Population Health and Artificial Intelligence

Corresponding Author

Introduction to population health and population health payment models

Population health refers to the focus on the health status and health outcomes of a group of individuals, with specific focus on the distribution of outcomes within the group. As such, population health increases emphasis on preventive care, chronic disease management, and health disparities. Artificial intelligence (AI) has enormous potential for accelerating population health approaches to health care because population health is inherently health care at scale. Over the past 2 decades, many factors have enhanced interest in population health, in contrast to traditional health care delivery focused on one patient at a time.

One such factor is the proliferation of big data in health care, enabling population health approaches. Understanding the distribution of health outcomes within a population requires a massive amount of discretely captured data. While advances in computing in the 1990s and data storage via the cloud in the early 2000s augmented the potential of big data, the passage of the American Recovery and Reinvestment Act of 2009 (the Stimulus) accelerated the application of big data within health care. The Stimulus mandated the meaningful use of electronic health records, effective January 1, 2014, as a prerequisite for Medicare and Medicaid reimbursement. As a result, the majority of patient clinical and financial data previously documented on paper charts is now captured electronically.

In addition, the introduction of alternative payment models has created a burning platform for health care organizations to invest in population health competencies. Alternative payment models tie reimbursement to quality and utilization outcomes, in contrast to fee-for-service payment models which pay clinicians and health systems for the volume of services provided. Health care organizations that deliver better health at lower costs can earn additional revenue under alternative payment models by sharing in the savings. The Affordable Care Act provided for the creation of the Centers for Medicare and Medicaid Services (CMS). CMS has developed and deployed payment models designed to improve care outcomes while lowering health care expenditures. Commercial payers have followed in kind. In 2022, over 40% of health service payments flowed through an alternative payment model. The U.S. Federal government is taking further steps to accelerate this transition; in January 2023, CMS announced the goal of 100% of traditional Medicare beneficiaries in an accountable care relationship by 2030.

This simultaneous acceleration of data availability and change in payment models has created enormous disruptive pressure in health care. Health systems must significantly reorient how care is delivered and prioritized. These shifts create immense data processing and operational changes—creating multiple openings for AI to have an impact on population health.

General framework for population health approaches

While strategies for addressing population health vary, the schematic in the Figure 1 serves as a general framework. In population health frameworks, health systems are often reimbursed based on risk-adjusted outcomes, so the ability to measure risk and outcomes at all stages of this continuum is particularly crucial to sustain services.

Figure 1.

General Framework for Population Health Initiatives

Population health initiatives generally begin with robust methods to identify patients who will benefit from population health services. To improve patient outcomes, health systems generally develop parallel pathways for pre-emptive health management and caring for patients who are at very high risk. Patient reporting of outcomes is important to both measure the success of programs and for receiving payment from payers. Shown in italics, artificial intelligence can improve the effectiveness of population health at multiple places on this continuum. AI = artificial intelligence.

Patient identification

First, health systems must identify patients eligible for population health services. These services are generally bifurcated into two strategies. The first is a strategy to pre-emptively manage chronic conditions such as diabetes or hypertension before they become morbid acute diseases like myocardial infarction. The second is a strategy to manage patient populations that have already developed high-risk diseases and gradually reduce that population’s risk with secondary or tertiary preventive services. Both strategies are served by optimization pathways leveraging health care teams operating at the top of license. Finally, when patients are well-optimized, this state must not only be maintained, but the risk level of patient conditions and outcomes of the patient must be well-captured and articulated back to payers.

AI holds promise to both increase the quality of care delivered and reduce the cost of care at each stage of the continuum depicted above, thereby improving population health. Patients eligible for pre-emptive health management or who are already at high risk may be better identified by AI-supported approaches. While routine computing can easily identify patients with certain discrete-field characteristics in the electronic medical record, significant information resides in free text, imaging, or even the cadence of patient interaction with the health system. AI algorithms are well-poised to recognize and exploit these patterns. For example, in our health system, we have implemented a machine learning algorithm that continuously scans free-text information in clinician notes and echocardiograms to find patients at very high risk for heart failure.¹ This algorithm continuously learns from clinician feedback.

Furthermore, the risk patients experience has better potential for capture with AI-powered algorithms. Large language models are particularly well-suited to extract meaning from free text or conversation. This information can then be input into discrete fields upon which payers rely. Beyond clinical data, we understand now that health care and socioeconomic inequity drive a significant component of risk for poor outcomes. While these risk factors are not currently well accounted for by payers in reimbursement models for population health, AI algorithms hold promise for identifying these risk factors and reporting them back to payers or allowing for health systems to have the information needed to change policy.

