Where the gap actually was

The system had no shortage of data and no shortage of clinical talent. What it didn’t have was a single place where the people responsible for the discharge-to-day-30 window could see, at a glance, who was most likely to come back and why. Readmission risk lived in three or four parallel systems: a vendor risk score embedded in the EHR that the floor teams had stopped trusting, a periodic claims-driven report that arrived weeks after the fact, an ambulatory follow-up tracker maintained largely in spreadsheets, and a payer-shared population view that only a handful of analysts could open.

Care managers were spending more time figuring out which discharged patients were theirs that day than actually intervening with them. Social workers were learning about high-risk discharges from phone calls rather than from a worklist. The clinical leaders we worked with were direct about it: the issue wasn’t model accuracy, it was that nothing accurate ever made it to the bedside or to the post-discharge call queue in time to matter.

What we built

We started by getting the data into one place. Inpatient encounters from the EHR, ambulatory visits, claims-derived prior utilization, medication fill patterns, social determinants indicators captured at admission, and post-discharge follow-up status were unified into a governed analytics layer with patient-level lineage back to source. That layer became the single substrate for everything downstream — the risk model, the care manager worklists, the executive readmission dashboards, and the payer-facing reports.

On top of that we built a readmission risk score validated against the system’s own twelve months of outcomes rather than against a generic benchmark cohort. The model was deliberately boring: a calibrated, interpretable score that clinical reviewers could challenge feature by feature. Interpretability mattered more than a fractional point of AUC, because the care managers needed to know why a patient was flagged in order to choose the right intervention.

The most important piece wasn’t the model. It was where the score showed up. We surfaced it inside the workflows the transitional care team and ambulatory clinicians already used — the discharge planning view, the post-discharge call queue, the ambulatory care manager’s daily list — rather than asking anyone to log into a new tool. High-risk patients showed up automatically on the right worklists, with the top contributing factors and the suggested intervention pathway already attached.

What changed

Twelve months in, the 30-day all-cause readmission rate was down 22% across the three flagship hospitals, with the largest absolute gains in the heart failure and COPD cohorts where the post-discharge intervention had the most leverage. The share of high-risk discharges reached by a care manager within 72 hours moved from 41% to 86% — that single operational shift accounts for a meaningful fraction of the headline number. Modeled against the prior baseline, the system avoided roughly 1,400 readmissions over the period.

Internally, the more durable change is how the transitional care team spends its day. Care managers reported a 58% reduction in the time spent locating and prioritizing patients, redirected into actual intervention work — medication reconciliation calls, follow-up visit scheduling, SDOH-driven referrals. The work the team was hired to do went from being the second half of the day to being the whole day.

Why it stuck

The risk score is recalibrated on the system’s own ongoing outcomes, so it tracks reality instead of decaying against a frozen training set. The data layer underneath is governed, version-controlled, and tested, which means the next analytics initiative — the system is already extending it to ED utilization and post-acute placement — starts from a working foundation rather than from another data-collection project. And because the score lives inside existing clinical workflows rather than in a separate tool, adoption didn’t depend on anyone changing their habits. The 22% number is the externally visible result; the internal result is a transitional care program that finally has the operational handle it was supposed to have all along.