An EHR Triage Score for Cancer Patients With ACS

Predictive Model DevelopmentPredictive External Validation

Authors:Florian A Wenzl; Kok Weng Ow; Matthijs A Velders, et al.

Journal:The Lancet

DOI:10.1016/S0140-6736(25)02020-3

Cancer patients with ACS face simultaneous risks of death, bleeding, and recurrent ischaemia—yet most risk tools exclude them. An 11-variable competing-risks XGBoost score (ONCO-ACS) externally validated across England, Sweden, and Switzerland achieved 6-month tAUC ~0.84 for mortality (0.67–0.74 for bleeding), enabling EHR triage to target pharmacist/cardio-oncology review and tailor DAPT decisions.

January 31, 2026

Quick Take

ONCO-ACS is an 11-variable machine-learning score for patients with cancer presenting with acute coronary syndrome (ACS) that predicts 6‑month all-cause mortality, major bleeding, and ischaemic events using a competing‑risks XGBoost framework. Key 6‑month time‑dependent AUCs (tAUC) on internal validation were: mortality 0.84 (0.83–0.85), major bleeding 0.70 (0.68–0.73), and ischaemic events 0.79 (0.78–0.81); observed 6‑month event rates in the development cohort were mortality 27.8% (95% CI 27.3–28.3), major bleeding 7.3% (7.0–7.5), and ischaemic events 16.1% (15.7–16.4).

Designed for EHR integration (an online calculator is provided), ONCO-ACS can serve as a triage dashboard to prioritise pharmacist and cardio‑oncology review for antithrombotic decisions (DAPT intensity and duration) and to flag cases where bleeding versus ischaemia trade‑offs may warrant early multidisciplinary input — but local validation and reliable oncology data capture are prerequisites for safe deployment.

Why it Matters

Patients with cancer who present with ACS face competing, high‑magnitude short‑term risks — mortality, major bleeding, and recurrent ischaemia — which complicate decisions about invasive strategy and antithrombotic therapy.

Major ACS trials and common risk tools systematically excluded or under‑represented cancer patients, so clinicians and pharmacists often extrapolate from non‑cancer populations, producing inconsistent practice and difficult bedside trade‑offs.

A standardised, scalable way to quantify these three competing 6‑month risks could enable CDS‑supported prioritisation of scarce cardio‑oncology and pharmacy resources and help standardise stewardship of antithrombotic therapy.

What They Did

Assembled linked, multisource ACS cohorts from England, Sweden, and Switzerland (2004–2023) and identified patients with a cancer diagnosis within 5 years of presentation for inclusion.

Trained XGBoost models in a competing‑risks framework to predict 6‑month all‑cause mortality, major bleeding, and a composite ischaemic endpoint from admission data using a single parsimonious set of 11 routinely available variables: tumour type, time since cancer diagnosis, metastatic disease, age, haemoglobin, heart rate, estimated glomerular filtration rate (eGFR), body‑mass index (BMI), Killip class, cardiac arrest, and major bleed within the prior 6 months.

Developed models in an English training cohort (excluding the Midlands) and externally validated them in held‑out English Midlands, national Swedish, and a smaller adjudicated Swiss cohort; used SHAP for per‑patient explainability and reported calibration and decision‑curve analyses.

What They Found

In the development cohort cancer patients with ACS had high 6‑month cumulative incidences: all‑cause mortality 27.8% (95% CI 27.3–28.3), major bleeding 7.3% (7.0–7.5), and ischaemic events 16.1% (15.7–16.4).

ONCO‑ACS discrimination was high for mortality (6‑month tAUC 0.84 [0.83–0.85] internal), good for ischaemia (tAUC 0.79 [0.78–0.81] internal), and moderate for major bleeding (tAUC 0.70 [0.68–0.73] internal). External validation showed generally similar performance across cohorts (mortality ~0.80–0.84; major bleeding ~0.67–0.74; ischaemia ~0.70–0.76).

Risk stratification concentrated events: 52.3% of cancer patients were classified high/very high for mortality, 63.1% for major bleeding, and 61.6% for ischaemia; compared with the low‑risk group the very‑high group had large effect sizes (mortality HR 90.4 [72.2–113.2]; bleeding subdistribution HR 7.8 [5.5–11.1]; ischaemia subdistribution HR 20.5 [17.2–24.3]).

Mapping ONCO‑ACS to guideline cutoffs suggested most patients (79.1%) would be candidates for invasive management and long DAPT (clopidogrel), while 20.9% might be considered for conservative care; combined profiling identified small actionable subgroups (5.4% high‑bleed/low‑ischaemia; 3.3% high‑ischaemia/low‑bleed).

ONCO‑ACS outperformed established non‑cancer scores (GRACE for mortality, PRECISE‑DAPT for bleeding, PARIS for ischaemia) on discrimination and decision‑curve analyses and showed generally good calibration across the risk spectrum and key subgroups.

Takeaways

Treat ONCO‑ACS as a three‑dial dashboard (mortality, bleeding, ischaemia) rather than a single composite verdict — clinical decisions in cancer+ACS should balance all three competing 6‑month risks.

For inpatient pharmacy, consider integrating ONCO‑ACS into the EHR to automate triage: use it to flag Very‑High and High risk patients for early pharmacist or cardio‑oncology review to inform DAPT intensity, duration, and invasive versus conservative strategy — but only after local validation and ensuring oncology fields are reliably captured.

Do not over‑interpret the bleeding prediction: its discrimination was only moderate externally (tAUC ≈0.67–0.74). Use bleeding risk outputs for prioritisation and hypothesis‑generation, and verify flagged high‑bleed risk patients with bedside data (platelet count, recent procedures, active bleeding, current cancer therapy) before changing therapy.

Before local adoption confirm structured capture of tumour type, metastatic status, and time since diagnosis in your EHR; use SHAP explainability outputs as audit/triage aids rather than as sole justification for treatment changes.

Strengths and Limitations

Strengths:

Very large, multisource development dataset with geographically distinct external validation (English Midlands, Sweden, Switzerland) and a deliberately parsimonious 11‑variable predictor set compatible with routine EHR data.

Methodological transparency and rigour: competing‑risks XGBoost models, per‑patient SHAP explainability, calibration checks, decision‑curve analyses, subgroup and temporal transportability assessments, and adherence to reporting standards.

Limitations:

Limited oncological granularity and missing bedside laboratory data across cohorts — no systematic histology/molecular subtype, limited cancer therapy regimen data, and inconsistent platelet counts (important for inpatient antithrombotic decisions).

Observational registry design with predominantly European cohorts and limited ethnic diversity, a relatively small Swiss adjudicated cohort, potential data capture variability across systems, and no prospective randomized or implementation impact evaluation to date.

Bottom Line

ONCO‑ACS is a validated, EHR‑friendly triage dashboard that improves discrimination for 6‑month mortality, bleeding, and ischaemic risk in patients with cancer and ACS compared with non‑cancer tools; it can support prioritisation of pharmacy review and personalised antithrombotic trade‑offs, but requires reliable oncology data capture, local validation, and prospective impact monitoring before broad operational rollout.