From Device Data to Trusted Decision Support: Building the Foundation for AI in Hospital Insulin Management

Abstract

The adoption of artificial intelligence (AI) tools for hospital insulin management is currently limited by data fragmentation and difficult integration into clinical workflows. This commentary examines the data infrastructure requirements for safe AI deployment in clinical settings. Device-mediated and clinician-administered dosing are the two methods by which insulin is managed in hospitals. In device-mediated dosing, glucose and insulin data often remain siloed within proprietary device ecosystems outside the electronic health record (EHR). In clinician-administered dosing, relevant data elements typically exist within the EHR but are distributed across workflows in ways that limit their usefulness for decision support. The Integration of Connected Diabetes Device Data into the Electronic Health Record (iCoDE) initiative is a standard for integrating device-generated diabetes data into clinical systems, which can lay the foundation for organizing hospital data in support of the development of trustworthy AI. A staged roadmap for hospitals building towards AI-ready insulin management infrastructure is presented along with governance requirements for trustworthy deployment. The value of iCoDE is that it helps define the conditions under which such AI can become clinically meaningful, trustworthy, and scalable.

1. Introduction

In hospital insulin management, the limiting factor for artificial intelligence (AI) is increasingly not whether a model can generate a recommendation, but whether the surrounding data environment is sufficiently structured, interoperable, clinically usable, and trustworthy to support safe implementation. Interest in AI for inpatient insulin management has grown substantially in recent years, as the field has advanced beyond theory to practical applications such as machine learning-based dose prediction, reinforcement learning approaches, and real-time decision support systems. This momentum reflects both the clinical importance of glycemic management in the hospital and the growing recognition that insulin dosing is a high-value target for data-driven support. At the same time, the promise of these methods has outpaced their routine use in practice.

Broader adoption remains constrained by limited external validation, difficulty integrating recommendations into the clinical workflow, and persistent concerns regarding privacy, transparency, and oversight. In this commentary, we argue that the central challenge in hospital insulin AI is not simply model performance, but the readiness of the underlying clinical data environment to support safe, scalable, and trustworthy decision support. To make this case, we distinguish device-mediated from clinician-administered insulin dosing, explain why inpatient insulin management is a uniquely demanding use case, and argue that the primary bottleneck to adoption is the data substrate rather than the model itself. We then examine how the Integration of Connected Diabetes Device Data into the Electronic Health Record (iCoDE-2) project contributes most directly to the standardization and integration of device-generated diabetes data, while also offering a broader informatics framework for thinking about data organization, workflow, and trust in hospital-based insulin decision support.

2. Why AI Implementation in Inpatient Insulin Management Is Difficult

Inpatient insulin management is a clinically complex and operationally demanding problem. Acute illness, surgical procedures, changes in nutritional intake, corticosteroid exposure, and transitions between levels of care can all alter insulin sensitivity in ways that are difficult to predict [1]. At the same time, insulin dosing decisions often depend on information distributed across multiple data sources, including point-of-care glucose values, continuous glucose monitoring (CGM) data, medication administration records, and other elements of the clinical record. These decisions may need to be revisited several times within a single day and across different members of the care team, making it difficult to maintain a complete and temporally coherent picture of the factors that should inform dosing.

The safety implications of this complexity are substantial. Insulin is consistently recognized as one of the highest-risk medications used in the hospital setting, and errors in timing, dose, or interpretation of the clinical context can have immediate and clinically meaningful consequences [2]. Accordingly, AI implementation in inpatient insulin management is best understood across two broad insulin-dosing paradigms: device-mediated dosing and clinician-administered dosing. Each of these presents distinct challenges for data integration, workflow implementation, and trust (

Table 1. A comparison of device-mediated and clinician-administered insulin dosing paradigms.

Dosing Mechanism	Device-Mediated Dosing	Clinician-Administered Dosing
Description	Onboard algorithms and interfaces that allow the patient to administer insulin through an insulin pump or automated insulin delivery system, using pre-programmed basal rates, bolus settings, CGM-linked automation, or adaptive dosing algorithms embedded in the device ecosystem.	Insulin is ordered, guided, or adjusted by clinicians and administered by nursing staff or other hospital personnel, typically through paper-based protocols, EHR-based calculators, or order-set logic, dedicated glycemic-management platforms, or more advanced software-enabled decision support.
Examples	Medtronic MiniMed 780G/SmartGuard [3] Tandem Control-IQ+ [3] Insulet Omnipod 5 [4] Beta Bionics iLet Bionic Pancreas [5]	Paper-based sliding scales and protocols EHR-based insulin calculators, order sets Glytec—Glucommander [6] Monarch Medical/Glooko—EndoTool [6] MDN—GlucoStabilizer [7] Clinical Software—GlucoTab [8]

).

