Presentation
Predicting Readmission Risk at the Point-Of-Care: A Personalized AI Generated Report Leveraging Multi-Source Healthcare Data
SessionPoster Session 1
DescriptionHospital readmissions remain a costly and persistent challenge in the US. Hospital clinicians often lack timely, personalized information to anticipate this risk early in the admission episode. Most existing prediction tools operate late in the hospital stay and require detailed discharge data, leaving clinicians without actionable information at the point when triage, care planning, and discharge preparation begin. Moreover, many models are trained within a single health system, failing to capture pre-hospital utilization and fragmented care across networks. These gaps leave clinicians dependent on raw data from health information exchanges that are organized but not systematized, demanding manual interpretation under time pressure.
This study introduces the Admission Readmission Risk Report (ARRR), a proof-of-concept prototype designed to introduce readmission risk analysis that considers pre-admission data (i.e., data from primary and other community care providers) and can be generated upon hospital arrival. Using a large regional health information exchange dataset of over 66,000 admissions, we developed a model to predict 30-day readmission. SHAP explanations were applied to identify global and patient-specific drivers of risk, emphasizing medication burden, chronic disease complexity, and utilization history. The ARRR integrates four coordinated components: a risk snapshot, a percentile-based risk distribution, individualized risk feature contributions, and a structured clinical and social profile. Together, these elements transform scattered data into an interpretable, multimodal summary aligned with clinician needs.
A preliminary evaluation with practicing clinicians demonstrated that the report improved comprehension of patient risk and was perceived as useful for supporting care planning, though participants also suggested ways to enhance actionability. These findings highlight the potential of embedding explainable AI predictions into admission summaries to bridge pre-hospital and in-hospital perspectives, inform discharge readiness, and guide follow-up strategies.
Presenting structured, explainable, and arrival-time risk information can better support clinicians in making early, resource-sensitive decisions. The take-away point for attendees is that prototypes such as the ARRR not only improve predictive accuracy but also make risk information actionable by embedding it in formats clinicians can interpret and use from the very beginning of the hospital stay.
This study introduces the Admission Readmission Risk Report (ARRR), a proof-of-concept prototype designed to introduce readmission risk analysis that considers pre-admission data (i.e., data from primary and other community care providers) and can be generated upon hospital arrival. Using a large regional health information exchange dataset of over 66,000 admissions, we developed a model to predict 30-day readmission. SHAP explanations were applied to identify global and patient-specific drivers of risk, emphasizing medication burden, chronic disease complexity, and utilization history. The ARRR integrates four coordinated components: a risk snapshot, a percentile-based risk distribution, individualized risk feature contributions, and a structured clinical and social profile. Together, these elements transform scattered data into an interpretable, multimodal summary aligned with clinician needs.
A preliminary evaluation with practicing clinicians demonstrated that the report improved comprehension of patient risk and was perceived as useful for supporting care planning, though participants also suggested ways to enhance actionability. These findings highlight the potential of embedding explainable AI predictions into admission summaries to bridge pre-hospital and in-hospital perspectives, inform discharge readiness, and guide follow-up strategies.
Presenting structured, explainable, and arrival-time risk information can better support clinicians in making early, resource-sensitive decisions. The take-away point for attendees is that prototypes such as the ARRR not only improve predictive accuracy but also make risk information actionable by embedding it in formats clinicians can interpret and use from the very beginning of the hospital stay.
Event Type
Poster Presentation
TimeMonday, March 234:45pm - 6:15pm EDT
LocationRhinelander Gallery
Digital Health