ORION Surgical Scheduling Targets Manitoba’s Backlog With AI-Assisted Planning

University of Manitoba researcher Dr. Jay Toor is developing ORION surgical scheduling to address orthopedic backlogs, with projected capacity gains still awaiting local validation. The initiative targets a practical hospital problem: matching surgical demand with available operating time. Its central question concerns how much additional care better planning can deliver within existing resource limits.

The university outlined the project on October 9, 2026, describing staged testing at Winnipeg’s Health Sciences Centre. Researchers predict a 5% to 8% increase in orthopedic surgical capacity and a 10% to 15% reduction in wait times. Those figures describe expected benefits, rather than measured improvements from a completed Manitoba deployment.

Related work at Toronto’s Sunnybrook Health Sciences Centre provides a useful reference. Researchers there have investigated AI-assisted scheduling under the SMART-S name, alongside published work combining predictions with mathematical planning. Together, these projects put a specific proposition under scrutiny: better estimates and schedules can recover usable capacity. Hospital leaders still need evidence that those gains survive staffing constraints, daily disruptions, and clinical priorities.

ORION Surgical Scheduling Moves Into Local Validation

ORION stands for the Operating Room Intelligence and Optimization Network. According to the university’s account of its surgical scheduling research, Toor leads the initiative with support from a Research Manitoba Solutions Grant. The planned evaluation begins with historical information before moving into prospective testing, which evaluates a system as new events occur.

The first year will examine 10 years of Health Sciences Centre patient data to assess predictions of past surgical demand. The second year is intended to focus on prospective testing at the hospital. The announcement establishes a research sequence; it does not establish that AI-generated schedules already determine routine patient bookings.

The immediate change is a funded local evaluation of scheduling technology, with patient benefits remaining a hypothesis to test.

That distinction affects how readers should interpret the project’s ambition. A historical evaluation asks whether a model can make useful predictions from records. A live evaluation asks whether people can act on those predictions under actual working conditions. Success at the earlier stage supports further testing, but does not settle the later question.

For hospital managers, the distinction separates a promising planning instrument from dependable operational infrastructure. Recommendations must arrive early enough to influence decisions. They must also reflect the resources people can actually deploy, including constraints that change after a schedule has been prepared.

Consider a hypothetical operating day with an opening that appears long enough for another procedure. Filling that opening requires more than identifying a patient. The relevant team, equipment, preparation, and recovery arrangements must align. A forecast becomes useful only when the hospital can convert it into a feasible booking.

This example illustrates the implementation question without claiming to describe an ORION booking already completed. The project’s value depends on whether better information changes these decisions. A more accurate forecast that never influences scheduling would remain a research result rather than an access improvement.

The local setting matters for another reason. Hospital records describe past practices as well as patient needs. An evaluation should examine which patterns remain relevant and which reflect earlier conditions. Otherwise, a system risks learning how work used to happen without adequately supporting how it must happen next.

For ORION surgical scheduling, a credible early milestone would therefore combine prediction results with a clear account of intended operational use. Readers should look for the decisions each output supports. Predicting demand, estimating procedure time, and choosing bookings are related tasks, but each requires its own evidence.

More Procedures Do Not Automatically Mean Shorter Waits

The pressure facing hospital leaders extends beyond recovering empty minutes. Patients need access that improves consistently, while staff need schedules they can deliver safely. A planning system has to connect those goals rather than treating the operating room as an isolated calendar.

National data show why activity alone is an incomplete measure. In its priority procedure findings, the Canadian Institute for Health Information reported increased procedure volumes alongside persistent access problems. Canada performed 26% more hip replacements and 21% more knee replacements in 2024 than in 2019. Nevertheless, timely access remained below pre-pandemic levels for those procedures.

In 2024, 68% of hip replacement patients and 61% of knee replacement patients received surgery within the recommended six months. These national figures provide context, rather than a Manitoba baseline or a prediction of ORION’s performance.

