The model's figures are not invented — and they don't work in just one place. Trained and validated in Latin America on ~3.94 million clinical records from 382,589 patients, and further confirmed on NIH All of Us (626,396 US patients): calibration and the ability to rank risk hold up in a different population, without retraining. All under the TRIPOD+AI protocol, with the derivation paper published in IJCRP 2026 (open access).
Corpus AI was trained on ~3.94 million clinical records from 382,589 Latin American patients (Colombia, 1990–2025) and externally validated in Latin America on an insurer never seen during training. An additional independent external validation in the NIH All of Us program confirms that calibration and risk-ranking hold in the US too — the same model, no retraining.
Pre-specified protocol frozen before execution (TRIPOD+AI style). Validation inside the secure All of Us Researcher Workbench; only aggregate results were exported (groups ≥ 20 people), in line with its privacy policies.
A complete cardio-metabolic-renal risk suite: five canonical models delivered —AMI, stroke, hypertension, DM and CKD— plus mental health. Pick a model and see its evidence. Hold-out figures with 95% bootstrap CI (n=1000, seed 2026), verified calibration (ECE) and SHA-256 traceability; AUC ranks, but what matters is that the probability is real.
The flagship model: derived and validated on a cardiovascular cohort of 382,589 patients, published in IJCRP 2026 and externally validated on All of Us. It is the only one in formal production — +241 AMIs prevented in a real deployment.
Amorocho-Morales JD, Parra-Guevara S, Quintero-Muñoz E, Dimas G, Correa-Morales JE. Int J Cardiol Cardiovasc Risk Prev (IJCRP). 2026.
Vol. 30, art. 200659
We present a predictive model for acute myocardial infarction trained and validated on a Latin American cardiovascular cohort of 382,589 unique patients followed between 1990 and 2025, with 3,940,059 clinical encounters and 15,511 observed AMIs. The model operates on routine clinical data, no advanced imaging, no special biomarkers, and reports near-perfect calibration (O/E 0.998), robust discrimination (AUC 0.869, C-index 6m 0.836, C-index 12m 0.846), and significant external validation on an insurer never seen during training (n=5,602, F=147.6, p<.001). The separation between operating bands (observed rate 54.6% / 38.1% / 21.0%) shows that calibration survives the population shift. The reporting follows the TRIPOD+AI standard (BMJ 2024).
The model was trained on real clinical practice, not on a trial. That is what lets the probabilities reflect the population they are applied to.
Deduplicated identifiers across the 35-year window.
Visits, hospitalizations, and events recorded in the system.
Confirmed AMI events (coding + clinical verification).
Insurer never seen during training · F = 147.6 · p < .001.
The model's metrics in academic form with citations. Computed on the full cohort and reproduced in the external validation.
O/E (observed over expected) compares how many events actually occurred against how many the model predicted. A value near 1.0 means the model's probabilities are numerically correct: when it says 22% risk at 12 months, the patient really does have a 22% chance of AMI in the actual cohort, not 5%, not 60%. That is what lets you use the probability as a clinical threshold, share it with the patient, and define operating bands. Discrimination (AUC, C-index) tells you the order is right; calibration tells you the numbers are real.
External validation on an insurer never seen during training. The separation between bands holds, calibration survives the population shift.
5,602
Patients
147.6
F statistic
<.001
p value
Classic models (Framingham, SCORE2, PROCAM, PREVENT) measure 10-year risk on European or American cohorts. Applied as-is in LATAM they overpredict: O/E < 1. The system ends up prioritizing people who won't have the event.
Systematic LATAM pattern: Framingham and SCORE2 overpredict (O/E < 1). Sex-adjusted PROCAM partially closes the calibration gap in Colombia. Caritas calibrates almost perfectly on the real cohort.
The best discriminator and best capture in the suite: the top 10% by risk concentrates 58.5% of strokes, nearly six times the base rate.
Predicts incident-hypertension risk so you can intervene before it sets in.
Predicts incident —not prevalent— glycemic dysregulation, with a clean label and no data leakage. Internally validated and on its way to production.
Under construction with academic cohorts in the US and Europe.
Renal progression. In the top 10% by risk the positive predictive value is 81%: of every 10 flagged patients, ~8 progress.
