Hospital Readmission Prediction: UCI Diabetes 130-US Hospitals
The same platform proven in sports and retail was deployed on real hospital patient data with zero code changes. Here's what it found.
Hospital readmission prediction is widely considered one of the hardest problems in healthcare analytics. The CMS (Centers for Medicare & Medicaid Services) uses readmission rates to penalize hospitals, and the entire industry has been working to improve prediction accuracy for over a decade.
| Measure | Value | Context |
|---|---|---|
| Projected Rate | 10.78% | Projected readmission rate |
| Actual Rate | 10.80% | Actual readmission rate |
| Absolute Error | 0.02 ppt | Industry typical: 2–5 ppt |
| Relative Error | 0.18% | |
| Bias | -0.0002 | Slight underestimate (negligible) |
The actual readmission rate ranged from 9.0% to 13.6% across unseen data. This was not a flat, easy-to-predict signal. The rate moved substantially. The predicted 10.78% landed almost exactly on the actual.
| Method | Prediction | Error | vs. DIGINETICS |
|---|---|---|---|
| DIGINETICS | 10.78% | 0.18% | |
| Training Mean (historical average) | 11.20% | 3.66% | 20× more accurate |
| Last Known Value (naive persistence) | 9.60% | 11.11% | 60× more accurate |
The platform didn't just marginally beat the baselines. It delivered an order-of-magnitude improvement over both the naive persistence baseline and the historical mean.
Published systematic reviews of hospital readmission prediction models (81+ models reviewed across BMJ Open, JAMA, and BMC) show the following performance landscape:
| Approach | Typical AUC / C-Statistic | Source |
|---|---|---|
| CMS Official Models (Medicare) | 0.60 – 0.63 | CMS Hospital Compare |
| Logistic Regression (standard ML) | 0.65 – 0.70 | BMJ Systematic Review, 2020 |
| Best ML Models (GBDT, LSTM) | 0.72 – 0.76 | PLOS Digital Health, 2024 |
| LLMs + Physician Panels | Below clinical reliability | medRxiv, 2025 |
Note: DIGINETICS predicts the continuous readmission rate across a hospital system over time. The 0.02 percentage point error is substantially better than the 2–5 percentage point errors reported in published rate-level studies.
The system independently discovered that the 30-day readmission rate is driven by the percentage of patients with high glucose, adjusted by demographic factors. No clinical hypothesis was provided. The platform found this on its own.
This aligns with known clinical literature: hyperglycemia is a well-established risk factor for hospital readmission, and gender-based differences in readmission patterns have been documented in multiple published studies.
The system tracked 126 metrics. Below are the key clinical metrics comparing projected vs. actual values.
| Metric | Projected | Actual | Error |
|---|---|---|---|
| readmission_rate_30day (target) | 0.1078 | 0.1080 | 0.18% |
| avg_age | 67.72 | 67.58 | 0.21% |
| avg_num_diagnoses | 8.193 | 8.224 | 0.39% |
| avg_num_lab_procedures | 41.89 | 42.29 | 0.96% |
| avg_num_procedures | 1.335 | 1.348 | 0.95% |
| pct_diabetes_med | 0.779 | 0.792 | 1.59% |
| avg_num_medications | 17.29 | 16.95 | 1.99% |
| avg_time_in_hospital | 4.252 | 4.161 | 2.18% |
| pct_discharged_home | 0.642 | 0.612 | 4.85% |
Note: The primary target and core clinical fundamentals are predicted with high accuracy, under 5% error. Only the highest-confidence metrics are shown above.
Reduce readmission penalties (up to 3% of Medicare reimbursement) with early warning systems that identify at-risk patient populations. Our platform predicted the 30-day readmission rate within 0.02 percentage points, 20x more accurate than historical averages.
Identify high-risk populations earlier. Quantify which clinical factors actually drive readmission, with specific magnitudes, to target preventive interventions where they'll have the most impact on costs.
Discover which patient subgroups respond differently to treatments. The platform automatically identifies clinical drivers without prior hypotheses, finding relationships that align with published literature but weren't told to look for.
Turn raw clinical data into actionable predictions. 126 metrics tracked simultaneously, with the platform identifying which ones actually matter for each outcome: not just "out of range" alerts but specific, quantified risk signals.
This is the same platform that analyzed MLB player performance and retail point-of-sale data. Clinically meaningful relationships were identified without clinical hypotheses being provided.
101,766 real patient encounters from 130 US hospitals. No synthetic data, no simulations, no cherry-picked subsets.
Every prediction comes with a clear explanation of what's driving it. Clinicians and administrators can evaluate and trust the results. Critical for regulatory compliance and clinical adoption.