Readmission Prediction
“Which patients will be readmitted within 30 days?”
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
By submitting, you accept the Terms and Privacy Policy.
Loved by data scientists, ML engineers & CXOs at
A real-world example
Which patients will be readmitted within 30 days?
CMS penalizes hospitals up to 3% of Medicare reimbursements for excess readmissions. For a 400-bed hospital processing 25,000 discharges per year, each avoidable readmission costs $15,000 on average. A 15% reduction in 30-day readmissions saves $5.6M annually in penalties and direct care costs. Traditional LACE scores miss the cross-patient signals hidden in shared providers, medication overlaps, and procedure histories.
Quick answer
Graph-based AI predicts 30-day hospital readmissions by connecting patient records, encounter histories, diagnoses, procedures, and medications into a relational network. Unlike LACE scores that evaluate each patient in isolation, relational models detect cross-patient patterns: patients sharing the same attending physician with specific procedure-diagnosis combinations have correlated readmission risk. A 400-bed hospital reducing readmissions by 15% saves $5.6M annually in CMS penalties and direct care costs.
Approaches compared
4 ways to solve this problem
1. LACE Index
A validated scoring tool using four variables: Length of stay, Acuity of admission, Comorbidities (Charlson score), and Emergency department visits in prior 6 months. The most widely used readmission risk tool in US hospitals.
Best for
Quick risk stratification at discharge when you need a simple, interpretable score that nurses and case managers can act on immediately.
Watch out for
LACE uses only four variables and ignores medication complexity, provider patterns, and cross-patient signals. Its discriminative ability (C-statistic 0.68-0.72) is not much better than a coin flip for individual patient decisions.
2. Logistic Regression on EHR Features
Custom models trained on 20-50 EHR features: demographics, diagnoses, lab values, prior utilization. Built by hospital analytics teams using their own data.
Best for
Hospitals with mature analytics teams that can extract and maintain EHR feature pipelines. Better than LACE because it uses more features and is trained on local data.
Watch out for
Feature engineering from EHR data is slow and fragile. Each new predictor requires clinical validation, ETL pipeline changes, and model retraining. The models still treat each patient as an independent row.
3. Commercial Risk Scores (Epic, Cerner)
EHR vendor-provided readmission risk scores embedded in the clinical workflow. Trained on multi-site data with periodic updates. Integrated into discharge planning dashboards.
Best for
Hospitals that want a turnkey solution integrated into their existing EHR without building custom models.
Watch out for
Black-box scores trained on general populations may not reflect your specific patient mix, provider patterns, or community factors. You cannot customize them for your highest-value use cases.
4. Graph Neural Networks (Kumo's Approach)
Builds a heterogeneous graph across patients, encounters, diagnoses, procedures, medications, and providers. Learns cross-patient readmission patterns from the full relational structure of the EHR.
Best for
Detecting network-level risk signals: medication interaction patterns across the patient population, provider-specific readmission correlations, and procedure-diagnosis combinations that predict complications.
Watch out for
Requires connected EHR data across multiple tables. If your hospital's data is trapped in disconnected clinical systems, the integration work comes first.
Key metric: LACE scores achieve a C-statistic of 0.68-0.72. Graph-based models exceed 0.82 by capturing cross-patient signals. A 15% readmission reduction saves $5.6M annually for a 400-bed hospital.
Why relational data changes the answer
Flat readmission models score each patient based on their own clinical data: age, diagnoses, lab results, prior visits. They can identify that an 81-year-old with ESRD and 4 recent inpatient encounters is high risk. But they cannot see that this patient's attending physician has a 34% readmission rate across all their patients (versus a 18% hospital average), that the specific combination of hemodialysis plus the 7 active medications this patient takes has a known interaction pattern, and that other patients discharged from the same unit in the same week had elevated readmission rates suggesting a systemic issue (perhaps staffing or discharge-process related). These cross-patient signals live in the relationships between patients, providers, medications, and encounters.
Relational learning connects these entities directly. The model walks from patient to their encounters, to the providers on those encounters, to other patients seen by those providers, to the medication patterns across those patient cohorts. It discovers that shared-provider readmission rate is the third most predictive feature for Patient P1003, contributing 18% of the prediction attribution. This is a signal that no patient-level model can capture because it requires looking across the patient network. The result is a readmission score that accounts for both individual clinical risk and systemic network factors, producing a C-statistic above 0.82 compared to 0.68-0.72 for LACE.
Predicting readmissions from individual patient records is like predicting which students will fail a class by looking only at their GPA. You miss that the professor has a 40% failure rate across all sections, the textbook was recently changed to a harder edition, and students in the Monday section consistently underperform Tuesday students. The individual student matters, but the network they are embedded in matters just as much. Relational readmission models see both the student and the classroom.
How KumoRFM solves this
Graph-learned clinical intelligence across your entire patient network
Kumo builds a heterogeneous graph across patients, encounters, diagnoses, procedures, and medications. It learns that patients sharing the same attending physician with specific procedure-diagnosis combinations have correlated readmission risk. The model captures medication interaction patterns across the patient network, not just individual patient features. One PQL query replaces months of manual feature engineering across siloed EHR tables.
From data to predictions
See the full pipeline in action
Connect your tables, write a PQL query, and get predictions with built-in explainability — all in minutes, not months.
