Artificial intelligence in Venezuelan hospitals: what works today and what doesn't yet

Almost every medical software vendor approaching a Venezuelan clinic today mentions artificial intelligence. Very few explain what conditions that AI needs in order to work, and practically none mention that a model performing well at one hospital can perform badly at the hospital next door.

This article goes in the opposite direction from those proposals. The thesis is simple and backed by published evidence and by real projects: in healthcare, the bottleneck isn't the algorithm — it's the data and the institution that will use it. That completely changes the order in which a clinic should spend its budget.

What "artificial intelligence" means in a Venezuelan hospital

When a vendor talks about "AI in hospitals," they usually blend three different things under the same label:

• Clinical AI: models that process clinical information (images, vital signs, physician notes) to support diagnostic or treatment decisions.
• Administrative AI: general language and classification models that help transcribe, sort messages, answer inquiries, and extract data from documents.
• Automation without AI: deterministic systems following clear rules (scheduling, reminders, billing, study reconciliation).

In Venezuela, what can actually produce measurable results today is concentrated in the last two categories: hospital automation and administrative AI with human supervision. Clinical AI, apart from specific imaging cases, still depends on conditions most institutions don't yet meet.

What kind of medical AI is actually approved today

The FDA list: AI-enabled medical devices

The US Food and Drug Administration (FDA) maintains a public list of AI- or machine-learning-enabled medical devices. That list identifies products authorized for marketing in the United States and is updated periodically.

The FDA itself warns the list is not intended to be exhaustive, but the proportion is stable and hasn't moved in years. As of the March 2026 cut, the list had accumulated 1,524 devices authorized since 1995, of which 1,164 — 76% — are radiology, according to The Imaging Wire's analysis of that same update. The rest is split between cardiology, neurology, and a handful of other specialties.

What is not on that list: any system that "manages a hospital" end to end, or that replaces general clinical judgment. What regulation has authorized are specific devices: nodule detection in CT scans, targeted diagnostic aids, MRI reconstruction, support for radiology and cardiology workflows.

Radiology dominates approved medical AI

The dominant category of real medical AI is imaging AI: radiology (X-ray, CT, MRI) and, to a lesser extent, cardiology. Most authorized systems analyze digital images already sitting in a PACS, triage urgent studies, help detect subtle findings, or reconstruct images to reduce radiation dose.

This matters far more than it seems in the Venezuelan context. That kind of AI presupposes a starting point very few institutions have: modern digital imaging equipment, a working PACS, stable internal connectivity, and enough study volume for the model to make economic sense. A clinic still handing out studies on paper or CD isn't one license away from using imaging AI: it's one digitalization project away.

Why a model works at one hospital and fails at another

A clinical model's performance depends on each institution's data and population, not just on the algorithm. The best-documented case in the sector is Epic's sepsis prediction model (Epic Sepsis Model, ESM), implemented at hundreds of US hospitals.

Epic Sepsis Model v1: external validation with poor results

In 2021, a University of Michigan team published in JAMA Internal Medicine a retrospective external validation of the ESM across 38,455 hospitalizations. The result was far below what the vendor reported:

• Area under the curve (AUROC) of 0.63 at the hospitalization level.
• Sensitivity of 33% — two out of every three sepsis cases were missed.
• Positive predictive value of 12%: out of every 100 alerts, 88 were false.

The editorial commentary published in the same journal concludes that the model predicts sepsis poorly, and that its widespread adoption despite that performance raises serious concerns about sepsis management at a national scale.

Epic Sepsis Model v2: better discrimination, same need for local validation

In February 2026, the same lead investigator headed a multicenter prospective study published in JAMA Network Open on version 2 of the model (ESM v2), analyzing 227,091 hospital encounters across four US health systems. The new version discriminates much better:

• Encounter-level AUROC between 0.82 and 0.92 depending on the site.
• Positive predictive values between 0.13 and 0.26, with high institutional variability.
• High alert burden and a number needed to evaluate (NNE) in the dozens of patients per sepsis case detected.

