question test to any new health technology.
- Explain why a technology can be accurate and still make outcomes worse.
- Health app — software delivering a health function: tracking, coaching, symptom checking, medication reminders, therapy programs or clinical tools.
- Artificial intelligence — systems that find patterns in data and use them to classify, predict or generate. In health, most often applied to images, records and risk prediction.
- Training data — the examples a system learned from. Its limits become the system's limits.
- Algorithmic bias — systematically worse performance for a group under-represented in the training data, or the reproduction of an inequity already present in it.
- Assistive technology — any device or system that maintains or improves a person's functioning: hearing aids, wheelchairs, prosthetics, communication devices, home modifications, and increasingly software.
- Validation — testing a tool against a known standard, in the population it will be used on.
- Clinical decision support — a system that advises a clinician rather than replacing their decision.
- Automation bias — the tendency to accept a machine's output even when other evidence contradicts it.
New technologies arrive faster than any syllabus can list them, so what you need is a method rather than a catalogue.
- 1. What has to be true for this to help? Every technology has hidden requirements: a device, connectivity, literacy, a clinician who will act on the output, funding for the follow-up. List them, because each one is a place the technology fails silently.
- 2. What happens when it is wrong, in each direction? A false reassurance and a false alarm have completely different consequences, and a technology can be acceptable in one direction and dangerous in the other.
- 3. Who gets it first, and who would benefit most? When these are different people — and they usually are — the technology widens a gap unless something deliberate is done.
These three questions turn “investigate new technologies” from a description task into an analysis task, and they are what an examiner is looking for when a question names a technology you have never studied.
Quick check: Why is “what has to be true for this to help?” a more useful question than “what does it do?”
Enormous in number, tiny in the proportion that has been evaluated. Sort them by what they are trying to do, because the evidence differs sharply between the groups.
- Tracking and coaching. Activity, sleep, diet, quitting smoking, mood. Best evidence in this group, and the effective ones share features: goal setting, feedback, reminders, and some form of social or professional accountability. An app that only displays numbers changes little.
- Condition management. Medication reminders, symptom diaries, action plans for asthma or heart failure. Useful, and most useful when connected to a clinician rather than standing alone — the “data without a responder” problem from the previous lesson.
- Symptom checkers and triage. The riskiest group. Accuracy varies widely, and errors matter in both directions: false reassurance delays care for something serious, and excessive caution sends people to emergency departments unnecessarily.
- Therapy programs. Structured psychological programs delivered digitally. Good evidence for several, particularly for anxiety and depression, and they extend a workforce that cannot meet demand — especially valuable where there is no local clinician.
- Clinical tools for practitioners, such as dose calculators and guideline references, which are regulated more like devices than like consumer software.
Regulation. Software that performs a medical function — diagnosing, monitoring a condition, calculating a dose — is regulated as a medical device in Australia. Wellbeing and general fitness apps largely are not. So the same phone can hold a regulated tool and an unregulated one that looks identical, and the user cannot tell them apart by appearance. That is a real consumer problem and it links directly to Lesson 13.
Three general limitations. Most apps are never evaluated; engagement drops sharply within weeks, so an app that works in a trial may not work in life; and privacy is often poor, with health data collected and shared in ways users do not expect.
True or false: Software that diagnoses or monitors a condition is regulated as a medical device in Australia, while general wellbeing apps largely are not — and a user cannot tell the difference by looking.
Separate what these systems are actually good at from what they are claimed to be good at. The distinction is not subtle.
Where it is genuinely strong.
- Pattern recognition in images. Reading scans, retinal photographs, skin lesions and pathology slides. This is the most mature application, and performance on well-defined image tasks can match or exceed human readers.
- Prediction from records. Identifying patients at risk of deterioration, readmission or a specific complication, so attention can be directed before the event.
- Handling volume. Triaging screening images so humans review the uncertain ones, drafting documentation, and finding cases in large datasets. Extending scarce expertise is where the value is clearest, and it addresses the workforce constraint from Lesson 14.
Where it fails, and the failures are specific.
- It learns from data, so it inherits the data's inequities. A system trained mostly on one population performs worse on others. Where a dataset reflects unequal past care, a system trained on it can reproduce that inequality and present it as objective — which is worse than a human bias, because it is invisible, consistent and hard to argue with.
- It generalises poorly. Performance in the setting it was developed in frequently does not transfer to a different hospital, scanner or population. Validation in the population it will be used on is not a formality.
- Automation bias. Clinicians tend to accept machine output even against contrary evidence, so an error can propagate rather than being caught.
- Explainability and accountability. If a system cannot say why, a clinician cannot weigh its advice properly; and when it is wrong, who is responsible is genuinely unsettled.
The safest framing, and the one the evidence supports, is decision support: the system narrows the field and a clinician decides. Accuracy on a benchmark is not the same as improving outcomes in a real service, and the second is the only thing that matters.
