- A readmission model scores each patient from the current stay, diagnoses, prior use, medications and social factors recorded in the EMR, before discharge.
- Reason codes show which factors raised each score, so care managers know what to address in the discharge plan.
- A simple app puts scores, reasons and trends in front of care managers, not only data scientists.
- The score orders a worklist. Your clinicians decide every step.
Discharge planning happens before you know who will come back
Your care management team has a fixed number of follow-up calls, home visits and transition-of-care slots. Under value-based contracts, every avoidable readmission costs you, and the patients who return often come back sicker.
Most teams rely on a checklist score or a nurse's judgment at discharge. That misses patients whose risk sits in the pattern of the stay rather than in a single diagnosis. A model that reads the whole record gives your team an ordered list, with reasons, while there is still time to act.
How predicting readmission works
A classification model is trained on your own history and estimates the chance a patient is readmitted within a window you set, usually 30 days.
Step 1: connect the history
Pull encounters, diagnoses, procedures, medications, prior admissions and recorded social factors from the EMR you already run. The model sits on top of the EMR; it does not replace it.
Step 2: build and compare models
H2O Driverless AI runs automatic feature engineering and ranks candidate models on held-out discharges.
Step 3: explain each score
Interpretability methods and Shapley values show which factors raised each score.
Step 4: put it in front of care managers
A simple app shows scores, reasons and readmission trends by unit and service line.
Step 5: run it in shadow mode, then monitor
Score patients beside your current process first, then watch for model drift as your patient mix and programs change.
Where the decision stays
The model orders a discharge worklist with the reasons attached. Your care managers and clinicians decide the plan for each patient, and no clinical efficacy claim is made for the score. Protected health information never has to leave your boundary, in any of four deployment options: on-premises, air-gapped, private cloud or cloud.