Bangladesh signed up for AI-enabled healthcare, what should that actually mean?

In December 2025, in a conference room in Tokyo, Bangladesh signed a document that researchers in my field had been waiting years to see. The National Health Compact, agreed at the UHC High-Level Forum, lists six pillars for reforming the country's health system by 2030 (National Health Compact, Bangladesh, UHC High-Level Forum, Tokyo, December 2025).

Tucked into the sixth one is a line that most readers probably skimmed past: the government commits to supporting "the scale-up of digital health solutions, AI-enabled tools, and other emerging technologies that enhance equitable, efficient, and high-quality service delivery."

I read that sentence more than once. It's the first time I've seen AI named directly in a Bangladeshi health policy document rather than treated as something for another country to figure out first. But a compact is not a blueprint. It says AI-enabled tools should scale up. It doesn't say which tools, for which pillar, solving which specific problem. That gap is where my own work sits, and I think it's worth spelling out what filling that gap could actually involve.

Bangladesh's public health spending was 0.40 per cent of GDP as of 2021, well behind Bhutan at 2.21 per cent, Nepal at 1.80 per cent, and even Pakistan at 0.84 per cent. Out-of-pocket costs make up 73 per cent of total health spending, one of the highest shares in South Asia, and that burden has been pushing people into poverty. Health costs pushed 3.74 per cent of the population below the 2.15 dollar a day poverty line in 2016, up from 3.11 per cent in 2010. The health budget for FY2024-25 came to just 0.74 per cent of GDP (Centre for Policy Dialogue, Policy Brief 2025(2), April 2025).

In a system this constrained, a badly targeted tool doesn't just underperform. It actively wastes money the country doesn't have to waste. That's the argument for getting the "how" right before scaling anything up.

The compact's second pillar calls for expanding community-based and telehealth service delivery, especially in hard-to-reach areas. I built a chatbot for dengue symptom triage in Bangladesh, designed to run on simple decision-tree logic in low-bandwidth clinical settings in Bengali and English. In testing on 2019-2023 case data, the model correctly flagged severe cases well enough to be useful as a first filter, and in a small pilot with 50 users, three in four said they were satisfied with it. It''s a small example, but it points at something the compact gestures toward without detailing: triage tools that work where the internet is slow and the nearest doctor is an hour away, not just in Dhaka.

The third pillar talks about redesigning the Essential Service Package around life-course prevention, including maternal health. My colleagues and I built a tool for maternal health risk prediction that combines two approaches: a set of medical rules a doctor would recognise, and a machine-learning model trained on patient data. We tested it against clinical data and then asked actual doctors, 14 of them, to review its recommendations. When we showed them the combined version, the one that explains its reasoning in terms they recognize rather than just a risk score, more than half said they would trust it enough to use in practice. That distinction between a tool that shows its work and one that doesn't seems small on paper. To a doctor deciding whether to refer a patient, it isn't.

The fifth pillar is where things get more complicated, and more interesting. It commits to establishing a National Health Security Office to manage a National Health Fund and lead what it calls strategic purchasing, while targeting vulnerable groups, urban slum residents, widows, people over 70, people with disabilities, for prioritized services.

This is not a new idea in Bangladesh. A draft National Health Protection Act has existed since 2014, proposing something similar: a National Health Protection Authority, a health card system, and lists sorting citizens into income categories, below poverty, marginal income, middle income, and above, to determine who gets subsidized care (Draft National Health Protection Act, 2014). The draft law even requires annual audits and treats misuse of the system as a punishable offense. More than a decade later, the compact is effectively picking up where that draft left off, this time with an actual signed commitment and a 2030 deadline attached.

Here's the part I keep coming back to. Any list sorting citizens into income categories is a judgment call, whether or not anyone calls it that. Someone, or something, decides who counts as "below poverty level" versus "marginal income," and that decision determines who gets a subsidized health card and who doesn't. If this system gets digitized without checking for bias first, and I've seen this exact pattern before in disaster aid allocation, it risks locking in existing unfairness at a much larger scale and a much faster pace than a paper-based system ever could. The bias-correction techniques I used to fix historical unfairness in post-flood aid rankings, adjusting a model so it stops treating location as a shortcut for actual need, apply directly to this kind of eligibility list too.

I want to be careful here about a second idea some of my recent work touches on: using AI to flag suspicious claims once a national health fund is up and running. That's a real and useful application in principle. But I''ll say plainly that my own research on this, which mixes rule-based logic with pattern recognition, has so far been tested only on simulated financial data built for a research paper, not on real claims from any health system, and my own paper is explicit that those results shouldn't be read as proof the method works on real-world fraud. It's a promising direction. It is not a finished tool, and Bangladesh shouldn't wait for it to be one before doing the more basic work of auditing its eligibility data for bias.

The compact's first and sixth pillars also call for modernizing infrastructure and expanding private sector participation, including private diagnostic centers. Bangladesh's health data will increasingly sit across public hospitals, NGO clinics, and private providers that have no particular incentive to share patient records with each other, and plenty of legal and ethical reasons not to.

This is the kind of problem a technique called federated learning is built for, where separate institutions train a shared model together without any of them handing over their raw patient data. I built a framework called MedHE to test exactly this idea, combining that data-sharing approach with encryption strong enough that even the central server coordinating the process can't read what any single institution contributed. In our tests, simulating five separate clients working together on a health-related text classification task, the encrypted, privacy-protected version performed statistically as well as an ordinary shared model, while cutting the amount of data that needed to move between institutions by more than 97 per cent.

It's still a research demonstration rather than something running across real Bangladeshi hospitals, but it shows the privacy problem and the efficiency problem can be solved together, not traded off against each other. Bangladesh's own Digital Health Strategy, in place since 2023, will need something in this direction eventually if it's serious about connecting a fragmented network of providers without asking any of them to give up their patients' privacy first.

The compact already has a monitoring framework with real targets: raising the UHC service coverage index from 54 to 65 by 2030, cutting the share of households facing catastrophic health spending from 42 per cent to 35 per cent, and more (National Health Compact, Bangladesh, 2025). A few additions would make the AI-enabled tools commitment more than a sentence in a document.

First, any tool used to determine eligibility or priority for the National Health Fund should be checked for bias before it goes live, the same way financial and clinical outcomes are already tracked.

Second, fairness and explainability deserve their own line in the compact''s list of indicators, not just a mention under one pillar. What gets measured tends to be what gets built.

Third, the government should draw on the research already happening inside the country. Bangladeshi researchers have working models for triage, maternal risk prediction, and privacy-preserving collaboration across institutions, at different stages of readiness. None of this needs to be imported or reinvented from scratch.

Fourth, the long-dormant National Health Protection Act should be revisited with this compact in mind. A legal framework for a health card system already exists in draft form. Updating it with explicit rules for checking algorithms for bias would save the country from relearning lessons other systems have already paid for.

I think about the years between that 2014 draft act and this 2025 compact, and how much could have been built in between if the tools had been treated as ready rather than hypothetical. Some of them are ready now. Others still need more testing before anyone should trust them with real decisions, and it matters to be honest about which is which. Bangladesh has already signed the intention. What remains is turning one sentence in a Tokyo hotel conference room into something that actually makes a difference to a patient in Sunamganj or Satkhira.