Beyond Burnout: How AI is Powering the Future of Hospital Workforce Management

A charge nurse looks over the schedule for the following week. Two nurses called out on an already understaffed unit. The float pool has also been depleted. Agency staff are available but do not know the unit workflows. The nurse manager makes decisions based on who picked up the phone, who owes them a favor, and what has worked before. The process is informal and reactive. This is done several thousands of times daily across healthcare systems in the U.S.

Considering the largest operational expense for healthcare systems is labor, 50 percent or more of total costs, managing labor usage is one of the least optimized and most data poor domains in an industry with heavy investment in data infrastructure for almost every operational area. The disproportionate investment in clinical data systems and the basic, primitive status of workforce planning tools is one of the most important operational inefficiencies in healthcare.

A combination of insufficient pipeline, increasing voluntary attrition, poor and unsustainable scheduling, and operational models that view labor as a cost to be minimized, as opposed to a resource to be expanded, leads to the healthcare workforce crisis.

The 2022 Surgeon General Advisory on Health Worker Burnout listed the systematic causes of workforce burnout and attrition, citing documentation burdens, understaffing, and poor physical work conditions. Shortages in both nursing and allied health professions are expected, according to the Bureau of Labor Statistics. The American Hospital Association has stated that the current workforce shortage is the most concerning operational issue that health systems are facing.

Health workforce solutions that utilize artificial intelligence have entered the workforce in response to the current environment, and predictive staffing is one of the leading concerns. Anticipatory demand and staffing predictions are based on past data and external factors. Advanced burnout predictions and clinician matching systems have been developed and implemented.

What is most at stake is the competency of the organizations implementing these systems, and the operational capability to maintain these systems.

Predictive demand forecasting is the most advanced facet of AI applications in workforce management. Health systems are realizing improved schedule stability and decreased reliance on agency staffing after implementing predictive models of patient volume and acuity. UPMC has documented a 30 percent reduction in agency utilization after the implementation of AI-informed forecasting in staffing. LeanTaaS has provided predictive staffing models which have improved boarding time in the emergency department.

The results provided operational value but a limited scope in the methodology. The majority of published outcomes from the vendors are related to their own case studies, or single-site implementations, and are not from controlled research. The Journal of the American Medical Informatics Association has published reviews on the use of AI in operational healthcare and has noted the gap between reported outcomes achieved in controlled studies and what are the actual use cases in operational settings. There are many tools that are being developed for the use of AI in a workforce setting, however, the external validity for these tools is a concern.

Dynamic scheduling, the second most significant area of application, adds another level of complexity. Scheduling in healthcare is a complex, multi-dimensional problem, rather than a single, optimal solution. There are union contracts, seniority, clinician preference, continuity of care and service, and regulatory staffing requirements by state. An AI optimized for one dimension, such as cost, and sub-optimizing others, such as clinician preference or continuity of care, will produce a schedule that is technically optimal but operationally suboptimal, and will be rejected by the operational leaders. This is the gap between the recommendations of the model and what the organization is able to implement, and this is where most AI scheduling solutions fail.

The most complex and most ambitious application of workforce AI is the prediction of burnout for employees. The hypothesis is that a combination of scheduling patterns, EHR interactions, overtime, and sick leave will indicate when a clinician is at risk for burnout, so that clinicians do not have to be reactive and leave. This will allow HR and clinical leaders to proactively provide support before the clinician resigns.

There is potential merit in the argument. The Mayo Clinic Proceedings discuss the correlation of burnout indicators with specific patterns of workload. The data signals exist, but the real problem is what the operationalization of those signals results in.

Burnout prediction models create surveillance dynamics that most healthcare systems have been unable to address. How is access to the risk scores determined. What constitutes a case justifying an intervention. How are burnout flags protected from affecting performance evaluations, shifts, and promotions in ways the clinician did not consent to. The health systems applying these model tools have not articulated the policy that separates supportive intervention from workforce monitoring.

The ANA has articulated positions that stress openness, data usage, and protection against punitive use in data collected pertaining to workforce management. Healthcare systems that monitor and model burnout with data that is unguarded concerning clinician consent, data usage, and intervention protocols are exposed to liability that is greater than the retention value the tools are designed to provide.

Workforce AI tools share the same gap with oversight, the same way all other tools examined in this newsletter have. The procurement process focuses on functionality, integration, and pricing. Oversight, where it exists, focuses on risk, ownership, and operational control. In most healthcare systems, the first process is the one well established. The second is not.

For a predictive staffing model to be implemented, there has to be ownership in case the model poorly estimates demand, leaving a unit short-staffed. If a dynamic scheduling tool is implemented, there has to be a decision made in advance regarding patient safety concerns raised by the charge nurse and the possibility of the model recommendation being overridden. In the case of a burnout detection system, there has to be a decision made in advance regarding what the organization will or will not do with the information prior to system implementation.

These types of decisions must be made before implementation and cannot be made after the fact. Such decisions must be woven into the deployment framework guidelines designed and led by a committee of clinical operations, nursing leadership, human resources, compliance, and legal departments. Any healthcare system treating workforce AI as an IT and operations problem will face the same legacy oversight challenges impacting AI implementations in other clinical and operational areas.

AI will not solve the healthcare workforce crisis. There are structural and systemic issues with the healthcare system that are the source of the problem and AI will do nothing to alter the regulatory environment, the compensation system, the restrictive guidelines, and the poor working conditions. What is possible, if the same level of management pursued with clinical AI is implemented, is to improve the way health systems distribute their workforce. The operational benefits of fewer hours wasted, fewer burnout events, and fewer shifts that are staffed with clinicians who are not familiar with the unit will also be improvements.

The risk is the same as shown throughout this newsletter: that health systems will use the tools without the oversight structures that make the use defensible. Workforce AI that improves scheduling results while creating surveillance concerns, or that simply reduces costs while ignoring the automation of labor, does not resolve the workforce issue. It adds strain to a workforce problem.

The 2022 Surgeon General Advisory on Health Worker Burnout documents systemic workforce drivers. The Bureau of Labor Statistics provides healthcare workforce supply and demand projections. The American Hospital Association reports outline workforce as a top operational concern. Operational results from AI workforce tools have been published by UPMC and LeanTaaS. The Journal of the American Medical Informatics Association has covered AI in healthcare operational management. Research on workload and burnout has been published by Mayo Clinic Proceedings. The American Nurses Association has outlined positions regarding responsible data use in workforce management.

Christopher Hutchins
Founder & CEO, Hutchins Data Strategy Consultants