Scheduling improvements should be judged by access gains alongside surgical volume, because a busier service can still leave patients waiting.

The analytical issue is the relationship between demand and completed care. Additional procedures help reduce a queue when they outpace new additions, with other factors held constant. If demand also rises, a hospital can increase activity without producing a comparable reduction in waiting.

This means the project’s capacity and wait-time projections require separate evaluation. A capacity increase describes additional work the service can perform. A wait-time reduction describes a change in patients’ experience. The connection between them depends on who enters the queue, which cases receive priority, and how the queue changes.

Hospital administrators consequently face a measurement problem as well as a scheduling problem. They need to explain what an improvement means. Does it represent more completed operations, fewer cancellations, shorter waits, less overtime, or a combination? Each outcome answers a different public concern.

The distinction also matters for clinicians deciding whether to support a new process. A fuller schedule might help their patients receive treatment sooner. It might also create additional coordination demands. Evaluation should make those effects visible rather than letting a single activity measure stand for overall success.

For patients, averages alone would offer limited reassurance. A service can improve its typical wait while leaving some people waiting much longer. Any future results should therefore explain how benefits are distributed, including whether patients with more complicated needs experience comparable improvements.

The pressure falls directly on hospital leadership to translate a technical experiment into an understandable service outcome. Researchers can report prediction accuracy, but administrators must decide whether the resulting process deserves sustained use. That decision requires evidence people outside the modeling team can interpret.

The strongest public account would connect each stage: a better forecast changed a booking decision, the booking became a completed operation, and access improved. Without that connection, claims about efficiency risk stopping before the outcome patients care about most.

Better Time Estimates Change Which Cases Fit

The technical precedent is more specific than a general claim that AI can improve healthcare. A peer-reviewed study published on January 23, 2026, examined scheduling for arthroplasty, surgery that replaces or reconstructs a joint. Its authors included Toor and researchers associated with Sunnybrook and the University of Toronto.

The study combined machine learning, which learns predictive patterns from data, with mathematical optimization, which selects arrangements under defined constraints. Its scheduling study compared predicted procedure durations and generated schedules against an approach using rolling average times.

Researchers reported that the generated schedules reduced unused operating-room time by 56.2%, equivalent to 13.3 minutes per day. Overtime increased by 17.2%, equivalent to 3.6 minutes per day. These results belong to that study’s evaluation and should not be presented as completed Manitoba outcomes.

The published finding supports a scheduling mechanism while exposing its tradeoff: less unused time came with additional overtime.

Prediction and scheduling contribute differently to that mechanism. A prediction estimates how long a case will take. A scheduling process then decides how cases should fit together. Even a useful estimate does not specify which patient should receive a particular slot or how much uncertainty staff should accept.

A hypothetical example makes the distinction concrete. Suppose an existing estimate leaves a gap that appears too small for another case. A more accurate estimate might reveal a feasible booking. Alternatively, it might show that the original schedule was already too ambitious and should contain fewer cases.

Both outcomes can represent better planning. The value lies in making decisions that match actual conditions, rather than always producing a fuller calendar. An evaluation focused solely on additional bookings could miss improvements that prevent unrealistic commitments.

The overtime result deserves the same attention as the reduction in unused time. Relative percentages describe changes from a baseline, while minutes show their absolute size. Readers need both to understand the operational significance. Neither figure alone determines whether the tradeoff would be acceptable elsewhere.

A hospital might accept additional overtime under some circumstances and reject it under others. That choice depends on local priorities and constraints. It should remain explicit, especially when an algorithm can produce different schedules depending on the objectives it receives.

The practical question for ORION surgical scheduling is therefore how its recommendations will balance competing demands. Filling time, limiting overruns, accommodating clinical urgency, and maintaining workable schedules do not always point toward the same arrangement.

The study offers a reason to investigate that balance, rather than a universal answer. Its findings support the proposition that better estimates can improve scheduling decisions. They do not establish that every hospital will gain the same usable capacity or that every recovered minute becomes another completed procedure.