Translated: the new model sees sepsis better than its predecessor, but it still generates many alerts that don't end in a real case, and its actual performance varies from hospital to hospital. The authors explicitly recommend that any institution adopting the model run a local validation before trusting it, and design the clinical workflow to handle false positives.

That's the point no vendor will explain to you. Even in the best measured case, between 74% and 87% of alerts are still false positives, and the same software behaves differently at each hospital. The "validate locally" recommendation isn't paperwork: it's an institutional capability that demands your own historical data, someone who knows how to measure, and time.

What a Venezuelan clinic needs before it can use clinical AI

Before thinking about sepsis prediction models, ICU risk scores, or automated diagnostic support, a Venezuelan clinic needs something more basic: structured, historical, traceable data. Without that foundation, clinical AI isn't expensive or risky — it's simply inapplicable, because there's nothing to train it on and nothing to verify it against.

These are the six layers that have to be solved, in order:

Layers 1 and 3 sink the most projects in Venezuela, and they're exactly the ones I covered in the guide to electronic health records in Venezuela and in the article on why healthcare interoperability projects fail. Layer 4 is no minor local detail: an early-warning model that isn't available during an outage is not an early warning.

What I saw in the field. In the interoperability project I designed for a hospital network, with more than 10 legacy systems connected, 40% of the total project time was data cleanup, mapping, and validation — not building the event architecture.

The hardest problem wasn't technical but semantic: the same patient had different identifiers in each system — what admissions called "patient 00145" was "PA-2891" in the laboratory. We found duplicate patients, lab results with no physician identifier, and admissions with inconsistent timestamps between systems that were supposed to be in sync. That is exactly layer 1 and layer 3 of the table above, and it happened in a network that already had ten systems running. Any AI model trained on that data without the prior cleanup work would have learned the errors, not the clinic.

What does work today in a Venezuelan clinic

What produces measurable results today isn't general-purpose clinical AI: it's deterministic automation and, second, administrative AI with human supervision. It's worth separating the three categories, because vendors sell them blended together.

1. Automation without AI: the immediate return

Appointment reminders, physician scheduling, billing, lab study reconciliation, authorization follow-up. There's no model, there are rules. It's boring, it's cheap by comparison, and it's the only thing that guarantees results in weeks.

In a mid-sized Venezuelan clinic, nearly all the short-term return available is here: fewer calls to confirm appointments, fewer no-shows, fewer billing errors, less time lost moving paper. This is the layer where nearly every point of the analysis of clinic automation costs and priorities lines up.

2. Supervised administrative AI: when the error never reaches the patient

Note transcription, email and message classification, first-line WhatsApp response, data extraction from documents (orders, discharge summaries, reports). Here AI does add value, because:

• The cost of an error is relatively low.
• There's always a human validating before the error reaches the patient.
• The system can be switched off at any moment without leaving the hospital without life support.

It's close to what I described in the case of AI applied to logistics and healthcare: generic text AI in administrative processes, with human supervision and clear boundaries.

3. Clinical AI: imaging only, and only where the conditions exist

General-purpose clinical AI — models that "diagnose better than a doctor" or that "run the hospital" — not yet. The only viable case in Venezuela today is imaging AI at institutions with:

• Modern digital imaging equipment.
• An operational PACS.
• Sufficient study volume.
• Local validation capability (layer 6).

This isn't pessimism, it's sequence. Clinical AI becomes viable again once layers 1 through 6 are solved at your institution.

Questions to ask a vendor selling medical AI

Five questions separate a serious proposal from a sales presentation:

1. What population was the model trained and validated on? If training used data from another country and another health system, performance at your clinic is an unknown, not a promise.
2. What positive predictive value should we expect at our prevalence? This is the decisive question. If they answer with AUROC, with "95% accuracy," or by changing the subject, either they don't understand the problem or they'd rather you didn't. The ESM v2 studies show high AUROC but low positive predictive values, between 0.13 and 0.26 depending on the hospital.
3. How will it be validated locally before production, with what data, and who absorbs that cost? It's the explicit recommendation of the authors of the prospective ESM v2 study: without local validation and without understanding the alert burden, the risk of alert fatigue and over-reliance on the model is high.
4. Who is clinically accountable for a false negative? It must be written into the contract, not settled with a reassuring sentence in a meeting.
5. What happens when there's no internet or the power goes out? In the Venezuelan context, this question isn't hypothetical. The answer defines whether the system is usable or a luxury that only works when everything else does.

