The AI That Heals the System

A healthcare system spends 18 months implementing an AI-powered sepsis prediction model. The model performs well during validation. Alerts are created in the EHR. Clinicians engage with it sporadically. A year later, the organization evaluates the impact of the model and finds a small decrease in sepsis mortality for the units where the model was adopted the most. The leadership team celebrates this achievement. However, no one poses the more challenging question: what if the same resources had been invested in the systemic barriers that allowed undetected sepsis to develop?

There is a focus on deploying AI in the healthcare sector on clinical use cases that operate at the point of care. Predictive models, diagnostic assistants, and documentation tools work at the interface of the clinician and the patient. These tools certainly add value. However, what is less discussed is that these tools address a system that is producing preventable harm through its design.

Healthcare systems harm patients and that harm is a result of systems that are structural and not clinical. Delays in patient discharges create boarding in the emergency room. This increases the wait time, and increases the likelihood that a patient will leave without being seen. Patients that leave without being seen return to the hospital in a worse condition. The cycle just continues to repeat itself.

A study from the Annals of Emergency Medicine has shown the link between boarding in the Emergency Department and negative patient outcomes, including increased mortality for patients who are admitted from the emergency department. This occurs during periods of high occupancy. The issue is not that the clinicians are poorly deciding. The issue is that the system is creating conditions that good decision making is difficult.

The National Academy of Medicine estimates the US healthcare system wastes somewhere between $760 billion and $935 billion each year. This waste is spread across the various categories of administrative complexity, pricing, failure in care delivery, and overtreatment. This waste is not random; it is a product of the system itself, the processes of which are identifiable and measurable, and for which effective analytical tools are readily available.

Clinical AI works at the level of the individual encounter. System-level AI works at the level of the process. This distinction is important because the most preventable harm in health care is caused by process problems, not encounter problems.

A health system that uses AI to predict which patients are at risk for readmission is answering a clinical question. A health system that uses AI to analyze why its discharge process is perpetually delayed and leads to readmissions is answering a system question. The clinical question results in an alert notification for a clinician. The system question results in a redesign opportunity for organizational leadership.

Studies done by The Dartmouth Institute analyzed healthcare service usage done by patients across various systems and locations and most of the variation was due to supply and demand system factors, such as practices and local cultures, rather than the clinical needs of patients. AI systems that explain and document variability give decision-makers access to evidence that was, until now, only available through retrospective analysis done months or years after the decision was implemented.

The most immediate impact of system-level AI will be felt in operational functions that have a ripple effect throughout the entire organization. Capacity management, staff scheduling, supply chain management, and coordination of patient flow are all areas where process inefficiencies create clinical challenges.

The American Hospital Association reports that hospitals operating at more than 85 percent capacity experience increases in adverse events, length of stay, and staff turnover. AI systems that are able to anticipate capacity limits hours or days prior to the critical point allow operational staff to take preventive actions before the cascading effect begins.

Research published in Health Affairs shows that health systems that have implemented real-time predictive capabilities in patient flow management have seen shorter ED boarding times, improved bed turnover, and fewer elective procedure cancellations. Though these may not be clinical outcomes in the traditional sense, they are operational outcomes that create the environment that allows clinical outcomes to be better.

Health system leaders typically do not face this question explicitly when planning investments; should the next dollar of AI spend go to a clinician support decision tool that can help a clinician make a better decision in a compromised environment, or a system vertical tool that can help a clinician make a decision in a less compromised decision environment?

The question is not about investing in one option or the other; it is about the proportions of investment. Most health systems have made a large investment in the former, while the latter is to a large extent underinvested. This has created a clinical AI tool portfolio, while the system continues to produce the conditions those tools were created to mitigate.

Organizations that have made a considerable shift in their AI investments to system-level applications are likely to observe a considerable improvement in the clinical tools that have already been deployed, as the environment in which those tools operate has improved. The system is not healed through a single intervention, but through the continuous application of analytical capability to the systemic conditions of the health system that are produced by the system.

The Annals of Emergency Medicine has published studies on patient outcome and ED boarding. The National Academy of Medicine has estimated the magnitude of waste in the US healthcare system. The Dartmouth Institute has studied regional differences in the use of healthcare services. The American Hospital Association has studied the impact of hospital overcrowding on negative events. Health Affairs has published studies on predictive analytics for patient flow management. This edition links to the operational and system-level analyses in Editions A, J, O, and P of this newsletter.

Christopher Hutchins
Founder & CEO, Hutchins Data Strategy Consultants