Fill the gaps. Systematically worse performance for a group under-represented in the examples a system learned from is called algorithmic . The tendency to accept a machine's output even when other evidence contradicts it is called bias.
Assistive technology attracts the least attention of the three and has the clearest, most immediate effect on the thing this course cares about: what a person can actually do.
- Mobility. Wheelchairs, walking aids, prosthetics and orthotics, and vehicle modification.
- Sensory. Hearing aids, cochlear implants, screen readers, magnification, captioning. Untreated hearing loss is associated with isolation and with cognitive decline, so a hearing aid is not only a convenience.
- Communication. Speech-generating devices and communication software, which for some people are the difference between participating and not.
- Daily living and home modification. Rails, ramps, bathroom modification, adapted utensils. Cheap, unglamorous, and among the most cost-effective interventions available for maintaining independence.
- Monitoring and safety. Fall detection, medication dispensers and reminder systems, which support ageing in place.
Why it matters so much for this course. Assistive technology acts directly on functional ability, which is the definition of healthy ageing from Lesson 7. It does not cure anything; it changes what a person can do with the body they have, which is often the outcome that matters most.
The barriers, and they are not technical. Cost, and eligibility rules that differ between the disability, aged care and health systems, so the same device can be funded for one person and not another with the same need — the seam problem again. Long assessment waits. Devices supplied without training, which end up unused. And in remote areas, no local service to fit, maintain or repair a device, which makes the supply pointless.
Odd one out. Three of these are genuine reasons an artificial intelligence system can perform worse in a new hospital than in testing. Which one is not?
Four observations that hold for apps, artificial intelligence and assistive technology alike, and that make an evaluation coherent rather than a list of three descriptions.
- The technology is rarely the constraint. Funding, workforce, connectivity, training and eligibility rules are. A device with no one to fit it, an app with no clinician to act on it, and a prediction with no service to respond to it all fail for the same reason.
- Accuracy is necessary and not sufficient. A tool can be accurate and still worsen outcomes, if it produces action nobody can deliver, anxiety nobody resolves, or findings nobody follows up.
- Each one reaches the best-resourced first. Apps need a phone, data and confidence; AI is deployed where the data and money are; assistive technology needs assessment, funding and local service. The distribution rule from Lesson 15 holds for all three.
- Each one shifts work, usually toward the patient. Self-tracking, self-management, navigating funding systems, learning to use a device. That is agency for people who can carry it, and an additional barrier for people who cannot.
The strongest single sentence available on this dot point: a new technology changes health outcomes only to the extent that the system around it changes to use it. Everything else is capability without benefit.
Activities · investigate properly
For each, answer all three questions briefly: what has to be true, what happens when it is wrong, and who gets it first versus who needs it most.
An app that screens photographs of skin lesions and advises whether to see a doctor.
An artificial intelligence system that flags scans likely to show a fracture, for radiologist review.
A digital psychological therapy program offered to people on a waiting list.
A subsidy for hearing aids for older people.
Match each technology to the limitation most characteristic of it.
- Symptom checker app
- Image-reading AI
- Assistive device
- Tracking app
- Errors matter in both directions, and false reassurance delays care
- Performance may not transfer to a different population, scanner or hospital
- Supplied without fitting, training or local repair, so it goes unused
- Engagement falls sharply within weeks, so trial results overstate real-world effect
A health service is offered an artificial intelligence system that predicts which patients are likely to be readmitted within a month, so that extra follow-up can be arranged.
Write your advice. Cover: what has to be funded besides the software; how you would check that it works for every population the service treats, not just on average; what could go wrong if clinicians trust it too much; and what you would measure to know whether it improved outcomes rather than only predicted them accurately.
Lock it in. In one sentence, explain why an accurate health technology can still fail to improve health.
Investigate the contribution of new technologies and treatments to the Australian healthcare system.
Investigate asks for a planned inquiry ending in a supported conclusion — the question, the evidence, what it shows, and what it cannot show.
Plan: frame the question, name the evidence, report the findings, state the limitations, then conclude.
The inquiry asks what new technologies and treatments contribute to health outcomes in Australia, and whether that contribution is distributed across the population or concentrated within it.
The evidence base is clinical trial and health technology assessment data used by the Medical Services Advisory Committee and the Pharmaceutical Benefits Advisory Committee, utilisation and expenditure data from Australia's Health, and workforce and infrastructure distribution data.
Three contributions are consistent. Health apps and wearables extend monitoring and self-management beyond the clinic, improving adherence in chronic conditions. Artificial intelligence in image analysis and triage improves diagnostic speed and consistency, particularly where specialist reporting capacity is scarce. And assistive technology — prosthetics, communication devices and mobility aids — restores function and independence, which are determinants of health in their own right rather than merely conveniences.
