Case Studies
Mayo Clinic Cuts Cardiovascular Duration Prediction Error by ~50% With AI
Mayo Clinic Cuts Cardiovascular Duration Prediction Error by ~50% With AI
Mayo Clinic Cuts Cardiovascular Duration Prediction Error by ~50% With AI
In a real-world evaluation of 643 cardiovascular procedures, an Opmed deep learning model cut average duration prediction error from approximately 68 minutes to 34 minutes, outperforming clinician estimates.
~50%
~50%
~50%
~50%
Reduction in duration prediction error
Reduction in duration prediction error
Reduction in duration prediction error
34 min
34 min
34 min
34 min
Average prediction error, down from ~68 minutes
Average prediction error, down from ~68 minutes
Average prediction error, down from ~68 minutes
643
643
643
643
Cardiovascular procedures in the evaluation
Cardiovascular procedures in the evaluation
Cardiovascular procedures in the evaluation
Cardiovascular procedure schedules are difficult to get right
Cardiovascular procedure schedules are difficult to get right
Cardiovascular procedure schedules are difficult to get right
Cases can involve multiple interventions, complex patient conditions, and clinical decisions made during the procedure itself. All of these factors can change how long a case actually takes.
Cases can involve multiple interventions, complex patient conditions, and clinical decisions made during the procedure itself. All of these factors can change how long a case actually takes.
Cases can involve multiple interventions, complex patient conditions, and clinical decisions made during the procedure itself. All of these factors can change how long a case actually takes.
When scheduled time does not match reality, the effects extend beyond a single procedure. One case running long can delay the next patient, shift staffing needs, disrupt room availability, and force teams to rebuild the schedule throughout the day.
When scheduled time does not match reality, the effects extend beyond a single procedure. One case running long can delay the next patient, shift staffing needs, disrupt room availability, and force teams to rebuild the schedule throughout the day.
When scheduled time does not match reality, the effects extend beyond a single procedure. One case running long can delay the next patient, shift staffing needs, disrupt room availability, and force teams to rebuild the schedule throughout the day.
To protect against that uncertainty, procedure times are often intentionally overestimated. That creates a different problem: valuable room time goes unused, fewer procedures can be scheduled, and patient access becomes constrained.
To protect against that uncertainty, procedure times are often intentionally overestimated. That creates a different problem: valuable room time goes unused, fewer procedures can be scheduled, and patient access becomes constrained.
To protect against that uncertainty, procedure times are often intentionally overestimated. That creates a different problem: valuable room time goes unused, fewer procedures can be scheduled, and patient access becomes constrained.
Mayo Clinic and Opmed set out to examine a fundamental scheduling question: can procedure duration be predicted more accurately without sacrificing operational reliability?
The Challenge
Where traditional duration estimates fall short
Where traditional duration estimates fall short
Where traditional duration estimates fall short
Procedure schedules have traditionally relied on clinician estimates and historical averages. But cardiovascular procedures are not average.
Procedure schedules have traditionally relied on clinician estimates and historical averages. But cardiovascular procedures are not average.
Procedure schedules have traditionally relied on clinician estimates and historical averages. But cardiovascular procedures are not average.
Duration can vary significantly based on the procedure being performed, the patient, clinical complexity, and what happens once the case begins. That variability makes it difficult for static estimates to consistently reflect how long an individual procedure will actually take.
Duration can vary significantly based on the procedure being performed, the patient, clinical complexity, and what happens once the case begins. That variability makes it difficult for static estimates to consistently reflect how long an individual procedure will actually take.
Duration can vary significantly based on the procedure being performed, the patient, clinical complexity, and what happens once the case begins. That variability makes it difficult for static estimates to consistently reflect how long an individual procedure will actually take.
The result can be:
The result can be:
The result can be:
Procedures running beyond their scheduled time
Procedures running beyond their scheduled time
Procedures running beyond their scheduled time
Delays cascading across the procedural day
Delays cascading across the procedural day
Delays cascading across the procedural day
Uneven room and staff utilization
Uneven room and staff utilization
Uneven room and staff utilization
Frequent day-of schedule adjustments
Frequent day-of schedule adjustments
Frequent day-of schedule adjustments
Unused capacity when cases are overestimated
Unused capacity when cases are overestimated
Unused capacity when cases are overestimated
The challenge was particularly pronounced for longer, less frequent, and more variable procedures, where an inaccurate estimate can have an outsized effect on the rest of the schedule.
The Approach
Moving from averages to case-level predictions
Moving from averages to case-level predictions
Moving from averages to case-level predictions
In collaboration with Mayo Clinic, Opmed evaluated whether AI could provide a more accurate foundation for cardiovascular procedure scheduling.
In collaboration with Mayo Clinic, Opmed evaluated whether AI could provide a more accurate foundation for cardiovascular procedure scheduling.