Your data
The relational tables Kumo learns from
PATIENTS
| patient_id | age | gender | insurance |
|---|---|---|---|
| P1001 | 72 | M | Medicare |
| P1002 | 58 | F | Commercial |
| P1003 | 81 | M | Medicare |
ENCOUNTERS
| encounter_id | patient_id | admit_date | discharge_date | type |
|---|---|---|---|---|
| E5001 | P1001 | 2025-02-10 | 2025-02-16 | Inpatient |
| E5002 | P1002 | 2025-02-20 | 2025-02-23 | Inpatient |
| E5003 | P1003 | 2025-02-25 | 2025-03-02 | Inpatient |
DIAGNOSES
| diagnosis_id | encounter_id | icd10_code | description |
|---|---|---|---|
| D001 | E5001 | I50.9 | Heart failure, unspecified |
| D002 | E5002 | J44.1 | COPD with acute exacerbation |
| D003 | E5003 | N18.6 | End-stage renal disease |
PROCEDURES
| procedure_id | encounter_id | cpt_code | description |
|---|---|---|---|
| PR001 | E5001 | 93306 | Echocardiogram |
| PR002 | E5002 | 94640 | Nebulizer treatment |
| PR003 | E5003 | 90935 | Hemodialysis |
MEDICATIONS
| rx_id | patient_id | drug_name | start_date | active |
|---|---|---|---|---|
| RX01 | P1001 | Furosemide 40mg | 2025-02-10 | Y |
| RX02 | P1002 | Albuterol inhaler | 2025-02-20 | Y |
| RX03 | P1003 | Epoetin alfa | 2025-01-15 | Y |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT BOOL(ENCOUNTERS.*, 0, 30, days) FOR EACH PATIENTS.PATIENT_ID WHERE ENCOUNTERS.TYPE = 'Inpatient'
Prediction output
Every entity gets a score, updated continuously
| PATIENT_ID | DISCHARGE_DATE | READMIT_30D | PROBABILITY |
|---|---|---|---|
| P1001 | 2025-02-16 | True | 0.74 |
| P1002 | 2025-02-23 | False | 0.18 |
| P1003 | 2025-03-02 | True | 0.89 |
Understand why
Every prediction includes feature attributions — no black boxes
Patient P1003 -- 81yo Male, ESRD
Predicted: True (89% readmission probability)
Top contributing features
Number of inpatient encounters (last 90d)
4 visits
31% attribution
Active high-risk medication count
7 drugs
24% attribution
Shared-provider readmission rate
34%
18% attribution
Diagnosis complexity (HCC score)
3.2
15% attribution
Days between last two admissions
12 days
12% attribution
Feature attributions are computed automatically for every prediction. No separate tooling required. Learn more about Kumo explainability
PQL Documentation
Learn the Predictive Query Language — SQL-like syntax for defining any prediction task in 2–3 lines.
Python SDK
Integrate Kumo predictions into your pipelines. Train, evaluate, and deploy models programmatically.
Explainability Docs
Understand feature attributions, model evaluation metrics, and how to build trust with stakeholders.
Frequently asked questions
Common questions about readmission prediction
How does AI predict hospital readmissions?
AI predicts readmissions by analyzing connected patient data: diagnoses, procedures, medications, provider patterns, and encounter history. Graph-based models go beyond individual patient features to detect cross-patient signals like shared-provider readmission rates and medication interaction patterns. This produces more accurate risk scores (C-statistic above 0.82) than traditional tools like LACE (0.68-0.72).
What is the CMS Hospital Readmissions Reduction Program penalty?
CMS penalizes hospitals up to 3% of Medicare reimbursements for excess 30-day readmissions for targeted conditions (heart failure, pneumonia, COPD, hip/knee replacement, coronary artery bypass graft, and acute MI). For a large hospital, penalties can reach $3-5M annually. Reducing readmissions by even 10-15% can eliminate the penalty entirely.
What is the LACE score and how accurate is it?
LACE is a readmission risk score using four variables: Length of stay, Acuity (emergency vs. elective), Comorbidities (Charlson score), and Emergency visits in prior 6 months. Its C-statistic ranges from 0.68-0.72, meaning it correctly ranks a readmitted patient above a non-readmitted patient only 68-72% of the time. Graph-based models improve this to 82%+ by incorporating medication, provider, and cross-patient signals.
How much does a hospital readmission cost?
The average cost of a 30-day readmission is $15,000 in direct care costs. On top of that, CMS penalties for excess readmissions can reach 3% of total Medicare reimbursements. For a 400-bed hospital processing 25,000 discharges per year, a 15% reduction in readmissions saves $5.6M annually from the combined effect of avoided care costs and eliminated penalties.
Bottom line: A 400-bed hospital reducing 30-day readmissions by 15% saves $5.6M per year in CMS penalties and direct care costs. Kumo learns cross-patient signals from shared providers, medication overlaps, and procedure patterns that LACE scores miss entirely.
Related use cases
Explore more healthcare use cases
Topics covered
One Platform. One Model. Infinite Predictions.
KumoRFM
Relational Foundation Model
Turn structured relational data into predictions in seconds. KumoRFM delivers zero-shot predictions that rival months of traditional data science. No training, feature engineering, or infrastructure required. Just connect your data and start predicting.
For critical use cases, fine-tune KumoRFM on your data using the Kumo platform and Research Agent for 30%+ higher accuracy than traditional models.
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
