> ## Content Index
> Fetch the complete content index at: https://www.healthdatacon.io/llms.txt
> Use this file to discover other available public pages before exploring further.

# Global Healthcare AI Needs Local Nursing Intelligence, Curtis and Riley Say
- URL: https://www.healthdatacon.io/global-healthcare-ai-needs-local-nursing-intelligence-curtis-and-riley-say/
- Published: 2026-10-16T03:19:34.000Z
- Updated: 2026-10-16T03:19:34.000Z
- Description: At HealthDataCon 2026, Sonya Curtis and Vanessa S. Riley explained why healthcare AI must fit local patients, resources and cultures, and why nurses need a formal role in its design and oversight.
- Author: Health Data Management
- Tags: Americas

An AI alert can travel across the world without changing a line of code. The conditions needed to act on it may change completely. A recommendation that assumes laboratory access, frequent monitoring and a rapid response team means something different to a nurse covering a rural ward overnight.

That gap was the focus of Sonya Curtis and Vanessa S. Riley’s presentation at HealthDataCon 2026\. Their argument placed nursing knowledge at the center of responsible implementation: someone must determine whether a technically functioning tool fits the patients, resources and relationships surrounding it.

“Care is always local,” Curtis, DHA said.

The conference’s [speaker directory](https://www.healthdatacon.io/speakers/) identifies Curtis as an AI nurse strategist and nurse manager in telehealth services at the Veterans Health Administration, and Riley, MSN as CEO of KhonsuAI and president of the Global Nursing AI Alliance. Riley also drew on her experience as an infectious diseases nurse practitioner, where local organisms, resistance patterns and available resources shape clinical decisions.

![Presentation chart showing United States and China shares of dataset sources in Celi and colleagues’ review of 2019 clinical AI literature.](https://storage.ghost.io/c/91/db/91db0927-dd9a-4578-9944-921da70f6885/content/images/2026/10/curtis-riley-clinical-ai-dataset-origins.png)

The presenters cited a 2022 review of 2019 research. Country shares came from manually reviewed subsamples. Slide credit: Sonya Curtis and Vanessa S. Riley / HDM’s HealthDataCon 2026.

## Where clinical AI learns

Curtis began with the geography of the evidence. A [global review](https://journals.plos.org/digitalhealth/article?id=10.1371/journal.pdig.0000022&ref=healthdatacon.io) by Leo Anthony Celi and colleagues, published in 2022, identified 7,314 eligible clinical AI papers from 2019\. In manually reviewed subsamples, the United States accounted for 40.8% of dataset sources and China for 13.7%.

Those figures describe a historical research sample. They nevertheless illustrate the question Curtis wanted organizations to ask: how closely do the people and circumstances represented in a model’s evidence resemble those where it will be used?

She identified three potential gaps. A data gap arises when patients differ from the population represented in development. A workflow gap appears when a recommendation assumes staffing, equipment or professional responsibilities that are unavailable locally. A context gap concerns culture, consent, family relationships, regulation and trust.

The workforce distribution adds another dimension. [WHO data](https://www.who.int/news-room/fact-sheets/detail/nursing-and-midwifery?ref=healthdatacon.io) estimate 29.8 million nurses worldwide in 2023, alongside a shortage of 5.8 million. Approximately 78% of nurses work in countries containing 49% of the world’s population. Riley argued that organizations looking to AI to relieve shortages must examine whether its assumptions fit settings with fewer people and resources.

![Illustrative sepsis alert used to discuss whether recommendations fit the resources available in different care settings.](https://storage.ghost.io/c/91/db/91db0927-dd9a-4578-9944-921da70f6885/content/images/2026/10/curtis-riley-illustrative-sepsis-alert-1.png)

This hypothetical alert supported a discussion of local resources and workflows; it was not a specific product or treatment instruction. Slide credit: Sonya Curtis and Vanessa S. Riley / HDM’s HealthDataCon 2026.

## One alert meets different care settings

The presenters used a hypothetical sepsis warning to make those assumptions visible. Their examples were composite scenarios, rather than reports of deployments in named countries or institutions. The same alert entered environments with different capabilities and expectations.

In an urban teaching hospital, laboratory services and a rapid response team might be available. Even there, the nurse must consider the patient’s appearance, the family’s concerns and changes across a shift. A score supplies information that still requires interpretation.

In a rural district hospital, the recommendation might call for tests that cannot be performed or monitoring that existing staffing cannot support. Referral may take hours. The relevant implementation question becomes whether the organization has a safe, feasible response pathway for the alert it has introduced.

