Physician Cognitive Burden, Telemedicine, and Low-Value Care

Presentation Type: Podium Abstract
Presentation Time: 03:30 PM - 03:42 PM

Abstract Keywords: Telemedicine, Healthcare Quality, Evaluation
Programmatic Theme: Clinical Informatics

Using EHR metadata from 449,799 ambulatory encounters, we examined how physician cognitive burden relates to low-value care (LVC) and how telemedicine moderates this relationship. Higher cognitive burden was associated with greater LVC in in-person visits, whereas telemedicine encounters showed the opposite pattern, with increasing burden linked to lower LVC and a clear dose–response effect, suggesting that strategically deploying telemedicine and cognitive burden–aware decision support may help health systems reduce LVC without reducing access.

Speaker(s):
A J Holmgren, PhD
University of California, San Francisco

Modeling Hospital Financial Risk under Capitated and Bundled Payments Using Medicare Inpatient Data

Presentation Type: Paper - Student
Presentation Time: 03:42 PM - 03:54 PM

Abstract Keywords: Artificial Intelligence, Machine Learning, Quantitative Methods, Real-World Evidence Generation, Population Health
Programmatic Theme: Clinical Research Informatics

This study develops a predictive model to estimate hospital financial risk under value-based payment systems using CMS Medicare Inpatient Hospital data. Using provider-level data from over 3,000 U.S. hospitals, we evaluated multiple supervised learning approaches to predict payment-to-charge ratios, a proxy for reimbursement efficiency under bundled and capitated payment models. Predictor variables included measures of patient complexity, demographics, utilization, and hospital characteristics. Random-Forest regression outperformed linear regression, achieving stronger predictive performance (correlation coefficient 0.79 vs. 0.61). Feature-importance analysis identified chronic-condition prevalence, case-mix index, and average length of stay as the strongest predictors of lower reimbursement efficiency. Results suggest that hospitals serving more clinically complex populations face greater financial vulnerability under value-based payment arrangements. This framework demonstrates how machine-learning methods applied to administrative healthcare data can support hospital administrators and policymakers in anticipating financial risk and improving payment-model design.

Speaker(s):
Zareen Taj Shaik, Health Informatics
George Mason University

Measuring the Financial and Utilization Impact of Hospice Enrollment Among Medicare Advantage Decedents

Presentation Type: Paper - Regular
Presentation Time: 03:54 PM - 04:06 PM

Abstract Keywords: Population Health, Chronic Care Management, Transitions of Care
Programmatic Theme: Public Health Informatics

This retrospective cohort study examined the impact of hospice enrollment on healthcare utilization and total cost of care (TCOC) among 7,554 beneficiaries in a regional Medicare Advantage plan during the last 90 and 180 days of life. Using administrative claims data from January 2024 through September 2025, decedents were categorized by hospice enrollment and adjusted for demographics and comorbidities. Hospice utilization was observed in 44% of the cohort, with a higher prevalence of short hospice stays than national benchmarks. Propensity score matching and regression analysis found that earlier and longer hospice enrollment is associated with greater reductions in 90-day total cost of care, with effect modification observed in patients with ESRD and older age. Each additional hospice day was associated with an estimated reduction of approximately $230, with this marginal effect attenuating monotonically. Findings highlight hospice’s role in patient-centered end-of-life care, with cost effects emerging as a secondary outcome of timely enrollment.

Speaker(s):
Andi Shapiro, MHA
Geisinger

Implementing Population Health Risk Analytics on a Standardized EHR Data Layer: Evaluation and Framework Development

Presentation Type: Podium Abstract
Presentation Time: 04:06 PM - 04:18 PM

Abstract Keywords: Population Health, Data transformation/ETL, Data Standards
Programmatic Theme: Clinical Research Informatics

Population health risk analytics are typically implemented outside standardized EHR data models, limiting scalability across healthcare systems. We implemented the Adjusted Clinical Groups (ACG) population health analytics framework within the OMOP CDM and evaluated outputs using raw EHR and OMOP data from the same patient population. Differences in utilization, morbidity burden, and risk stratification indicate that data standardization may influence analytic outputs. The study also demonstrates the feasibility of implementing population health analytics within OMOP.

Speaker(s):
Haeun Lee, MS
Johns Hopkins University

Predicting Opioid Overdose Mortality in Pennsylvania: A Multi-Level Analysis of Structural and Incident-Level Factors

Presentation Type: Paper - Student
Presentation Time: 04:18 PM - 04:30 PM

Abstract Keywords: Quantitative Methods, Public Health, Machine Learning, Data Mining, Racial Disparities
Programmatic Theme: Public Health Informatics

The opioid crisis has evolved rapidly over the past two decades, shifting from prescription opioid misuse to a complex illicit drug epidemic. This study examines the determinants of opioid overdose mortality in Pennsylvania between 2007 and 2023 using a multi-level analytical framework. County-level fixed effects models identify factors associated with variation in overdose deaths across counties and over time. Results show that mortality changes are largely unpredictable using observable county characteristics, with only a small number of variables, including heroin seizures, minority population share, and Medicare spending, remaining statistically significant under the most stringent tests. Notably, fentanyl seizures explain differences between counties but do not predict within-county mortality changes, suggesting heterogeneous saturation patterns. Individual-level analysis of more than 41,000 overdose incidents reveals that naloxone administration is the strongest determinant of survival, reducing death rates from 53% to 8%. However, access to naloxone varies systematically across race, geography, and drug type.

Speaker(s):
Dusan Ramljak, PhD
The Pennsylvania State University

Autonomy Budgets for Clinical AI: A Simulation Study of Performance-Gated Allocation of Human Oversight

Presentation Type: Paper - Regular
Presentation Time: 04:30 PM - 04:42 PM

Abstract Keywords: Clinical Decision Support, Artificial Intelligence, Human-computer Interaction, Governance
Programmatic Theme: Clinical Informatics

As clinical AI systems move from advisory tools to operational decision-makers, healthcare organizations lack principled methods for determining when autonomous action is appropri- ate versus when human review should occur. Because clinician attention is limited, escalation decisions must allocate scarce oversight capacity across AI-generated decisions. We introduce Performance-Gated Autonomy (PGA), an architectural framework that allocates human over- sight through sequential statistical certification (S-CPC), epistemic uncertainty detection (I- TEC), and decision-theoretic escalation (DTE). Oversight capacity is modeled as a constraint κ, representing the fraction of cases that can be escalated. Monte Carlo simulations show that escalation policies prioritize cases with highest value of information, producing diminish- ing marginal safety benefit as escalation increases. Governance policies therefore lie along an autonomy–safety frontier. These results suggest that safe clinical AI deployment requires explicit management of autonomy budgets governing how frequently AI systems may act autonomously.

Speaker(s):
Blackford Middleton, MD, MPH, MSc
Retired