At Temple University Hospital in North Philadelphia, an artificial intelligence program called EndoTool is now fully embedded in how nurses manage insulin for hospitalized diabetic patients, and hospital leaders say it has produced a more than twofold reduction in hypoglycemia, the dangerously low blood sugar episodes that can cause seizures, confusion and, in severe cases, death.
How EndoTool Works
EndoTool analyzes a patient’s height, weight, kidney function, metabolic markers and real-time blood glucose readings to generate a personalized insulin dosing recommendation, replacing the standardized sliding-scale charts that hospitals have relied on for decades. Every recommendation still requires sign-off from a nurse or physician before insulin is administered, keeping a human in the loop, but the system continuously refines itself as it accumulates more data on how an individual patient responds, becoming more precise the longer someone stays in the hospital.
From Pilot to Hospital-Wide Standard
Temple began testing the tool in 2022 in select units before expanding it across all Temple hospitals by 2025. Ben Slovis, Temple University Hospital’s chief medical information officer, has described the health system’s approach to adopting AI tools as deliberately unhurried: “Our goal is not to chase shiny objects, but to approach these tools with thoughtful evaluation.” That multi-year rollout, rather than a rapid deployment, is part of why hospital staff say they trust the system’s recommendations. Nurses including Samantha Messick, who works in Temple’s Neuroscience ICU, have described the tool as reducing the guesswork that previously went into adjusting insulin doses for complex, critically ill patients.
Why Hypoglycemia Matters So Much
Hypoglycemia in hospitalized patients is not a minor inconvenience — it is associated with longer hospital stays, higher mortality, and increased costs, and it disproportionately affects patients with kidney disease, irregular eating schedules from procedures, or fluctuating steroid doses that make blood sugar hard to predict with a fixed dosing chart. A more-than-twofold drop in these events at Temple, a large urban safety-net hospital that treats a high volume of complex and acute cases, represents a meaningful patient-safety improvement rather than a marginal statistical gain.
Background: A Broader Wave of AI Insulin Tools
Temple’s rollout is part of a broader shift in diabetes technology in 2026. DreaMed, an FDA-approved app for people with type 1 and type 2 diabetes on multiple daily injections, uses continuous glucose monitor data to recommend insulin dose adjustments that are then reviewed by a clinician. Separately, UpDoc received the first FDA clearance for a diabetes management software that uses a patient-facing large language model, letting patients interact by voice or text. These tools, along with automated insulin delivery systems showcased at the 2026 Advanced Technologies & Treatments for Diabetes conference in Barcelona, reflect a broader move away from static, one-size-fits-all dosing charts toward continuously learning, individualized systems.
Two Perspectives on Algorithmic Dosing
Proponents, including much of Temple’s nursing staff, say the tool has made a historically stressful and error-prone part of bedside care measurably safer, since insulin dosing errors are among the most common and consequential medication mistakes in hospitals. Slovis and colleagues argue that keeping a nurse or physician in the approval loop, rather than allowing fully automated dosing, strikes the right balance between efficiency and safety.
Nursing advocates, including Maureen May, a Temple nurse and president of the Pennsylvania State Nurses Association, have raised broader concerns shared across the profession about the pace of AI adoption in hospitals — specifically that tools introduced to reduce workload can also change the nature of nursing judgment and raise questions about accountability if an AI-generated recommendation turns out to be wrong. The union has pushed for continued nurse training and clear protocols on who is responsible when an algorithmic dose recommendation is followed and something goes awry.
Looking Ahead
Temple executives say they plan to study whether EndoTool’s benefits extend to reduced length of stay and lower rates of diabetic complications after discharge, not just fewer in-hospital hypoglycemic events. As more health systems face pressure to cut costs while improving safety metrics tracked by regulators and insurers, tools like EndoTool are likely to spread to other hospitals, particularly those serving large populations of patients with diabetes and other chronic conditions that complicate standard dosing protocols.
Temple’s patient population, drawn heavily from North Philadelphia, includes a disproportionate share of patients with advanced kidney disease, food insecurity that disrupts regular eating schedules, and multiple chronic conditions requiring steroids or other medications that make blood sugar notoriously difficult to predict with a static sliding-scale chart. Hospital leaders say that context is part of why a three-year, phased rollout rather than a rapid systemwide switch was necessary: nurses needed time to build trust in the tool’s recommendations on exactly the kind of complex, high-acuity patients where dosing errors carry the greatest risk. Temple has said it is now sharing implementation lessons with other academic medical centers considering similar tools, including how it structured nurse training and built in mandatory human sign-off at every dosing step.