Sunnybrook Provides a Comparison, Not a Manitoba Outcome

Sunnybrook offers a relevant institutional comparison because its researchers are pursuing closely related scheduling work. Its 2026 program account describes SMART-S as a system that predicts elective orthopedic procedure times and uses those estimates to support scheduling.

The hospital foundation’s SMART-S program update says a team led by Dr. Cari Whyne is comparing performance with the existing manual process. The account also identifies forecasting trauma and urgent case volumes as an aim. This places the work within an ongoing evaluation program.

The useful comparison concerns how different sites validate AI-assisted planning, rather than which project makes the larger promise.

Manual scheduling supplies an essential reference because a new system must improve the process people already use. That comparison should be fair. A model evaluated against an oversimplified description of current practice would provide less useful evidence than one tested against actual operational decisions.

For hospital buyers, the distinction changes the procurement conversation. A polished demonstration can show that software generates schedules. It cannot establish that those schedules outperform experienced coordinators working with the same information. That requires a comparison designed around the real service.

The relationship between SMART-S and ORION also calls for careful reporting. Shared research interests and overlapping collaborators provide relevant context. They do not justify assuming that the systems have identical software, inputs, implementation arrangements, or results.

A local evaluation should explain which components carry over and which require adaptation. Some transferable value might come from methods or lessons rather than an unchanged model. That would still represent useful progress, provided the account makes the boundaries clear.

The broader technology lesson is that operational AI needs a defined role. Hospitals should be able to state whether software estimates durations, proposes schedules, forecasts incoming demand, or supports another decision. Describing all those functions simply as AI conceals the work necessary to evaluate each one.

For developers, the demanding interface is between a recommendation and the person who must act on it. Useful outputs need enough explanation for staff to assess feasibility. They also need a way to record why an apparently reasonable recommendation was unsuitable.

That feedback can inform future development without treating every disagreement as user resistance. A coordinator may know about a constraint missing from the data. Alternatively, a recommendation may reveal an assumption worth reconsidering. The implementation needs room for both possibilities.

Sunnybrook’s ongoing comparison therefore gives Manitoba a meaningful reference point without supplying an advance verdict. The strongest evidence would show how each local workflow changes, which outcomes improve, and where the technology remains unable to help.

Staff Adoption and Fair Access Belong in the Evaluation

Evidence about implementation adds a perspective distinct from model performance. A July 2026 qualitative study examined preparations for machine-learning-enabled scheduling at a Canadian orthopedic center. Its participants included 17 people working in clinical, administrative, and leadership roles.

The scheduling implementation study identified concerns about fragmented information, existing processes, and resource constraints. Participants expressed varying support for the proposed technology while emphasizing workflow fit, training, usable design, and phased implementation. These findings describe perceptions before implementation, rather than measured patient benefits.

Staff acceptance is part of operational performance, and a system that adds work can undermine its own scheduling gains.

This is a practical issue rather than a referendum on whether clinicians like AI. Staff must understand when to consult a recommendation, who can change it, and how changes reach everyone affected. Ambiguity at those points can create additional coordination work.

The study’s findings support examining those questions early. They do not show that ORION has encountered the same barriers. For Manitoba, the appropriate inference is that implementation deserves explicit evaluation rather than an assumption that useful predictions will automatically enter routine practice.

Fair access introduces another design question. A scheduling objective expresses what the system tries to improve. If that objective emphasizes completed cases, evaluators should examine whether it creates incentives that conflict with clinical priorities. This is a risk to test, rather than an established finding about ORION.

For example, an evaluation could examine whether patients needing longer procedures benefit from the new process. It could also track cases repeatedly moved when schedules change. These checks would help distinguish a service that improves broadly from one whose headline gains conceal uneven effects.

Clinical judgment needs a clear place in that process. A recommendation should not become harder to challenge merely because software generated it. Staff should have a practical route to override unsuitable suggestions and record the operational or clinical reason.

