One patient comes in with fever.

No lab. No doctor. No second opinion.

Just symptoms.

It could be malaria.

It could be typhoid.

It could be meningitis.

It could be early sepsis.

These can look almost identical at first presentation. But the outcomes are very different.
 

Meningitis can kill within about 24 hours.

Sepsis can deteriorate within hours.

A wrong decision is not abstract. It is time lost. In practice, many of these cases get treated as malaria first. Sometimes correctly. Sometimes not.

I have been trying to understand whether AI based clinical decision support could meaningfully improve this kind of decision making in low resource settings. Not in a theoretical sense, but in the exact environment described above.

Constraints matter a lot here:

• No internet in many facilities
• Limited drug availability based on national essential medicines lists
• Health workers with varying levels of training
• High patient load and very little time per case

So any intervention has to work within those constraints or it does not work at all.

What I find difficult is not the technical side. It is the impact question. Even if a system can suggest better differentials or highlight danger signs, several things could still go wrong:

• The health worker may not trust the recommendation
• The required drug or referral pathway may not be available
• Over reliance could introduce new failure modes
• Improvements in decision quality may not translate into measurable outcome changes
   

There is also a broader question of comparison.

How does this approach compare to: Better training programs, Paper based triage tools, Increased supervision, Simple protocol reinforcement.

It is not obvious that AI is the highest leverage intervention here. I am especially interested in how to evaluate something like this properly.

If the goal is to reduce missed high risk cases among febrile patients, what is the most credible way to measure that in practice?

• Randomized rollout across facilities?
• Before and after comparisons?
• Proxy metrics like referral accuracy or time to escalation?

Each approach seems to have tradeoffs. More generally, I am trying to answer a simple question:

In this setting, does AI decision support meaningfully reduce harm, or does it just change how decisions are made without improving outcomes?

I would value perspectives from people who have worked on similar problems, especially around evaluation design and failure modes in real world healthcare systems.