Chronic disease management

Obesity, diabetes, hypertension, dyslipidemia, and smoking represent the root causes of chronic, morbid conditions like chronic obstructive pulmonary disease, renal failure, heart failure, myocardial infarction, and stroke. Proactively addressing these root causes with multidisciplinary care pathways that empower patients and get patients on optimal medical therapy quickly is critical to preventing downstream morbidity. Proactive management of chronic disease requires significant provider time and expertise to deliver well and at scale. Providing chronic disease management services also requires significant back-office time to schedule appointments and coordinate care. If we can overcome these operational challenges, the payoff for population health is immense: treating primary risk factors and managing chronic disease increases health span and compresses morbidity into fewer years at the end of life, reducing average lifetime health care spending.²

While AI-based technologies can help with both patient-facing and nonpatient-facing elements of population health services, the impact will first be seen in nonclinical functions. Functions like scheduling appointments will be handled by AI agents, augmenting human care coordinators, and freeing humans to enter patient-facing, clinical roles.

AI also shows promise for augmenting human teams to provide health coaching, answer patient questions, and enhance adherence. These developments hold vast potential to speed clinical workflows and improve throughput to optimize patient risk factors in entire populations simultaneously, rather than one patient at a time in the office. Digital health, with its advances in patient wearables and home-based devices, allows for rapid, scaled intake of health care data, without patients having to come to the office for measurements. Combining digital health advances with AI advances to compute data and offer meaningful patient interaction could revolutionize how we approach chronic disease management, and in turn population health.

Patients at high risk

Population health strategies also focus on patients who are already at high risk, such as hospitalized patients. The goals are the treatment of disease, mitigation of further risk, and prevention of recurrent exacerbation of disease that results in events like hospital readmission. One major area of promise is predicting clinical deterioration risk. Conventional regression-based risk prediction models may perform better when augmented by machine learning-based risk prediction.³ Future AI algorithms may be able to incorporate elements beyond discrete fields in the electronic medical record, such as patient appearance as adjudicated by computer vision, high-fidelity waveform data from pulse oximetry or cardiac telemetry, and direct evaluation of medical imaging such as computed tomography scans or echocardiograms. Patients recently discharged from the hospital also represent a population at high risk of clinical deterioration. During this period of transitional care, AI-based agents may act as health coaches for patients or help flag to clinical teams patients at risk of clinical deterioration.

Patient reporting

Measuring health outcomes represents a critical component of any population health strategy for two major reasons. First, outcomes reporting allows health systems to understand whether strategies implemented for population health are effective. Second, risk-adjusted outcomes are critical for determining the value of care delivered, which in turn impacts payments from commercial and government payers to health systems for implementing population health.

AI based on large language models can help abstract data from patient records to determine outcomes and identify risk factors that mediate these outcomes. In the current state, these data must be aggregated and input into discrete fields to effect payment. Currently, this work is carried out by laborious human chart review. Beyond this, AI holds promise for helping health systems articulate to payers elements of risk that may not be well captured using current systems.

The downsides of artificial intelligence

A critical downside of reliance on AI is the potential for introducing bias into population health management. AI algorithms often train on human-generated data sets, creating strong potential for encoding human biases into algorithms. Because population health is a field very concerned with addressing inequities in the distribution of health care outcomes, baking bias into AI population health tools would be antithetical to the goals of population health. To counter the potential for AI bias, we must be careful about representation and content in AI training datasets, and continuously monitor model performance.

A second downside of AI particularly relevant to population health is AI hallucination. AI has the potential to create misinformation. While this can be particularly dangerous in health care, misinformation is even more dangerous in population health. Much of the promise of population health is the promise of scale—the idea that we can deliver safe and effective medical care not to one patient at a time, but to an entire population at once. Misinformation distributed by AI in the form of chatbots or virtual health coaches could inadvertently spread disinformation. Effective human-supervised guardrails must be in place. Furthermore, because a fundamental aspect of AI algorithms is continual learning based on input, this supervision must be continuous as AI algorithms evolve.

Conclusions

Population health represents a promising approach to delivering health care at scale. AI has many practical applications at many stages of the care delivery continuum for population health.

Funding support and author disclosures

The authors have reported that they have no relationships relevant to the contents of this paper to disclose.