You cannot take away the human aspect of care
At Temple Health in Philadelphia, an artificial intelligence system is quietly reshaping how hospitals manage one of medicine's most delicate balancing acts — keeping blood sugar stable in patients whose bodies, disrupted by illness and unfamiliar routines, can swing toward crisis without warning. The tool, EndoTool Sub-Q, has cut dangerous low blood sugar episodes by more than half since its full deployment in 2025, not by replacing clinical judgment but by shifting insulin dosing from reaction to anticipation. This development arrives at a moment when federal regulators are beginning to hold hospitals financially accountable for diabetes complications, making the question of how much to trust algorithms in the care of vulnerable patients both urgent and unavoidable.
- Hospitalized diabetic patients face life-threatening blood sugar swings — seizures, kidney failure, death — when standard fixed-formula insulin protocols fail to account for each person's unique physiology.
- Temple Health's EndoTool Sub-Q has slashed hypoglycemia cases by more than half by learning individual patient characteristics over time and generating personalized dosing recommendations rather than relying on one-size-fits-all sliding scales.
- Federal regulators at CMS have begun requiring hospitals to report diabetes complication data and can impose financial penalties, turning AI-assisted glucose management from a clinical nicety into a strategic and economic necessity.
- Temple's rollout is deliberate — pilot programs first, AI strategy reviewed every three months — while peer institutions like St. Luke's, Jefferson Health, and Penn Medicine are deploying their own AI tools across cardiac, administrative, and clinical domains.
- Nurse advocates warn that algorithmic dependence risks hollowing out the intuitive clinical thinking that no dashboard can replicate, and they insist that every AI recommendation must still pass through a human hand before reaching a patient.
Temple Health has deployed an AI system that cuts dangerous blood sugar crashes in half — a result that reflects both the genuine promise of machine learning in medicine and the unresolved tensions it introduces.
The tool, EndoTool Sub-Q, analyzes each patient's height, weight, metabolism, and kidney function to generate customized insulin dosing recommendations, which nurses and doctors must review and approve before any dose is given. Temple began piloting the system in 2022 and completed its rollout across all hospitals in 2025. The shift it represents is fundamental: rather than relying on a fixed sliding-scale formula that responds only to a patient's current blood sugar reading, the AI learns from each individual over time, growing more attuned to their particular physiology. For patients in the hospital — eating little, far from their normal routines, with blood sugar that can swing unpredictably — that proactive approach has made a measurable difference.
The timing is not accidental. Federal regulators at the Centers for Medicare and Medicaid Services now require hospitals to report detailed data on blood sugar complications and can impose financial penalties on those that fall short. That pressure has made AI solutions attractive not only as instruments of better care, but as tools for protecting a hospital's metrics and financial standing.
Temple's chief medical information officer, Ben Slovis, has emphasized a cautious philosophy: pilot first, evaluate rigorously, review the institution's AI strategy every three months. Other regional health systems are moving in parallel directions, deploying AI to predict cardiac arrest, reduce unnecessary ICU transfers, and automate clinical documentation.
But the expansion has drawn scrutiny. Nurse advocates, including PASNAP president Maureen May, argue that over-reliance on algorithms risks eroding the intuitive clinical judgment nurses must exercise when a patient's condition changes faster than any system can model. EndoTool's built-in human review requirement is designed to preserve that check — but whether it is sufficient, as AI spreads deeper into hospital care, remains an open and consequential question.
Temple Health has deployed an artificial intelligence system that cuts dangerous blood sugar crashes in half, a development that reflects both the promise and the lingering uncertainty surrounding AI's role in modern medicine.
The tool, called EndoTool Sub-Q, analyzes patient data—height, weight, metabolism, kidney function, and other characteristics—to generate customized insulin dosing recommendations that nurses and doctors then review and approve before administration. Since Temple began testing the system in 2022 and completed its rollout across all hospitals in 2025, the health system has seen more than a twofold reduction in cases of hypoglycemia, the dangerous condition in which blood sugar drops to life-threatening levels. Hyperglycemia, the opposite problem, poses equal risk. Both can trigger seizures, kidney failure, and death if left unchecked.
