Predict 30-day hospital readmission

What the model found

The model scores every encounter; the dashboard above zooms in on the 20 with the highest risk. Read as a group, they share a clear profile.

A tight band at the very top

All 20 sit in a narrow risk-score band of 2.271–2.324 (mean 2.298), with encounter 120625968 at the top (2.324). A tight cluster at the extreme end is what you want to see — these are the model’s most-confident readmission risks, so follow-up effort spent here is effort spent well.

Who’s most at risk

Within this cohort of 20:

• Gender — 80% male (16 of 20), 20% female (4). But the highest average score skews female (2.307) over male (2.295).
• Age — the 70–80 group is the most common (30%, 6 of 20), while the 50–60 group carries the highest average risk (2.311).
• Race — 75% Caucasian (15), 25% African American (5); average risk is close across both (2.299 vs 2.294).

The clinical pattern

• Emergency admissions dominate — admission type 1 (emergency) accounts for 13 of 20 (65%).
• Most go home — and are still high-risk — discharge disposition 1 (home) covers 15 of 20 (75%). The rest split across special dispositions (two to 22, two to 3, one to 18).

That combination — an emergency admission discharged home — is the signature of this high-risk group. These are exactly the patients who leave without a structured hand-off and can bounce back within the month.

Highest-risk combinations

A few demographic-and-clinical combinations rise to the top:

• 40–50, female, African American — 2.321 (highest)
• 50–60, female, Caucasian — 2.321 (tied)
• 70–80, male, Caucasian — 2.296 (the most common high-risk group, 6 cases)

These patterns describe the 20 highest-risk encounters, not the whole population. Use them to target interventions — not to profile patients.

What to do with it

The score turns a discharge list into a priority queue. The clearest action: give emergency admissions who are discharged home enhanced post-discharge monitoring — a follow-up call, medication reconciliation, or a home-health referral — while they’re still inside the 30-day window.

Why it’s worth solving

A 30-day readmission is rarely just a number. It’s a patient who got worse after going home — and a bed that fills back up with a problem that was often preventable.

The stakes are clinical and financial at once. Readmissions are among the most scrutinized events in a hospital: Medicare’s Hospital Readmissions Reduction Program ties a share of hospitals’ reimbursement to their 30-day rates, and widely cited estimates put the cost of avoidable readmissions in the tens of billions of dollars a year. Behind each one is the human cost of a recovery that didn’t hold.

But the binding constraint usually isn’t money — it’s attention. No care team can give every discharge a follow-up call, a medication review, and a home-health referral. That capacity is finite, and today it’s spread evenly across a discharge list where the risk is anything but even.

Follow-up capacity is finite. A risk score decides where it goes.

That’s the leverage. Ranking every encounter by risk sends the same limited hours to the patients most likely to return — the emergency admissions discharged home, not the routine cases that were always going to be fine. Even a small, targeted drop in readmissions pays for the effort many times over: in avoided penalties, in beds freed for the next patient, and in outcomes that never show up on an invoice.

How Langsat builds this

This is a Data Science project with a binary-classification target (readmitted within 30 days: yes or no).

1. Point Langsat at your encounter data — and the tables around it: patients, diagnoses, medications, labs, prior visits. You don’t pre-join them.
2. Train on the relationships, not a flattened table. Langsat’s relational models read the patient graph directly — a patient’s diagnoses, medications, and prior encounters all inform the score — so signal that manual flattening throws away is kept.
3. Review the model, then serve it. Langsat reports the standard binary-classification metrics for you to check before anything ships. Once you’re satisfied, every encounter gets a live risk score through the prediction API and the dashboards you saw above.
4. Ask questions in plain English — “which 20 encounters have the highest readmission risk?” — and share a read-only dashboard with the discharge team by link.

The whole thing runs on Langsat’s secure cloud — your data isolated to your account, encrypted in transit and at rest, and never used to train models for anyone else.