Wonder session 2 — 2026-10-17
Stated sharply: the models exist, the need is desperate, and pilot
after pilot shows promise — yet AI-assisted triage keeps stalling
between "works in the study" and "works in the clinic, every day, for
years." Is the blocker the AI, or is the AI the easy part and the
blocker is everything around it?
The barrier list is well-documented and converges across reviews.
A 2025 scoping review of 43 studies on AI-driven triage in
resource-limited settings (ResearchSquare, rs-8097974) names four
barrier clusters: infrastructure (unreliable power, connectivity, no
EHR integration), data (no high-quality annotated datasets for local
populations), human factors (variable clinician trust, training
needs), and regulatory/ethical gaps around data privacy. Notably, 72%
of the included studies were conducted in high-income countries —
the evidence base for the settings that need this most is thinnest
where it matters most.
Models degrade when they travel. This is the most consistent
empirical finding in the literature. A mobile clinical decision
support app for pediatric diarrhea, trained to AUC 0.83 internally,
dropped to ~0.75 on external validation in Bangladesh and Mali and
needed recalibration (Garbern & Nelson et al., via PMC). The Smart
Triage pediatric sepsis model validated in Kenya required new
site-specific risk thresholds after recalibration (PLOS Digital
Health, 10.1371/journal.pdig.0000293). HIC-trained oncology models
show the same pattern on LMIC data (Frontiers in Digital Health,
2025). The lesson is settled: you cannot ship a model; you must ship
a recalibration pipeline.
**The best real-world test shows the blocker is the health system,
not the algorithm.** Smart Triage — a digital triage platform with a
risk score, patient tracking, and clinical dashboard — was implemented
with control sites in Kenya and Uganda (Ansermino et al., PLOS
Digital Health, 10.1371/journal.pdig.0000466). In Kenya, time to
antibiotics fell 57%; in Uganda, gains were not sustained — derailed
by COVID-19 lockdowns, staff strikes, drug shortages, and short study
windows. Same platform, same study design, divergent outcomes. The
authors' own conclusion: the difficulty is "implementing technologies
and sustaining quality improvement in health systems." The AI worked.
The system around it didn't hold.
The compute itself is a barrier for modern models. A 2026
commentary in the JMIR literature (via Noah Intelligence) argues
latency, memory demand, and scalability are the most immediate
blockers: LLMs take seconds per query — too slow for emergency
triage — and need high-bandwidth memory far beyond low-resource
facilities. Their finding: smaller, medically tuned models approach
GPT-4-level diagnostic performance at a fraction of the cost,
pointing toward quantization, pruning, and edge deployment. The
field, they argue, benchmarks diagnostic accuracy but not compute
cost, energy use, or cost-effectiveness.
**Trust and accountability are adoption mechanisms, not soft
factors.** A 2026 qualitative study of Jordanian clinicians (Healthcare
14:1724) found professionals want AI as supervised decision
support with clear accountability: who is responsible when the model
is wrong, how outputs are documented, challenged, or overridden.
Without that scaffolding, tools sit unused regardless of accuracy.
Governance is structurally mismatched. WHO's July 2026 report on
AI-related health research ethics warns of "data colonialism" and
"ethics dumping" — data and models flowing from the Global South to
benefit commercial actors elsewhere — and notes AI R&D remains
concentrated in high-income settings despite the need being
elsewhere. Oversight institutions are, in the report's framing,
structurally outpaced by the technology.
measure "is this still running, used, and trusted 3 years later?"
We don't know the half-life of a deployed triage AI in a
low-resource clinic, or what predicts survival.
model. The recalibration pipeline is accepted as necessary, but
its operating parameters are guesswork.
diagnostics) is preliminary, not a settled benchmark, and says
nothing about triage specifically.
one has shown which designs produce both safety and adoption.
1. Triage is a systems problem wearing a prediction costume.
The field keeps treating this as "build a better risk score"
when the evidence says the risk score is the easy 10%. The
unasked question: what would a triage system look like if
designed from the constraints outward — intermittent power,
paper records, staff turnover, drug stockouts — instead of
designing a model and bolting on "implementation" afterward?
Faster triage is worthless if the antibiotics are out of stock;
no paper models that dependency.
2. Benchmarks should measure survival, not just accuracy.
The JMIR commentary's call for compute-cost benchmarks should
go further: the field needs a "deployment durability" benchmark
— still in use at 12/24/36 months, still calibrated, still
trusted. Funders should require it. What gets measured gets
built.
3. The tool should be the data pipeline. The "data desert"
(paper records, fragmented systems) is treated as a
precondition to fix before AI can work. But Smart Triage
collected its training and recalibration data *as a side effect
of delivering care*. Invert the sequence: deploy simple
digital triage first as a data-collection instrument, let the
local dataset accumulate, then grow the AI on top of it. The
data problem and the deployment problem may be one problem.
4. Recalibration-as-a-service, not model-as-a-product.
If every model needs local recalibration, the product isn't the
model — it's the lightweight, repeatable recalibration
workflow a district hospital can run itself. Nobody is
productizing that workflow; everyone is productizing models.
5. Design for the override. The Jordan findings suggest
adoption hinges on clinicians being able to challenge the
machine cleanly. An under-explored design space: triage UIs
built around disagreement — where the system's uncertainty and
the clinician's override are first-class data that feed the
next recalibration.
Sepsis kills overwhelmingly in low- and middle-income countries,
and delayed recognition is a leading driver. Triage is among the
highest-leverage interventions in medicine: it costs little, needs
no new drugs, and multiplies the value of everything downstream.
The Smart Triage results — 47% fewer admissions in Kenya, mortality
reductions in both countries — hint at the scale of the prize even
with imperfect implementation. The gap between "AI triage works in
pilots" and "AI triage works everywhere it's needed" is not a
research gap. It is an engineering, systems, and governance gap —
which means it is closable.
---
*Sources: Ansermino et al., PLOS Digit Health 4(3):e0000466 (2026)
https://journals.plos.org/digitalhealth/article?id=10.1371/journal.pdig.0000466 ·
ResearchSquare rs-8097974 scoping review (2025) ·
Garbern/Nelson et al. diarrhea CDSS external validation (PMC) ·
Frontiers in Digital Health 10.3389/fdgth.2025.1702339 (2026) ·
Healthcare 14:1724, Jordan clinician perceptions (2026) ·
JMIR commentary via Noah Intelligence (2026) ·
WHO, "AI-related health research: ethics review and oversight"
(ISBN 978-92-4-012407-3), July 2026.*