What makes EndoTool different from the standard approach is its shift from reaction to anticipation. Traditionally, doctors and nurses rely on a sliding scale—a fixed formula that tells them how much insulin to give based on a patient's current blood sugar reading. The AI tool instead learns from each patient over time, becoming more attuned to their individual physiology and needs. A neuroscience intensive care unit nurse at Temple named Samantha Messick described the difference as moving insulin dosing from reactive to proactive, a distinction that matters most in the hospital setting, where patients are often away from their normal routines, eating little or nothing, and their blood sugar can swing unpredictably.
The timing of Temple's investment in AI diabetes management is not coincidental. As of this year, federal regulators at the Centers for Medicare and Medicaid Services have begun requiring hospitals to report detailed data on how many patients experience dangerous blood sugar swings. Hospitals that fail to prevent these complications face financial penalties. That regulatory pressure has made AI solutions suddenly attractive not just as tools for better care, but as a way to improve the metrics that determine a hospital's financial standing and reputation.
Temple's approach to rolling out EndoTool reflects a deliberate caution. Ben Slovis, the chief medical information officer at Temple University Hospital, said the institution's goal is not to chase trendy technology but to evaluate new tools thoughtfully. The system tests AI applications in small pilot programs first and reviews its overall AI strategy every three months. Other health systems are moving in similar directions: St. Luke's Health Network uses AI to predict cardiac arrest and reduce unnecessary ICU transfers, while Jefferson Health and Penn Medicine have deployed ambient listening tools that record doctor-patient conversations and automatically generate visit notes.
Yet the expansion of AI in hospitals has not gone unchallenged. Nurse advocates have begun raising concerns about whether these tools deliver the benefits they promise, and more fundamentally, whether increased reliance on algorithms might erode the critical thinking skills that nurses depend on when they must respond quickly and intuitively to a patient's changing condition. Maureen May, a nurse at Temple and president of PASNAP, the region's leading nurses union, emphasized that artificial intelligence can provide useful algorithms and tools, but the human eye remains irreplaceable. "You cannot take away the human aspect of care," she said.
EndoTool's design includes a safeguard meant to preserve that human element: a nurse or doctor must review and approve every insulin dose recommendation before it reaches the patient. The system learns and adapts the longer it is used, but it does not act independently. Whether that check is sufficient to address concerns about over-automation in healthcare remains an open question as more hospitals adopt similar tools.
Citações Notáveis
Our goal is not to chase shiny objects, but to approach these tools with thoughtful evaluation— Ben Slovis, chief medical information officer at Temple University Hospital
It takes insulin dosing from reactive to proactive— Samantha Messick, neuroscience intensive care unit nurse at Temple
A Conversa do Hearth Outra perspectiva sobre a história
Why does blood sugar management become so much harder inside a hospital?
Patients lose their normal rhythm. They're not eating regular meals, or eating at all if they're very sick. Their activity drops. All the things that keep blood sugar stable in daily life disappear, and the body responds unpredictably.
So the AI is essentially learning each person's unique metabolism?
Exactly. It's not applying a one-size-fits-all formula. It's building a profile—height, weight, kidney function, how their body processes insulin—and updating that profile as it watches how they respond.
The nurses still have to approve each dose, though. So where's the actual AI doing the work?
In the anticipation. A nurse using the old sliding scale reacts to what the blood sugar is right now. EndoTool predicts where it's heading and suggests a dose before the patient crashes. The human still makes the final call, but they're making it with better information.
What worries the nurses most?
That we'll start trusting the algorithm more than our own judgment. That we'll stop thinking critically because the tool is there. In an emergency, you can't wait for a recommendation. You have to know what to do.
Is that a real risk, or are they being cautious?
Both. The tool is new enough that we don't have years of data yet. And yes, there's always a risk when you hand part of your decision-making to a machine. But Temple is being careful—testing in small groups first, reviewing the strategy every quarter. They're not rushing.
What happens if the AI gets it wrong?
The nurse catches it. That's the whole point of keeping humans in the loop. But it also means the tool is only as good as the people using it.