Device-mediated insulin delivery includes both pump-based systems and automated insulin delivery (AID) platforms. In these environments, the sensor, algorithm, user interface, and insulin delivery mechanism are tightly coupled within a single device ecosystem. Systems such as Medtronic’s MiniMed 780G (Medtronic, Northridge, CA, USA) and Tandem’s Control-IQ (Tandem Diabetes Care, San Diego, CA, USA) use onboard algorithms to adjust basal insulin delivery, issue correction doses, and respond to CGM data in near real time. Because these systems are designed to operate within proprietary device environments, much of the relevant data remains outside the EHR unless specifically integrated. This creates a distinct informatics challenge: enabling device-generated glucose and insulin data to move into clinical records in standardized, interpretable, and interoperable forms.

Clinician-administered hospital insulin management, by contrast, relies on insulin that is ordered by clinicians and administered by nurses or other hospital staff. Dosing decisions may draw on EHR-based medication orders, medication administration records, point-of-care glucose values, CGM data when available, nutrition-related information, and other elements of the clinical record. In many hospitals, these core data elements already exist in structured form within the EHR; however, they are often distributed across documentation modules, workflows, and team roles in ways that complicate rapid longitudinal interpretation. In this setting, the challenge is less whether the data exist than whether they are sufficiently connected, temporally aligned, clinically interpretable, and workflow-ready to support advanced decision support. Within this context, iCoDE is best understood as a useful informatics framework for organizing and operationalizing data in support of more trustworthy insulin-related decision support.

3. AI for Insulin Dosing

The usefulness of AI for hospital insulin management depends less on model architecture alone than on whether the relevant clinical data can be assembled into a coherent and clinically interpretable longitudinal picture. To support both patient care and AI reasoning, the surrounding data environment must be able to link glucose measurements, insulin administration events, timing, nutritional context, and broader clinical status in ways that are meaningful at the bedside. In this context, AI readiness requires more than the presence of structured data fields. It requires data that is sufficiently complete to reflect the clinical situation, attributable to specific sources for provenance, temporally aligned, and accessible within the clinical workflow. These requirements differ across the two insulin-dosing paradigms. In clinician-administered hospital management, the challenge is often to organize, connect, and present already-available hospital data in ways that support trustworthy decision support. In device-mediated management, key data may remain within proprietary manufacturer environments, which can limit interoperability with hospital information systems.

Although machine learning models trained on EHR data have shown the ability to predict insulin doses more accurately than standard clinical calculators [9], broader clinical adoption of these tools remains limited. Key implementation barriers include limited external validation, concerns regarding transparency and privacy, cost, and persistent difficulty integrating recommendations into real-world workflows. In device-mediated management, the continued challenge is to standardize and integrate external device data into the EHR. In clinician-administered management, the challenge is to make hospital data more interoperable, temporally coherent, and clinically usable for decision support. In both settings, the principal bottleneck to deployment is less what the model can do than whether the surrounding data infrastructure can support its safe, interpretable, and meaningful use.

4. iCoDE: Standardizing Device Data Integration

The iCoDE project is a multi-stakeholder initiative organized by the Diabetes Technology Society that brings together clinicians, researchers, manufacturers, EHR vendors, government partners, and patient representatives to address the fragmentation of diabetes device data [10,11]. Its most direct contribution is in the device-mediated data pathway. In workflows involving insulin pumps, AID systems, and connected insulin pens, glucose and insulin delivery data may otherwise remain siloed within manufacturer platforms or be presented to clinicians through summaries that exist outside the EHR. iCoDE addresses this challenge by defining a canonical insulin data model to reconcile variability across devices and by advancing interoperability pathways based on existing standards such as Fast Healthcare Interoperability Resources (FHIR) [11]. The project also introduced the Insulin Dosing Profile, a one-page summary that displays insulin delivery and CGM data together to support clinical decision-making [11]. Collectively, these contributions help create a more standardized and clinically interpretable pathway for integrating external device data into the EHR.