Three limitations qualify this. Much of the evidence is short-term and drawn from trial populations that are younger, healthier and better connected than the population that will use the technology. Benefit depends on digital access and literacy, so the same technology can widen the gap it is intended to close. And regulatory approval establishes safety and efficacy but not cost-effectiveness at scale, so a technology can be genuinely effective and still displace more valuable spending.
This inquiry asks what new technologies and treatments contribute to health outcomes in Australia, and whether that contribution is distributed across the population or concentrated within it. The evidence base is clinical trial and health technology assessment data used by the Medical Services Advisory Committee and the Pharmaceutical Benefits Advisory Committee, utilisation and expenditure data from Australia's Health, and workforce and infrastructure distribution data. Three contributions are consistent across it: health apps and wearables extend monitoring and self-management beyond the clinic and improve adherence in chronic conditions; artificial intelligence in image analysis and triage improves diagnostic speed and consistency, particularly where specialist reporting capacity is scarce; and assistive technology such as prosthetics, communication devices and mobility aids restores function and independence, which are determinants of health rather than conveniences. Three limitations qualify this. Much of the evidence is short-term and drawn from trial populations younger, healthier and better connected than those who will use the technology. Benefit depends on digital access and literacy, so the same technology can widen the gap it is meant to close. And regulatory approval establishes safety and efficacy but not cost-effectiveness at scale, so a technology can be effective and still displace more valuable spending. The evidence therefore supports the conclusion that new technologies raise the system's capability substantially, but that their contribution to population health depends on distribution rather than on capability alone.
Assess the impact of artificial intelligence on the delivery of healthcare in Australia.
Assess requires a judgement against criteria. Describing what artificial intelligence can do, however impressively, is not an assessment of its impact.
Plan: criteria, the case for, the case against, a verdict that commits, then a qualification.
Artificial intelligence will be assessed against three criteria: whether it improves the accuracy and speed of care, whether it improves access for under-served populations, and whether its use is safe and accountable.
Against the first two criteria the impact is positive. Machine analysis of imaging and pathology detects abnormalities with consistency a fatigued human reader cannot match, and it does not vary across a shift, which raises the floor of diagnostic quality. Because it is not geographically constrained, it can extend specialist-level reporting to rural and remote services that cannot recruit a specialist at all, so it acts directly on the maldistribution that produces geographic health inequity. Administrative applications also reclaim clinician time for patient contact.
Against safety and accountability the impact is more troubling. Models trained on populations that under-represent Aboriginal and Torres Strait Islander Peoples or culturally diverse groups can perform worse for them, encoding existing inequity into a system that appears objective. Where a model's reasoning cannot be inspected, clinical responsibility for an error becomes unclear. And systems depend on large quantities of personal health data, raising privacy and consent questions the regulatory framework has not fully resolved.
On balance the impact so far is strongly positive where artificial intelligence supports a clinician's decision and materially risky where it replaces one. The technology raises consistency and extends reach, but it inherits the limitations of its training data and offers no mechanism of accountability of its own — so its value depends on it remaining an input to clinical judgement rather than a substitute for it.
Artificial intelligence will be assessed against whether it improves the accuracy and speed of care, whether it improves access for under-served populations, and whether its use is safe and accountable. Against the first two the impact is positive: machine analysis of imaging and pathology detects abnormalities with a consistency a fatigued human reader cannot match and does not vary across a shift, raising the floor of diagnostic quality; because it is not geographically constrained it can extend specialist-level reporting to rural and remote services unable to recruit a specialist at all, acting directly on the maldistribution producing geographic inequity; and administrative applications reclaim clinician time for patient contact. Against safety and accountability the picture is more troubling. Models trained on populations under-representing Aboriginal and Torres Strait Islander Peoples or culturally diverse groups can perform worse for them, encoding existing inequity into a system that appears objective. Where a model's reasoning cannot be inspected, responsibility for a clinical error becomes unclear. And these systems depend on large quantities of personal health data, raising privacy and consent questions the regulatory framework has not resolved. On balance the impact is strongly positive where artificial intelligence supports a clinician's decision and materially risky where it replaces one, since the technology raises consistency and extends reach but inherits the limitations of its training data and supplies no accountability of its own. Its value therefore depends on it remaining an input to clinical judgement rather than a substitute for it.
Return to card 01, the mole app. You described what happens in each direction when it is wrong. Which direction did you consider first? Most people think of the false alarm, because it is the visible one — and the false reassurance is the one that costs a life. Note which way your instinct ran, because it is the same instinct that makes overdiagnosis invisible.