The correct sequence: data, infrastructure, and then AI

The order isn't negotiable, and getting it wrong is the fastest way to spend a digitalization budget with nothing to show:

1. Structured, reliable data first. Electronic health records, coded diagnoses, lab and imaging records that can be read, queried, and grouped.
2. Then infrastructure and interoperability that let that data circulate. Clinical systems that talk to each other (HIS, LIS, RIS, billing) and that survive outages without losing information or consistency.
3. Only then, clinical AI models. With years of your own data, clear clinical workflows, defined governance, and the ability to measure performance.

Investing in AI before solving the first two layers produces a pattern that has already repeated many times across the region: a system purchased, an implementation that stalls halfway, a contract nobody wants to renew, and a clinical team that distrusts every technology proposal from then on. The damage isn't only the money: it's that the next initiative — the good one — starts with the institution against it.

This sequence is the same one behind the overall picture of digitalizing Venezuela's health system and the one I applied in the clinical data interoperability case.

Before signing any AI proposal

If there's an artificial intelligence proposal on the table at your clinic or hospital, the useful decision isn't to accept or reject it: it's to know which layer you're standing on.

An institution without a structured health record doesn't need a model, it needs to solve layer 1 — and that costs considerably less and produces more immediate impact. An institution with a health record but no interoperability needs to solve layer 3 before considering AI that depends on cross-system data. In the technical diagnostic I review the real state of your systems, identify which of the six layers holds the bottleneck, and tell you what would have to be in place before investing in AI. No commercial proposal attached and no commitment.

Frequently asked questions

Why does a medical AI system work at one hospital and fail at another?

Because performance depends on each institution's population and data, not just on the algorithm. The multicenter validation of Epic Sepsis Model v2 published in 2026 found an AUROC between 0.82 and 0.92 depending on the site, but with positive predictive values between 0.13 and 0.26 and high institutional variability. That is why the authors recommend local validation before adopting any clinical model.

Can a small clinic in Venezuela use artificial intelligence today?

It can use administrative AI with human supervision — note transcription, message classification, first-line responses — and non-AI automation for scheduling, reminders, and billing. Clinical AI requires a structured electronic health record, the institution's own historical data, and local validation capability, conditions most small clinics don't yet meet.

How is artificial intelligence used in Venezuelan hospitals?

Today, AI in Venezuelan hospitals is used mostly for administrative automation (scheduling, reminders, billing), administrative text AI (transcription, message classification, initial responses), and in some cases radiology imaging AI at institutions with digital equipment and an operational PACS.

Which hospitals in Venezuela already use artificial intelligence?

There is no public registry documenting which Venezuelan institutions use clinical artificial intelligence, and any list circulating without a verifiable source should be treated with skepticism. What is observable is the pattern: real adoption sits in administrative automation and in diagnostic imaging within private institutions that already have digital equipment and an operational PACS. If a vendor presents you with references of Venezuelan hospitals using their model, ask for the name of the institution and the contact details of whoever supervises it clinically.

How much does it cost to implement AI in a Venezuelan hospital?

The real cost depends less on the model license and more on the maturity of the six layers: electronic health records, interoperability, infrastructure, governance, and local validation capability. The investment needed in data and integration is often larger than the price of the AI software; that's why it makes sense to start with non-AI automation and supervised administrative AI.

What challenges does Venezuela face in adopting AI in healthcare?

The main challenges are incomplete data histories, records kept on paper or in free text, lack of interoperability between clinical systems, unstable infrastructure, and the absence of a local model validation culture. Until those are solved, clinical AI remains a promise rather than a tool.

If you have an AI proposal on the table and don't know which layer your institution is standing on, that is exactly the kind of decision I help sort out at Code by Meléndez — with direct experience from the real healthcare data interoperability case cited above.