In clinician-administered hospital insulin management, by contrast, the core data elements relevant to insulin dosing often already reside within structured EHR systems. In this setting, iCoDE should be understood less as a direct data-ingestion solution and more as a useful informatics framework. Its emphasis on standardized representation, interoperability, data quality, and workflow-aware visualization remains relevant to the task of organizing already-structured hospital data into forms that are more coherent, clinically interpretable, and trustworthy for advanced decision support. Framed this way, iCoDE’s relevance to hospital insulin AI is twofold: it contributes directly to the integration of externally generated device data, while also offering a broader model for how diabetes-related data can be organized and operationalized in support of more trustworthy insulin-related AI.

5. Trustworthy AI for Insulin Management

Interoperable and structured data provide the foundation for hospital insulin AI, but trustworthiness depends on whether those data are embedded within a governance structure capable of supporting safe, transparent, and accountable use. The transition from structured data to trusted decision support therefore requires more than technical readiness alone; it also requires clear expectations for oversight, transparency, and accountability. Recent work on medical AI in diabetes has identified six core attributes that are central to patient and clinician trust: accuracy, reproducibility, privacy and security, transparency, human oversight, and fairness [12]. In the context of hospital insulin management, these attributes are not abstract principles. They are operational requirements for implementing decision support around a high-risk medication in a fast-moving clinical environment.

Accuracy and reproducibility are especially important for insulin-related AI because errors in recommendation logic, timing, or contextual interpretation can have immediate clinical consequences. Trustworthy deployment therefore requires performance to be demonstrated against strong comparators and across multiple institutions, workflows, and patient populations, using clinically meaningful reference standards rather than technical metrics alone. Privacy, security, and transparency are also important. Hospitals need clear documentation of the data sources used by a model, the assumptions and limitations built into the model, and the safeguards in place to protect patient information. Clinicians should also be able to determine whether a recommendation reflects recent glucose trends, prior insulin administration, incomplete data capture, or other clinically relevant factors that may affect its interpretation.

Hospital insulin AI also requires governance that extends beyond initial validation. Post-deployment oversight should include monitoring for model drift, mechanisms for recalibration or review when performance changes, and clearly defined pathways for human oversight in higher-risk or low-confidence situations [13,14]. Monitoring plans should define minimum scheduled review intervals, such as weekly during early deployment and monthly or quarterly once performance is stable, plus event-triggered reviews. The appropriate cadence should be locally defined, depending on the tool’s intended use, level of automation, and institutional risk assessment. Event-triggered review should occur after major changes to insulin-management inputs or related workflows, or when safety, data quality, clinician feedback, concordance between model recommendations and clinical judgement, or other performance signals suggest that model recommendations may no longer be reliable. Many health systems do not yet have the institutional infrastructure to support this level of continuous monitoring and accountability. In this setting, building oversight capacity is not secondary to model development; it is part of the core work required to make insulin-related AI clinically trustworthy and operationally sustainable.

Fairness in insulin-related AI is not only driven by algorithm design. These tools may inherit inequities from the data used to train and deploy them, with potential consequences not only for fairness, but also the safety and reliability of insulin decision support [15,16]. For example, insulin requirements may vary across patient populations because of differences in insulin sensitivity, glycemic patterns, nutrition, comorbidities, care access, documentation practices, and diabetes technology use [17,18,19]. AI decision support could therefore amplify inequities if it is trained on data in which some groups are underrepresented, have less CGM or pump data, experience different documentation patterns, or receive systematically different care. Ensuring safe and equitable deployment requires subgroup validation, monitoring for differential safety outcomes, and governance processes that can detect when model performance differs across racial, ethnic, socioeconomic, age, sex, or technology-access groups. Data infrastructure modeled on iCoDE principles [10,11] can support this work by preserving provenance, tracking missingness and data latency, and linking glucose-insulin data to EHR-derived contextual variables (e.g., acuity levels, steroid exposure, race/ethnicity, age, etc.). Such variables are essential context for stratifying model performance, determining whether recommendations are safe, interpretable, and equitable across the population served by the hospital.