In collaboration with Mayo Clinic, Opmed evaluated whether AI could provide a more accurate foundation for cardiovascular procedure scheduling.
Opmed developed a deep learning model using routinely collected clinical and procedural data to predict the duration of individual cardiovascular procedures. Inputs evaluated included:
Opmed developed a deep learning model using routinely collected clinical and procedural data to predict the duration of individual cardiovascular procedures. Inputs evaluated included:
Opmed developed a deep learning model using routinely collected clinical and procedural data to predict the duration of individual cardiovascular procedures. Inputs evaluated included:
Procedure type
Procedure type
Procedure type
Patient characteristics
Patient characteristics
Patient characteristics
Comorbidities
Comorbidities
Comorbidities
ASA classification
ASA classification
ASA classification
Surgeon-level variation
Surgeon-level variation
Surgeon-level variation
Clinical notes and ECG data in select model configurations
Clinical notes and ECG data in select model configurations
Clinical notes and ECG data in select model configurations
Rather than assigning time based primarily on averages, the model produces a case-specific duration prediction based on the characteristics of the procedure and patient.
Rather than assigning time based primarily on averages, the model produces a case-specific duration prediction based on the characteristics of the procedure and patient.
Rather than assigning time based primarily on averages, the model produces a case-specific duration prediction based on the characteristics of the procedure and patient.
The goal is straightforward: give scheduling teams a more accurate picture of the day before the day begins.
The goal is straightforward: give scheduling teams a more accurate picture of the day before the day begins.
The goal is straightforward: give scheduling teams a more accurate picture of the day before the day begins.
The Results
~50% lower prediction error
~50% lower prediction error
~50% lower prediction error
In a real-world evaluation at Mayo Clinic involving 643 cardiovascular procedures, AI-based duration predictions significantly outperformed clinician estimates.
In a real-world evaluation at Mayo Clinic involving 643 cardiovascular procedures, AI-based duration predictions significantly outperformed clinician estimates.
In a real-world evaluation at Mayo Clinic involving 643 cardiovascular procedures, AI-based duration predictions significantly outperformed clinician estimates.
Measure
Measure
Measure
Clinician estimates
Clinician estimates
Clinician estimates
Best-performing AI model
Best-performing AI model
Best-performing AI model
Mean absolute error
Mean absolute error
Mean absolute error
1.13 hours (~68 minutes)
1.13 hours (~68 minutes)
1.13 hours (~68 minutes)
0.564 hours (~34 minutes)
0.564 hours (~34 minutes)
0.564 hours (~34 minutes)
R²
R²
R²
0.31
0.31
0.31
0.721
0.721
0.721
~50%
~50%
Reduction in prediction error
Reduction in prediction error
Reduction in prediction error
Average prediction error fell from approximately 68 minutes to approximately 34 minutes.
Average prediction error fell from approximately 68 minutes to approximately 34 minutes.
Average prediction error fell from approximately 68 minutes to approximately 34 minutes.
Strongest gains
Strongest gains
In high-variability cases
In high-variability cases
In high-variability cases
Improvements were particularly meaningful for longer procedures, less frequently performed procedures, and cases with greater duration variability. These are also the cases that create the greatest downstream disruption when their scheduled duration is wrong.
Improvements were particularly meaningful for longer procedures, less frequently performed procedures, and cases with greater duration variability. These are also the cases that create the greatest downstream disruption when their scheduled duration is wrong.
Improvements were particularly meaningful for longer procedures, less frequently performed procedures, and cases with greater duration variability. These are also the cases that create the greatest downstream disruption when their scheduled duration is wrong.
More accurate predictions give teams a better signal before those cases ever reach the procedure room.
More accurate predictions give teams a better signal before those cases ever reach the procedure room.
More accurate predictions give teams a better signal before those cases ever reach the procedure room.
Operational Impact
What better predictions change
What better predictions change
What better predictions change
Prediction accuracy is not the end goal.
Prediction accuracy is not the end goal.
Prediction accuracy is not the end goal.
The value comes from what teams can do with a more accurate understanding of how the day is likely to unfold.
The value comes from what teams can do with a more accurate understanding of how the day is likely to unfold.
The value comes from what teams can do with a more accurate understanding of how the day is likely to unfold.
Build schedules that better reflect reality
Build schedules that better reflect reality
More accurate case durations give schedulers a stronger foundation for determining how much time procedures actually require.
More accurate case durations give schedulers a stronger foundation for determining how much time procedures actually require.
More accurate case durations give schedulers a stronger foundation for determining how much time procedures actually require.
Reduce the risk of cascading delays
Reduce the risk of cascading delays
When scheduled durations more closely reflect actual procedure times, teams can reduce the likelihood that one underestimated case disrupts everything that follows.
