Hospital Readmissions Risk Prediction and Prevention (HARPP)
The HARPP model utilizes electronic health record (EHR) data combined with data used to compute the LACE score to predict risk of readmission and the likelihood of a patient benefiting from a post discharge intervention.
An unplanned readmission is a hospital readmission that occurs within 30 days of the initial admission. Reducing readmissions yields significant benefits for a hospital system. Initiatives such as the Blue Cross Blue Shield Pay-for-Performance program, the Center for Medicare & Medicaid (CMS)’s Hospital Readmission Reduction Program (HRRP), or value-based contracts hinge on the performance of this metric. High readmission rate reflects low patient care quality.
Frequent readmission leads to overall higher healthcare costs
Hospitals are financially incentivized to reduce hospital readmission rate
Performance on readmissions is included in every value-based contract in which Michigan Medicine participates
Meets strategic BASE goals at Michigan Medicine
These results will be immediately implemented in practice. Whichever model turns out to have the best performance will be used operationally at Michigan Medicine to support the Central TOC team’s goal to reduce readmissions.
In developing the HARP model, the team has made impactful technical contributions in the field of AI and Machine Learning. First, they introduced a novel technique for estimating heterogeneous treatment assignment effects under non-adherence (e.g., when an individual is prescribed a medication as treatment, but fails to take it) in which a multi-task neural network implements conditional front-door adjustments with joint modeling of nuisance parameters. This approach reduces estimation error across several semi-synthetic and real-world datasets compared to baselines, even when the true effect of treatment assignment is large. This work was published and presented at the Conference on Health, Inference, and Learning in June 2025.
Additionally, the team investigated multiple AI model evaluation strategies that leverage all data collected in an RCT (instead of only using data from the control group, which is currently the prevailing approach for unbiased evaluations of AI models). The team introduces a novel evaluation approach, Nuisance Parameter Weighting (NPW), and demonstrates that it yields better model selection compared to standard approaches that ignore data from the treatment group. This work has been accepted for publication and will be presented at AAAI under the AI for Social Impact Special Track in January 2026.
Digital Health Innovation Support
Digital Health Innovation is supporting a 3-month randomized control trial (RCT), where post-discharge phone calls will be performed based on an EHR developed lottery tool rather than the LACE score. The RCT will enable us to compare results from LACE, Epic Readmissions Model, and HARPP. We will evalueate data-drive approaches to target post-discharge interventions at Michigan Medicine to reduce readmissions as well as:
Develop an EHR-based lottery design tool in MiChart
Explore benefit- over risk-based approach for guiding interventions
Explore utility of hospital-developed (custom) vs. hospital-fine-tuned (Epic) vs. generic risk scores (LACE)
Principal Investigators:
Michael Sjoding, MD, MSc
Jenna Wiens, PhD,
Associate Professor, Computer Science and Engineering