Multiple choice · 5 questions
MC 1. Algorithmic bias in a health artificial intelligence system arises mainly because:
- A. the software is deliberately programmed to favour some groups
- B. the system learns from data that under-represents some groups or reflects unequal past care
- C. clinicians disagree with the system's outputs
- D. the system processes too many cases
MC 2. The riskiest category of consumer health app is generally:
- A. activity and sleep tracking
- B. medication reminders
- C. symptom checking and triage
- D. structured psychological therapy programs
MC 3. Assistive technology is particularly important in this course because it:
- A. cures the underlying condition
- B. acts directly on functional ability, which is what healthy ageing is defined by
- C. is the cheapest form of health technology in all cases
- D. requires no assessment or training to use
MC 4. Automation bias describes the tendency of a clinician to:
- A. refuse to use any automated system
- B. accept a machine's output even when other evidence contradicts it
- C. automate tasks that should be done manually
- D. prefer newer technology over older technology
MC 5. Which best explains why a technically accurate new technology may not improve health outcomes?
- A. Accuracy is impossible to measure in healthcare
- B. It changes outcomes only to the extent that the system around it changes to use it
- C. Patients generally refuse new technologies
- D. Regulators prevent accurate technologies from being deployed
Short answer · exam style
SA 1. Understand Band 4 Describe two new technologies being used in the Australian healthcare system, and explain one benefit and one limitation of each. (4 marks)
SA 2. Analyse Band 5 Explain how artificial intelligence could improve healthcare in Australia, and analyse one risk that would need to be managed. (5 marks)
SA 3. Evaluate Band 6 Investigate new technologies and treatments in the healthcare system, and evaluate the extent to which they are likely to improve the health of Australians. (6 marks)
📖 Comprehensive answers (click to reveal)
Multiple choice: 1 → B. 2 → C. 3 → B. 4 → B. 5 → B.
Activity 1 model answers: 1 → Must be true: it has been validated on the range of skin types it will meet, and a doctor is reachable when it says to go. When wrong: false reassurance delays diagnosis of a melanoma; a false alarm produces an unnecessary consultation and possibly an excision. First versus most: taken up by people already health-engaged with good phones, while the highest-risk group includes outdoor workers and older men who are the least likely users. 2 → Must be true: validated on this hospital's equipment and population, and radiologists retain the decision. When wrong: a missed fracture if flagging is trusted as exclusion; extra review time if it over-flags. First versus most: large, well-resourced departments, while the greatest workload pressure is often in regional services with fewer radiologists. 3 → Must be true: the person has a device, private space, connectivity and enough literacy, and there is a pathway to a clinician if they deteriorate. When wrong: someone needing more intensive care is left on a program that cannot help them. First versus most: people with the resources to engage online, while some of the highest need is among people without stable housing or connectivity. 4 → Must be true: assessment capacity, fitting, training and local repair exist. When wrong: an unfitted or unsupported device is abandoned, so the money produces nothing. First versus most: people who can navigate the assessment and funding system, while need is highest in the isolated, who are least able to.
SA 1 (4 marks): Two technologies described accurately from health apps, artificial intelligence, assistive technology or telehealth [1]. One benefit each, with a mechanism [1]. One limitation each, with a mechanism [2] — credit specifically for limitations of the system around the technology, such as no funded responder, no local fitting service, or no validation in the local population, rather than only for technical shortcomings.
SA 2 (5 marks): Improvement explained across at least two of: pattern recognition in images and pathology; prediction of deterioration or readmission so attention can be directed early; and extending scarce expertise by triaging volume, which addresses the workforce constraint [2]. One risk analysed in depth rather than named [2] — for example, algorithmic bias arising because the system learns from data that under-represents some populations or reflects unequal past care, so it can reproduce that inequality while appearing objective, which is harder to detect than a human bias. Management named for the fifth mark [1]: validation in the population it will be used on, monitoring performance by group, keeping the clinician as decision-maker, and clear accountability when it is wrong.
SA 3 (6 marks): At least three technologies investigated with specific capability described — health apps, artificial intelligence, assistive technology [2]. The hidden requirements identified for each: device, connectivity, literacy, a clinician who will act, funded follow-up, assessment and fitting [1]. Failure modes in both directions considered, and the point that accuracy is necessary but not sufficient [1]. Distribution addressed: each reaches the best-resourced first, and each shifts work toward the patient, so without deliberate action they widen gaps [1]. Judgement: the potential is real and largest in assistive technology and in extending scarce clinical capacity, but realised benefit depends on the system changing around the technology — funding, workforce, connectivity and eligibility — so an evaluation that assesses the device alone will overstate the gain [1].
The three questions are the durable content of this lesson. They will still work on technologies that do not exist yet, which is more than can be said for a list of current products.
Next, digital health takes one specific application — connecting health information across a fragmented system — and asks whether it has worked. It is the clearest test case for the claim that the constraint is the system rather than the technology.
Mark lesson as complete
Tick when you've finished the practice and review.