Reducing Hospital Readmissions through Predictive Modeling

Using predictive analytics to identify high-risk patients and enable proactive care decisions.

Case Study

Hospital readmission risk is a problem where the analytical method matters more than the dataset. This piece uses the New York SPARCS public hospital discharge data to show how we build a risk model end to end: from a star-schema data foundation, through feature engineering and model development, to the dashboard a director of nursing would actually use on a Monday morning. The approach, the documentation standard and the deliverables are exactly what a client engagement produces.

The challenge

Identifying patients at high readmission risk before discharge, rather than reacting after a return. Care approaches that are reactive by default. Resource prioritization made without a view of who is most likely to come back.

Why it matters

Readmissions carry financial penalties, worse patient outcomes, and operational strain on a facility already running close to capacity. The reporting that governs a hospital's performance on this measure is frequently the reporting it trusts least.

Our approach

1. Data preparation

A star-schema relational database built for analysis rather than transaction, so clinical and financial facts can be examined together.

2. Feature engineering

Identifying the patterns that capture utilization history and clinical risk drivers, documented so each feature's rationale is traceable.

3. Predictive modeling

Machine learning models for readmission prediction, with stated use cases and limitations.

4. Performance tracking

Dashboards and reports for monitoring, designed for the user making the discharge decision.

What the analysis showed

A small group of drivers accounts for a disproportionate share of readmission risk in this dataset: concentrated among elderly and Medicare patients, higher-severity cases, and those with major chronic conditions, confirming literature on readmissions.

The operational implication: A targeted transitional care program aimed at defined high-risk groups, other than one applied across all discharges.

Key takeaways

Early detection or readmission risk identified at discharge can be acted on, risk identified afterwards cannot.
Adoption and delivery depends on the predictive model of choice.
The data foundation does most of the work. Model performance was constrained far more by data structure and completeness.
A risk model influencing care decisions (reducing readmission risk and improving resource allocation and value-based care) must be explainable to clinician, compliance officer and board.

If this sounds like your organization

If you run a hospital, a senior living community or a care home and your occupancy report and your revenue report disagree, that is a smaller problem than this one, we solve it in ten business days.

Originally developed as a graduate capstone project for MSc Business Analytics at Lewis University, with Dev Arora, Tony Lordson and Anil Kumar Swamy Bandaru. Data source: New York State SPARCS, public dataset.