6. A Roadmap for Hospitals Seeking to Implement AI in Insulin Management

Hospitals considering AI for insulin management begin from very different points of maturity. Some continue to rely primarily on paper-based protocols, whereas others use EHR-based calculators, dedicated glycemic-management platforms, or partial device-data integration. Movement toward AI-enabled insulin decision support should begin with an honest assessment of local readiness rather than an assumed ideal end state. The process is inherently staged: hospitals must first understand their current insulin-management environment, and then strengthen the underlying data substrate, make those data clinically usable, establish appropriate governance, and introduce AI use cases in a deliberate and sequenced way.

Table 2. A staged roadmap for building AI-ready insulin management infrastructure.

Roadmap Step	Primary Objective	Key Activities	Implication for Hospital Insulin AI
1. Assess the current state	Define the local starting point	Characterize current use of paper protocols, EHR calculators, vendor platforms, and device data; map workflows and pain points	Clarifies local readiness and prevents premature AI deployment
2. Build the data substrate	Create a structured insulin-and-glucose data layer	Standardize key data elements, align timestamps, reconcile sources, and preserve provenance	Enables reliable linkage of insulin actions, glucose responses, and clinical context
3. Apply data quality controls	Ensure data are trustworthy	Evaluate conformance, completeness, plausibility, and currency; identify missing or conflicting records	Improves reliability for both clinical use and model development
4. Make the data clinically usable	Embed the data in workflow	Develop concise summaries, visualizations, and role-appropriate views for clinicians and nurses	Supports interpretation, usability, and adoption
5. Establish governance and trust mechanisms	Create the oversight structure for safe use	Define validation, transparency, human oversight, privacy/security safeguards, and monitoring processes	Makes decision support safer, more explainable, and more accountable
6. Start with narrow use cases	Match ambition to readiness	Begin with summarization, pattern detection, risk stratification, or protocol support	Reduces implementation risk and builds operational confidence
7. Evaluate with clinically meaningful measures	Assess performance in clinically relevant terms	Judge outputs against safety, glycemic priorities, workflow fit, and clinician judgment	Keeps evaluation tied to real-world usefulness rather than technical metrics alone
8. Monitor, refine, and scale	Treat deployment as a continuous capability	Track drift, recalibrate, review edge cases, incorporate feedback, and expand gradually	Supports durable, trustworthy scale-up over time

summarizes this roadmap and emphasizes that readiness for hospital insulin AI is cumulative rather than instantaneous.

Hospitals can apply this roadmap at different levels of digital maturity, using the suggested sequence as a guide from their starting point. A hospital relying primarily on paper protocols or static order sets might start with mapping insulin workflows, identifying the minimum glucose, insulin, nutrition, medication, and clinical-context data elements needed for safe review, and converting key elements into structured fields. A hospital with established EHR-based insulin orders, medication administration records, and point-of-care glucose values might focus next on timestamp alignment, provenance tracking, data-quality checks, and concise summaries that allow clinical teams to review insulin actions and glucose responses together. A more digitally mature hospital with CGM, insulin pump, AID, or glycemic-management platform data might extend this foundation by integrating device-generated data, validating ingested data, and beginning with narrow AI decision-support use cases such as pattern detection, risk stratification, protocol support, or pre-rounding summaries. Data models, detailed technical specifications, workflows, and implementation guides are being developed and published through iCoDE-2 [10,11] to support this work.

7. Conclusions

The successful application of AI to hospital insulin management depends not only on advances in model development, but on the strength of the surrounding data, workflow, and governance infrastructure. In clinician-administered hospital insulin management, many of the core data elements relevant to insulin dosing already exist within the EHR; however, their presence alone does not ensure that they are sufficiently connected, temporally coherent, clinically interpretable, or workflow-ready to support trustworthy decision support. In device-mediated insulin delivery, the challenge is more directly one of integrating externally generated diabetes data into clinical systems in standardized and usable forms. Within this landscape, iCoDE’s most direct contribution is in standardizing and facilitating the integration of device-generated diabetes data, while its broader contribution is to offer an informatics framework for thinking about representation, interoperability, visualization, and trust. Seen this way, the value of iCoDE is not that it solves hospital insulin AI on its own, but that it helps define the conditions under which such AI can become clinically meaningful, trustworthy, and scalable.