When does a smarter tool produce a healthier person?

A smarter health tool earns its place when it helps someone make a better decision or reduces a meaningful burden. To evaluate that benefit, I want to know what was measured, what the tool was compared with and what remains uncertain.

The studies below address different questions. Their findings can inform a buying or clinical decision, provided we keep those questions distinct.

Clinical AI under controlled conditions

A randomized study of 249 physicians in Indonesia, Kenya and the Netherlands reported better performance on standardized clinical vignettes with access to GPT-4o. The control group could not use the internet or clinical guidelines, and harms were not assessed. The findings concern performance under those test conditions, rather than demonstrated improvements in patient health. Multinational physician study

For a clinical team considering deployment, I would want a comparison with the resources clinicians normally use. I would also want the evaluation to measure the decisions the team expects the software to improve.

Wearables with a defined clinical purpose

FDA granted De Novo authorization to Abbott’s Libre Duo 10 Day system for diabetes management in people age two and older. It monitors glucose and ketone levels. The labeling says ketone readings should be interpreted with glucose levels, symptoms and other relevant factors. This is a defined diabetes indication; it does not establish a general longevity benefit. FDA classification order DEN250049

Before introducing monitoring into a care pathway, I would specify who reviews the information and what response is expected. Staffing, patient support and costs belong in that assessment.

Prevention research needs completed results

The DARE initiative describes studies examining wearable devices for health promotion and disease prevention. Its research overview is useful for understanding what is being investigated. It should not be presented as proof that the program has already improved prevention outcomes. DARE research overview

I would follow the individual studies for their results, including how consistently people use the devices, what actions follow the measurements and whether those actions improve health.

Aging clock changes need careful interpretation

A proteomic analysis used samples from 42 participants in a 12 week trial of rentosertib for idiopathic pulmonary fibrosis. Researchers examined six aging clocks and reported changes toward younger estimated ages. These are model estimates derived from protein measurements. They do not demonstrate lifespan extension or establish a benefit in healthy people. Several authors disclosed employment or leadership roles at the drug’s developer, Insilico Medicine. Proteomic clock analysis

My interpretation is that changes in a disease population require further study before they support a general claim about slowing aging. I would look for independent replication and clinical outcomes alongside the biomarker findings.

Testing patient outcomes in routine care

A Kenyan trial enrolled 9,691 patients across 16 primary care facilities. Clinical officers used an electronic medical record with or without AI assistance. The primary outcome was treatment failure within 14 days. The adjusted odds ratio was 0.77, with a 95% confidence interval of 0.55 to 1.08. The primary difference was not statistically significant. This result does not prove that the groups were equivalent or that the system had no possible benefit. Kenyan primary care trial

That is why I would choose the outcome before choosing the tool. Better documentation, better answers on test cases and fewer treatment failures are different findings.

A documentation benefit can still matter

In a randomized trial of 238 outpatient physicians, Nabla reduced measured time spent in the note by 9.5% relative to usual care. DAX Copilot did not show a statistically significant change on that primary measure. Clinicians reported occasional clinically significant inaccuracies with both systems. These results support a specific discussion about documentation, not a claim that patients became healthier. Ambient AI scribe trial

Before buying or expanding a tool, I would ask:

1. What decision or task should improve compared with the current process?
2. Who receives the output and owns the response?
3. What patient or operational outcome will we measure?
4. How will errors, missed responses and correction work count against the benefit?
5. What finding would make us stop or redesign the pilot?

Put the answers beside the proposal. They give the team a basis for deciding whether the tool is helping and when that decision needs to be revisited.

Originally published on my Substack. For hospital leaders evaluating AI, see my AI governance framework or explore my healthcare AI keynotes.