In mental health we start from validated, published screening instruments and peer-reviewed risk models. We are recalibrating them to the local population before reporting probabilities — true to our philosophy: first make the number real.
The instruments are validated and published; the models are peer-reviewed and under local recalibration. We do not report a production probability in mental health until that calibration is complete.
Muñoz OM, Rodríguez NI, Ruiz Á, Rondón M. Validación de los modelos de predicción de Framingham y PROCAM en una población colombiana. Rev Colomb Cardiol. 2014;21(4):202–212.
paperHageman SHJ, McKay AJ, et al.. SMART2 risk prediction algorithm. Eur Heart J. 2022;43(18):1715–1727.
paperMancini GBJ, Ryomoto A. Adoption of the PREVENT Risk Algorithm: Potential International Implications. JACC Adv. 2024;3(8):101122.
paperScheuermann B, Brown A, et al.. External Validation of the AHA PREVENT Cardiovascular Disease Risk Equations. JAMA Netw Open. 2024;7(10):e2438311.
paperWHO CVD Risk Chart Working Group. WHO CVD risk charts: revised models for 21 global regions. Lancet Glob Health. 2019;7(10):e1332–e1345.
guidelineCollins GS, Moons KGM, et al.. TRIPOD+AI statement. BMJ. 2024;385:e078378.
standardLiu T, Krentz A, Lu L, Curcin V. ML-based prediction models for CVD risk using EHR data: systematic review and meta-analysis. Eur Heart J Digit Health. 2024;6(1):7–22.
paperDamen JA, Pajouheshnia R, et al.. Performance of the Framingham risk models and pooled cohort equations. BMC Med. 2019;17(1):109.
paperAmorocho-Morales JD, Parra-Guevara S, Quintero-Muñoz E, Dimas G, Correa-Morales JE. A next-generation prediction risk model for acute myocardial infarction: Derivation and validation in a multi-centre cohort. Int J Cardiol Cardiovasc Risk Prev. 2026;30:200659. doi:10.1016/j.ijcrp.2026.200659.
paperKrittanawong C, Virk HUH, et al.. Machine learning prediction in cardiovascular diseases: meta-analysis. Sci Rep. 2020;10(1):16057.
paperKroenke K, Spitzer RL, Williams JB. The PHQ-9: validity of a brief depression severity measure. J Gen Intern Med. 2001;16(9):606–613.
DOI →Cassiani-Miranda CA, Cuadros-Cruz AK, et al.. Validity of the Patient Health Questionnaire-9 (PHQ-9) for depression screening in Colombia. Rev Colomb Psiquiatr. 2020.
DOI →Spitzer RL, Kroenke K, Williams JB, Löwe B. A brief measure for assessing generalized anxiety disorder: the GAD-7. Arch Intern Med. 2006;166(10):1092–1097.
DOI →García-Campayo J, Zamorano E, et al.. Cultural adaptation into Spanish of the GAD-7. Health Qual Life Outcomes. 2010;8:8.
DOI →Remor E. Psychometric properties of a European Spanish version of the Perceived Stress Scale (PSS). Span J Psychol. 2006;9(1):86–93.
DOI →Molinero Ruiz E, Basart Gómez-Quintero H, Moncada Lluis S. Validation of the Spanish version of the Copenhagen Burnout Inventory (CBI). Rev Esp Salud Publica. 2013;87(2):165–179.
DOI →King M, Walker C, et al.. Development and validation of an international risk prediction algorithm for episodes of major depression (predictD). Arch Gen Psychiatry. 2008;65(12):1368–1376.
DOI →Moreno-Peral P, Luna JD, et al.. Predicting the onset of anxiety syndromes at 12 months in primary care (predictA). PLOS ONE. 2014;9(9):e106370.
DOI →Fernandez A, Salvador-Carulla L, et al.. Development and validation of a prediction algorithm for the onset of common mental disorders (HILDA). Aust N Z J Psychiatry. 2018;52(1):47–58.
DOI →Inoue S, et al.. Predicting mental health–related sick leave using stress check data. J Occup Health. 2026;68(1):uiag011.
DOI →30 minutes with the clinical-technical team. We walk through the published paper, discuss the calibration methodology, and explore what the model would look like on your own cohort. We respond within 2 business days.
CORPUS