AI融入临床工作流程如何释放患者吞吐量

U.S. hospitals may be on the brink of a capacity crisis. A 2025 JAMA Network Open study found that average U.S. hospital occupancy now runs at 75%, up roughly 11 percentage points from the pre-pandemic baseline near 64%. If the trajectory holds, national adult occupancy will hit 85% by 2032, the threshold most experts associate with a functional bed shortage and material patient safety risk.

The pressure is compounded by a workforce stretched thin. Average hospital

RN turnover sits at 16.4% with a national vacancy rate of 9.6%, and each percentage point of turnover costs the average hospital roughly $289,000 a year. Fewer nurses managing more occupied beds means discharge planning, bed assignment, and care coordination fall to teams that do not have the bandwidth. This is where AI can ease the burden, not by replacing clinical judgment but by surfacing the patients ready to move, flagging bottlenecks before they form, and handing back the time staff lose to manual coordination.

To prepare, hospitals should seek new ways to accelerate patient throughput. However, the next leap in throughput will not come from another dashboard or another committee. It will come from artificial intelligence woven into the clinical workflows where discharge and placement decisions are actually made.

The data hospitals need to improve patient throughput already exists. It sits in census reports, discharge logs, referral records, and care management notes across every health system in the country.

The problem is not that the data is missing. Rather, the issue is that critical data is passive, trapped in retrospective dashboards rather than surfaced when a discharge or placement decision is being made.

For more than a decade, health systems have invested heavily in interoperability, data warehouses, and analytics platforms in the hope that better information would translate into better throughput. These investments were necessary, but they were also incomplete. Patients still wait for beds. Beds still wait for the right patient. And the report that explains exactly why arrives a week after it could have mattered. What is changing now is where the intelligence lives: not in a dashboard someone has to remember to open, but inside the clinical workflow itself, at the moment and point of care where the discharge and placement decisions are actually made.

That shift, from passive analytics to embedded artificial intelligence that delivers actions, monitoring, and performance, is already underway, and the operational benefits are starting to show up in three places that hospital leaders feel every day.

Sponsored Post

Where Will AI Deployment Benefit Payers the Most?

We are taking a look at how health insurers are using AI, defining success, and managing cybersecurity risks. Give us your opinions by completing our brief, anonymous survey.

By MedCity News

Forecasting daily discharge volume before bottlenecks form – Most health systems still manage bed capacity reactively. The morning huddle starts when the floors are already full, the emergency department (ED) is already boarding, and the perioperative schedule is already at risk. Operations leaders spend their day responding to bottlenecks that formed overnight rather than getting in front of the ones forming tomorrow.

Machine learning trained on historical discharge patterns changes that posture. By factoring in variables like time of day, day of week, seasonality, payer mix, and unit classification, predictive models can generate dynamic, continuously updated forecasts of how many patients are likely to leave each unit, each shift, and each day. The output is not a static projection produced at 7 a.m. It updates real-time as conditions change.

With visibility to capacity, that forward-looking view turns staffing, transfer coordination, and surge planning into proactive disciplines. Beds can be queued for the next patient before they are vacant. Transfers can be timed to land in the right unit at the right hour. Surge protocols can be activated based on what is coming rather than what has already arrived. That is a meaningful operational change, and it does not require a single new piece of data. It requires intelligence applied to the data already flowing through the system.

Predicting individual patient discharge readiness earlier in the stay –

Even when forecasting works at the unit level, discharge planning at the patient level is often still a downstream activity. The traditional sequence starts when a physician writes the discharge order. Only then do care  teams begin coordinating discharge plans, post-acute services, transport, durable medical equipment, family communication, and medication reconciliation. By the time those pieces fall into place, the bed has been occupied long after the clinical work was finished.

AI that monitors patient-level clinical signals throughout the stay across many patients and can flag approaching discharge readiness hours to days earlier. By watching the same indicators that experienced care teams learn to read over years of practice, such as trends in vital signs, lab values, mobility status, oxygen requirements, pain control, and order patterns, embedded models can prompt the care team when a patient is likely to be ready for discharge soon.

That head start is the difference between a discharge at 11 a.m. and one at 5 p.m. Across busy in-patient units and floors, that difference compounds. It shortens length of stay, opens beds earlier in the day, reduces ED boarding, and gives case managers the time they need to arrange the right post-acute transition at the right time, rather than the most convenient one.

ED boarding is no small consideration for hospitals, as a 2025 Health Affairs study found the practice has become “increasingly common,” with around 5% of patients admitted during peak times waiting 24 hours or longer for an inpatient bed.

Accelerating post-acute referral placements

– The third bottleneck sits at the end of the patient journey during post-acute referrals and placements. Once a referral is generated, it lands in a queue at a skilled nursing facility, an inpatient rehab, a long-term acute care hospital, or a home health agency. There, an intake clinician or admissions coordinator has to comb through a packet that can run dozens or even hundreds of pages before a yes-or-no decision is possible.

That review takes time, and time is exactly what the patient does not have. Additionally, every extra hour a referral sits in queue is another hour the upstream bed stays occupied, another hour the discharge plan remains in limbo, and another hour the receiving site is doing administrative work instead of clinical care.

AI applied to a referral packet can surface the clinical information that matters most for the acceptance decision: active diagnoses, relevant comorbidities, medication list, functional status, wound or infectious considerations, and equipment needs. The clinician still owns the determination.

The technology simply hands the clinician minutes of essential context instead of hours of document review. Across a network of post-acute partners, that compression of cycle time directly accelerates placement, frees upstream capacity, and keeps referral relationships from quietly breaking down.

From interoperability to decision support at the point of care

These three use cases share a structural feature that explains why they work. Each is specific. Each is embedded in a workflow that already exists. Each is tied to a decision that already has to be made. The AI is not asking clinicians to learn a new system or to interpret an abstract score. It is intervening at a moment when an experienced practitioner would already be reaching for the information.

That is the right standard for evaluating any AI investment in clinical operations. The question is not whether the technology is impressive. The question is whether it shows up at the right time, with the right information, inside the workflow where the decision is being made. Anything else is just another report.

Health systems have spent years building the data foundation. Throughput gains in the next decade will come from how effectively that data drives decisions at the point of care, not from how comprehensively it is visualized after the fact. The conversation must move from analytics to action.

From passive analytics to embedded intelligence

This is the version of AI that earns clinician trust. It does not interrupt. It does not gate. It does not pretend to make the call. It surfaces what the clinician would have wanted to know anyway, at the moment they need it, inside the workflow they are already in. That is how passive data finally becomes active, and that is where the next chapter of operational performance will be written.

Photo: Volha Rahalskaya, Getty Images Jonathan Shoemaker Jonathan Shoemaker joined ABOUT in 2023 as Chief Executive Officer, bringing more than 25 years of health system and information systems experience with a proven track record of transforming and delivering initiatives and solutions that improve healthcare delivery, operations, and growth.

Before joining ABOUT, Jonathan most recently was senior vice president of operations and chief integration officer as well as a member of the senior executive team leading Allina Health’s Performance Transformation Office. Before his most recent role at Allina, Shoemaker spent six years as Allina Health’s chief information officer and chief improvement officer. Prior to Jonathan’s tenure at Allina, he held leadership positions at prominent IT & healthcare firms, including NorthPoint Health and Wellness Center, BORN Consulting, and Hennepin County Medical Center.

This post appears through the MedCity Influencers program. Anyone can publish their perspective on business and innovation in healthcare on MedCity News through MedCity Influencers.

Click here to find out how