The resulting records could become useful evaluation material. Frequent overrides might indicate missing information, poorly chosen objectives, or constraints that change faster than the system can accommodate. Occasional overrides might simply reflect appropriate human responsibility for unusual situations.

Privacy questions also remain relevant because the planned historical evaluation involves patient information. Public descriptions should explain how researchers govern access and use. The available announcement does not provide enough detail to assess every technical or organizational safeguard.

A historical dataset is an opportunity for analysis, but its size does not settle whether the information supports the intended decisions. Evaluators should examine completeness, consistency, and relevance. More records cannot compensate for an important constraint that was never recorded.

The strongest skeptical position is therefore specific: prediction accuracy alone is insufficient evidence of service improvement. ORION needs evaluation across usability, feasible decisions, patient access, and staff workload. That standard allows promising results to receive credit while keeping untested benefits clearly identified.

Watch the Evaluation Design Before Expecting Backlog Results

The next meaningful developments concern how researchers define success and move toward testing. Over the coming months, readers should distinguish announcements about methods from evidence about patient outcomes. The university’s staged plan does not justify expecting a completed wait-time evaluation immediately.

The first signal to watch is a clearer evaluation design. Useful information would include the baseline, the decisions ORION supports, and the outcomes researchers intend to measure. A specific design would make later claims easier to assess and reduce ambiguity around the projected benefits.

That information should also distinguish historical testing from live use. A retrospective result, produced from past records, can strengthen the case for prospective evaluation. It cannot independently establish how staff and patients respond when recommendations influence actual scheduling.

The second signal is evidence of operational integration. Readers should look for an explanation of who receives recommendations, when they arrive, and how staff review them. A defined workflow would strengthen the proposition that prediction results can influence care delivery.

By contrast, a presentation limited to model accuracy would leave the implementation question open. That would not make the research worthless. It would mean the available evidence addresses a narrower claim than improved hospital capacity.

The third signal is balanced reporting of outcomes and tradeoffs. When live results become available, completed procedures and waiting times should appear alongside cancellations, overtime, and relevant measures of access. Gains accompanied by new burdens need to be understood together.

Independent context helps explain why that balance matters. CIHI’s joint replacement indicator measures the share of planned replacements completed within 26 weeks, or 182 days, from readiness for treatment. A clearly defined measure makes comparisons interpretable.

That definition also illustrates a reporting obligation for any future Manitoba results. Researchers should identify when a waiting period begins and which patients enter the calculation. Changing the definition can change the apparent outcome without changing a patient’s experience.

The evidence will become more persuasive when better predictions produce feasible schedules and measurable access improvements without hiding operational costs.

These signals would strengthen the project’s case in sequence. A transparent design would clarify the claim. A workable process would show that people can use the system. Balanced outcomes would establish whether that use delivers benefits worth sustaining.

The reverse would also be informative. Unclear baselines would weaken comparisons, persistent workflow problems would limit usefulness, and uneven outcomes would require changes to the approach. Reporting such findings would contribute more than treating every development as confirmation.

For AI developers and enterprise buyers, ORION surgical scheduling offers a practical example of what operational evidence should contain. The relevant question extends beyond whether a model predicts accurately. It includes whether the organization can turn that prediction into a better decision and verify the consequences.

Follow the project’s evaluation milestones before treating projected wait-time reductions as established benefits. Ask which decisions changed, which patients gained access sooner, and what happened to cancellations and staff workload. When results appear, compare the definitions and baseline before comparing the headline percentages. Keep historical testing separate from prospective findings, and examine how researchers explain cases where recommendations were rejected. Manitoba’s initiative deserves attention because it addresses a concrete service problem with a testable approach. Its contribution will become clearer through transparent reporting of useful gains, practical limits, and necessary adjustments. What evidence would convince patients and hospital staff that the new scheduling process makes care more accessible?