Evidence
Validation Studies
The Marqi Index methodology is documented in peer-reviewed manuscripts. Paper 1 is currently under submission at npj Digital Medicine (Nature Portfolio).
Paper 1: AMI 30-Day Readmission
Manuscript under submission at npj Digital Medicine (Nature Portfolio)
AUC on temporal validation
95% CI: 0.771 to 0.838
AMI patient cohort
TriNetX 2019-2024
AUC points above LACE/HOSPITAL
Published validation ranges
Paper 1 presents a 15-seed transformer ensemble with an XGBoost out-of-fold meta-learner for 30-day readmission prediction in acute myocardial infarction patients. The model was trained on 80% of the cohort (n=4,122) and validated on a temporally held-out 20% (n=1,031). All metrics reported are from the held-out validation set.
Validation Metrics
Discrimination
- AUC-ROC0.807 (0.771-0.838)
- Sensitivity at 80% specificity0.62
- Positive predictive value0.34 at top decile
Calibration
- Calibration intercept0.02
- Calibration slope0.98
- Brier score0.12
Methodology Overview
Model Architecture
15-seed transformer ensemble with an XGBoost out-of-fold meta-learner. Multiple random seeds provide training stability and enable uncertainty quantification. The meta-learner improves calibration and reduces variance across the ensemble.
Validation Strategy
Temporal validation with strict date-based split. Training cohort: admissions 2019-2022. Validation cohort: admissions 2023-2024. No temporal leakage. All metrics reported on held-out validation set only.
Data Sources
TriNetX federated research network. Structured EHR fields only. No clinical notes in Paper 1. Features include demographics, diagnoses (ICD-10), procedures (CPT), laboratory values (LOINC), medications (RxNorm), and utilization history.
Limitations Disclosure
We disclose limitations explicitly
- Single-condition validation (AMI only). Expansion to other HRRP conditions in progress.
- Internal validation only. External validation with academic medical center partners in progress.
- Structured EHR data only. Clinical note extraction is a v2 capability currently in research.
- Data integrity event caught and disclosed in the manuscript. Five-control verification protocol now standard.
Publication Roadmap
A transparent timeline for peer-reviewed validation across conditions and settings.
Paper 1: AMI 30-Day Readmission
30-day readmission prediction for acute myocardial infarction. 15-seed transformer ensemble. AUC 0.807 on temporal validation.
Paper 2: All-Cause Readmission
30-day all-cause readmission across HRRP and non-HRRP conditions. In active validation with academic medical center co-authors.
Paper 3: External Validation
Independent recomputation on validation partner data. Multi-site external validation with full calibration and discrimination metrics.
Paper 4: Clinical Note Integration
Cross-note longitudinal extraction for Marqi Index v2. Adding structured signals from clinical notes alongside existing structured EHR features.
Performance vs. Published Scores
How Marqi Index compares to LACE and HOSPITAL in peer-reviewed validation studies.
| Score | Published AUC Range | Method | Source |
|---|---|---|---|
| LACE | 0.62 - 0.70 | Logistic regression, 4 variables | van Walraven et al., 2010 |
| HOSPITAL | 0.65 - 0.73 | Logistic regression, 7 variables | Donzé et al., 2013 |
| Marqi Index | 0.807 (0.771-0.838) | 15-seed transformer + XGBoost | Paper 1 (in submission) |
LACE and HOSPITAL ranges from peer-reviewed validation studies in JAMA Internal Medicine, BMJ Open, and Journal of Hospital Medicine. Marqi Index from Paper 1 validation cohort.
Want the full evidence brief?
Schedule a 30-minute call to receive the Paper 1 evidence brief with full methodology, limitations, and calibration metrics.