Transforming Healthcare Decision Making Using Artificial Intelligence

Shengpu TangAAAI 2025 · DOI

Artificial Intelligence (AI) holds immense potential to revolutionize healthcare, yet a significant "deployment gap" persists between theoretical research and clinical practice. While many models show promise in retrospective studies, few are successfully integrated into hospital workflows to improve patient outcomes. Shengpu Tang’s research addresses this divide by developing "application-inspired" Reinforcement Learning (RL) methods designed specifically for the complexities of the bedside.

The Challenges of Clinical AI

Healthcare decision-making is high-stakes and characterized by high cognitive loads for clinicians. While RL is traditionally successful in "online" environments—such as games where an agent learns through trial and error—clinical settings demand an "offline" approach. Experimenting on patients to find optimal strategies is unethical; therefore, models must learn from historical, often sparse, electronic health record (EHR) data.

Beyond data limitations, two major hurdles exist:

  1. Methodological Gaps: Standard RL models often act as "black boxes," providing single recommendations that lack the nuance required for medical expertise.
  2. The Deployment Gap: Most clinical models are published in academic journals but never integrated into EHR systems, preventing them from ever assisting a patient in real-time.

Methodological Innovations

Tang proposes a synergistic approach that combines algorithmic innovation with a rigorous path to clinical deployment. Two key techniques form the core of this methodology:

Factored Action Spaces

To address the scarcity of clinical data, Tang leverages the "compositionality" of medical treatments. Rather than treating every unique combination of drugs and interventions as a distinct, unrelated state, the model decomposes them into individual factors. This structured approach allows the AI to learn more efficiently from limited data by understanding the underlying components of complex treatment plans.

Set-Valued Policies

Recognizing that AI should augment rather than replace human expertise, Tang developed "set-valued" policies. Instead of outputting a single "optimal" action, these algorithms identify a set of near-equivalent actions that are likely to lead to positive outcomes (e.g., patient survival). This allows clinicians to apply their own judgment and choose the most appropriate option from a curated, evidence-based list, increasing the likelihood of clinical adoption.

Validation and Deployment

The research emphasizes that a model's performance on a static dataset is only the first step. To bridge the deployment gap, Tang’s work involves building models directly into EHR systems and conducting quasi-experimental studies in live hospital environments.

Datasets

The models are trained and validated using diverse real-world data, including:

  • Multi-Center EHR Data: Information from 13 U.S. hospitals used to validate the M-CURES deterioration model.
  • Michigan Medicine (MM) Data: Local EHR data utilized for the deployment of infection prevention protocols and COVID-19 response.
  • Specialized Data: Insurance claims for opioid studies and longitudinal vital signs for predicting acute Graft-versus-Host Disease (aGVHD).

Key Results and Impact

The practical application of this research has already yielded measurable improvements in patient care:

  • M-CURES: This model, designed to predict early deterioration in COVID-19 patients, was successfully integrated into Michigan Medicine’s EHR. It is currently utilized by rapid response teams to inform daily rounding decisions and prioritize high-risk patients.
  • Infection Prevention: In a three-year quasi-experimental study, an AI-guided protocol for C. difficile infections led to a tangible reduction in infection rates. Furthermore, the study showed a measurable improvement in antimicrobial stewardship, ensuring more responsible use of antibiotics across the facility.

Future Directions

While these successes demonstrate the viability of application-inspired RL, several challenges remain. Future work will focus on Human-Centered Design to ensure AI tools are intuitive and do not contribute to "alert fatigue" among clinicians. Additionally, Tang aims to create a standardized Deployment Pipeline that can be replicated across different hospitals and medical conditions, scaling the impact of AI-driven decision support. By continuing to explore the compositionality of multi-stage clinical scenarios, this research seeks to make AI an indispensable partner in complex medical decision-making.

Summary updated

BibTeX
@inproceedings{tang2025transforming,
  author = {Shengpu Tang},
  title = {{Transforming Healthcare Decision Making Using Artificial Intelligence}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i27.35122},
}
Figure