In community and home care, a proposed hospital transfer may involve transportation costs, separation from family and relationships that shape decision-making. Riley described the nurse’s work of explaining risk, understanding concerns and involving family appropriately while respecting the patient’s preferences.

A different patient population creates further questions about whether a model’s inputs and recommendations remain suitable. The presenters’ examples showed that crossing a care boundary can matter even within one country. Staffing, available services and patient needs can vary substantially between neighboring organizations.

## Evidence travels with conditions

Riley grounded the discussion in studies that challenge assumptions about transferability. A [2021 evaluation](https://pubmed.ncbi.nlm.nih.gov/34152373/?ref=healthdatacon.io) by Andrew Wong and colleagues examined the then-deployed Epic Sepsis Model at one academic health system. At the studied alert threshold, it missed 67% of patients with sepsis while generating alerts for 18% of hospitalizations. Those findings concern that model and evaluation, rather than every sepsis prediction tool or subsequent version.

The operational concern is how clinicians respond to imperfect signals. Repeated unhelpful warnings can encourage dismissal, while excessive confidence in an automated output can displace independent assessment. Riley positioned nursing judgment between those risks.

She also drew on evidence from treatment studies. The 2011 [FEAST trial](https://www.gov.uk/research-for-development-outputs/mortality-after-fluid-bolus-in-african-children-with-severe-infection?ref=healthdatacon.io) found increased 48-hour mortality with fluid boluses among children with severe febrile illness and impaired perfusion in the studied African hospital settings. A separate [Zambian trial](https://jamanetwork.com/journals/jama/fullarticle/2654854?ref=healthdatacon.io), published in 2017, found higher hospital mortality with a specific early resuscitation protocol among adults with sepsis and hypotension, most of whom were HIV-positive.

These were treatment trials, not tests of AI. Riley used them to illustrate why recommendations must be evaluated alongside the populations and clinical resources supporting them. Embedding a protocol in software does not remove those conditions or establish that it will transfer safely.

## Contextual intelligence needs a formal role

“Validation is not the same as translation,” Riley said. The distinction broadens what implementation teams must examine beyond predictive performance.

She described contextual intelligence as the ability to combine clinical judgment, advocacy, communication and knowledge of local circumstances. This work includes recognizing when an alert does not fit, understanding why a recommended action is impractical and communicating concerns to someone who can respond.

Riley described three dimensions of translation. Clinical translation considers practice, workflows, populations and resources. Cultural translation examines expectations about consent, privacy, family roles and autonomy. Social translation concerns the trust, relationships and community conditions around care. Together, they require implementation teams to understand how a recommendation will be received and acted on.

Riley connected that capacity to Christine Tanner’s [judgment model](https://pubmed.ncbi.nlm.nih.gov/16780008/?ref=healthdatacon.io), which describes noticing, interpreting, responding and reflecting. AI may contribute to noticing a concerning pattern. Nurses still need opportunities to interpret findings, assess the person and learn from what happens next.

That point returned during the audience discussion about preserving expertise. The presenters emphasized direct patient assessment and verification of concerning measurements. Their response placed nursing skills inside the workflow, with an expectation that clinicians exercise and develop judgment as they use the tool.

International collaboration, Riley argued, should bring diverse nursing perspectives into development, governance, evaluation and implementation. Local knowledge should influence which tools are selected, how they are adapted and what organizations measure after deployment. Nurses also need a dependable channel for reporting a mismatch and an accountable person who acts on it.

She cautioned against treating international participation as a search for one uniform nursing perspective. Its value lies in retaining differences that would otherwise disappear from a standardized design. Sharing those differences can help teams identify assumptions about practice before a tool reaches patients.

The role continues as staffing, populations and practice change. Giving nurses responsibility to use an alert without influence over its design or oversight leaves much of their knowledge outside the improvement process.

## Six questions before deployment

Curtis and Riley closed with six questions for procurement and governance conversations, each addressing a condition that can determine whether an AI tool works in practice.

**Population:** Was the tool trained on patients like those the organization serves, and what evidence supports its use locally?

**Practice:** Does the recommendation fit nursing responsibilities and the way care is delivered?

**Resources:** Can the team carry out what the tool recommends with the people, equipment and services available?

**Culture:** How does implementation account for consent, privacy, family relationships and local expectations?

**Oversight:** Who can override the tool, and how is that person supported when clinical judgment differs from its output?

**Feedback:** How can nurses safely and simply report that a recommendation does not fit, and who will respond?

“Governance is not finished at go live,” Curtis said. Her final question gave leaders a practical place to begin: examine whether bedside observations can reach decision-makers and produce a change.

Riley closed with the design challenge underlying the session: “We should build AI capable of understanding and adapting to the world in which nurses actually practice.”