When scheduled durations more closely reflect actual procedure times, teams can reduce the likelihood that one underestimated case disrupts everything that follows.
When scheduled durations more closely reflect actual procedure times, teams can reduce the likelihood that one underestimated case disrupts everything that follows.
Make better use of procedure room capacity
Make better use of procedure room capacity
Reducing unnecessary padding can uncover time that would otherwise remain unavailable for additional procedures.
Reducing unnecessary padding can uncover time that would otherwise remain unavailable for additional procedures.
Reducing unnecessary padding can uncover time that would otherwise remain unavailable for additional procedures.
Improve patient flow
Improve patient flow
A more predictable schedule can reduce downstream waiting and create a more reliable experience for patients moving through the procedural day.
A more predictable schedule can reduce downstream waiting and create a more reliable experience for patients moving through the procedural day.
A more predictable schedule can reduce downstream waiting and create a more reliable experience for patients moving through the procedural day.
Reduce day-of operational friction
Reduce day-of operational friction
When the starting schedule is more accurate, clinical and operational teams have fewer avoidable disruptions to manage throughout the day.
When the starting schedule is more accurate, clinical and operational teams have fewer avoidable disruptions to manage throughout the day.
When the starting schedule is more accurate, clinical and operational teams have fewer avoidable disruptions to manage throughout the day.
Inside the model
What drives procedure duration
What drives procedure duration
What drives procedure duration
The research also provided insight into which factors contribute most to prediction accuracy.
The research also provided insight into which factors contribute most to prediction accuracy.
The research also provided insight into which factors contribute most to prediction accuracy.
Among the strongest signals were procedure type, patient condition, and comorbidities.
Among the strongest signals were procedure type, patient condition, and comorbidities.
Among the strongest signals were procedure type, patient condition, and comorbidities.
Clinical notes and ECG data provided additional predictive value in some model configurations, but much of the signal was already present in routinely available clinical and procedural information.
Clinical notes and ECG data provided additional predictive value in some model configurations, but much of the signal was already present in routinely available clinical and procedural information.
Clinical notes and ECG data provided additional predictive value in some model configurations, but much of the signal was already present in routinely available clinical and procedural information.
There will always be factors that cannot be fully predicted before a procedure begins, including intraoperative decisions and variations in clinical workflow. The objective is not to eliminate uncertainty. It is to give teams a better starting point for managing it.
There will always be factors that cannot be fully predicted before a procedure begins, including intraoperative decisions and variations in clinical workflow. The objective is not to eliminate uncertainty. It is to give teams a better starting point for managing it.
There will always be factors that cannot be fully predicted before a procedure begins, including intraoperative decisions and variations in clinical workflow. The objective is not to eliminate uncertainty. It is to give teams a better starting point for managing it.
The Bigger Picture
A better signal for a better schedule
A better signal for a better schedule
A better signal for a better schedule
For Mayo Clinic, AI reduced cardiovascular procedure duration prediction error by approximately 50%, creating a more accurate foundation for scheduling.
For Mayo Clinic, AI reduced cardiovascular procedure duration prediction error by approximately 50%, creating a more accurate foundation for scheduling.
For Mayo Clinic, AI reduced cardiovascular procedure duration prediction error by approximately 50%, creating a more accurate foundation for scheduling.
But duration prediction is only one part of the operational equation. Once health systems have a clearer view of how long cases are likely to take, that intelligence can inform the decisions that follow: how cases are sequenced, how rooms and staff are allocated, where capacity exists, and how the schedule should adjust as conditions change.
But duration prediction is only one part of the operational equation. Once health systems have a clearer view of how long cases are likely to take, that intelligence can inform the decisions that follow: how cases are sequenced, how rooms and staff are allocated, where capacity exists, and how the schedule should adjust as conditions change.
But duration prediction is only one part of the operational equation. Once health systems have a clearer view of how long cases are likely to take, that intelligence can inform the decisions that follow: how cases are sequenced, how rooms and staff are allocated, where capacity exists, and how the schedule should adjust as conditions change.
That is where prediction becomes operational.
That is where prediction becomes operational.
That is where prediction becomes operational.
Forecast
Forecast
Predict what is likely to happen, including how long each case will take.
Predict what is likely to happen, including how long each case will take.
Predict what is likely to happen, including how long each case will take.
Allocate
Allocate
Deploy rooms, staff, and capacity around those predicted needs.
Deploy rooms, staff, and capacity around those predicted needs.
Deploy rooms, staff, and capacity around those predicted needs.
Adjust
Adjust
Adapt the schedule as conditions change through the day.
Adapt the schedule as conditions change through the day.
Adapt the schedule as conditions change through the day.
Better signals. Better decisions. More resilient operations.
Better signals. Better decisions. More resilient operations.
Better signals. Better decisions. More resilient operations.