Reclaiming Time: Designing AI That Gives Back the Day

After completing rounds on fourteen patients, a hospitalist has a steep pile of inbox messages, lab results, questions from nursing, alerts from pharmacy, and prior authorization requests awaiting their attention. There is a cognitive cost of task switching that no electronic health record system metric can account for. EHR systems record clicks, but there is no record of the mental cost incurred from switching contexts for a complex discharge plan or a simple medication order due thirty times during a shift.

In the burnt-out medical profession, the cost of cognitive load can interfere with the safe care of patients. When the systems that physicians rely on demand more attention than the clinical decisions they are making, the system has become the primary consumer of the resources it was designed to support.

Cognitive load theory was developed by Sweller and colleagues in the field of educational psychology, and has been the focus of much research in the healthcare context. Studies in BMJ Quality and Safety have examined the effect of cognitive load in clinical tasks and demonstrated that with growing cognitive load, the number of errors made during completion of tasks, the time taken to complete the task, and the level of situational awareness all decrease. This is a non-linear relationship. In the context of cognitive overload, there is a point where the degradation of judgment results in catastrophic outcomes, not just slowness.

Cognitive load has been studied in EHR usability, and research funded by the American Medical Association has started to analyze these effects in the context of the EHR. Physicians in outpatient practice settings process over 4,000 data points during a single shift interacting with the EHR system. Processing those data points and interruptions in a clinical setting is borne entirely by the clinician. The system does not share the load. It creates it.

Research from the Mayo Clinic points out specific patterns of EHR workload associated with burnout. Most affected are physicians who have the greatest proportion of administrative work as compared to clinical work. The work surrounding the actual practice of medicine becomes the greatest burden.

The AI tools most relevant to the reduction of cognitive load fit into the following three categories: ambient documentation, intelligent summarization, and workflow-aware decision support. Each one attempts to address a different facet of the attention tax that clinicians carry while working.

The burden of charting and documentation, which was discussed in detail in previous editions of this newsletter, will be removed from the clinical encounter. Instead, clinicians will be able to concentrate on the patient while the AI captures the encounter and generates the note in the background. The cognitive benefit is retention of attention for the encounter, which is the most critical period for clinical decision-making.

Intelligent summarization works on the burden of information retrieval. Most physicians in day-to-day clinical work have to synthesize documents, often hundreds of pages, encompassing notes, results, and correspondence. AI summarization tools that distill a longitudinal record into a clinically relevant briefing can reduce the cognitive burden of preparation while simultaneously providing the clinician with access to the underlying documentation.

Workflow-aware decision support goes some way toward addressing the interruption burden. Traditional clinical decision support systems fire alerts without consideration of context, resulting in large notification volumes that clinicians learn to ignore. AI decision support that accounts for context, relevance, and timing can reduce notification volumes while preserving clinically relevant alerts. Publications in the Journal of the American Medical Informatics Association have shown that contextual alert systems can reduce alert fatigue and improve clinician engagement with important alerts.

Each of these tools brings with it an oversight dimension that goes past the original technical purpose. Ambient documentation entails some form of attestation. Summarization tools will need some form of validation to determine whether the summary is accurate and complete. Decision support tools will require monitoring of the amount of suppressed alerts and the resulting clinical outcomes.

The unifying element is that cognitive load mitigation tools work by filtering information before it reaches the clinician. Every filtering tool that processes clinical information is claiming a position in the clinical decision-making process. The real question is whether the organization is clear on that position, how it will assess that position, and what will happen if the filter removes an important clinical element.

Health systems that implement cognitive load mitigation tools without these oversight structures in place are swapping one burden for another. Clinician attention is preserved, but the ability of the organization to audit and explain the decision-making process is eroded. The need for performance and operational ownership to evolve in tandem is as relevant here as it is in every other instance this newsletter covers.

The majority of healthcare systems evaluate the performance of AI tools using adoption metrics, which include deployment metrics, usage numbers, and the overall duration of the user session. While these metrics state whether the tool has been utilized, they say nothing about whether the tool diminishes the cognitive burden on the clinician.

Assessing the cognitive burden requires other instruments. These include time-and-motion studies which assess documentation time before and after deployment, clinician-reported outcome measures that evaluate the burden that clinicians perceive, monitoring error rates to know whether decreased cognitive load coincides with improved clinical outcomes, and EHR usage patterns after hours to know whether the tool has shifted the burden of documentation, and to what extent.

Those organizations that commit to investing in these measurement methods will build the necessary evidence to substantiate their ongoing commitment and to ascertain when a tool fails to meet expectations. Organizations that only focus on adoption metrics will be aware of the usage of the tool, but will not see the benefit.

The American Medical Association has investigated the usability of EHRs and clinician workloads. BMJ Quality and Safety has published studies examining the relationship between cognitive load and clinical error rates. Studies on EHR workload patterns and clinician burnout have been published in the Mayo Clinic Proceedings. Context-aware clinical decision support systems and alert fatigue have been analyzed in the Journal of the American Medical Informatics Association.

Christopher Hutchins
Founder & CEO, Hutchins Data Strategy Consultants