Development and Validation of a Predictive Scoring System for Medical Interventions and ICU Admission in Cardiac Emergency Department Patients

Aim and Research Question(s)

This thesis aimed to develop and evaluate a machine learning–based scoring system for early risk stratification of patients with suspected cardiac emergencies using routinely available admission data.

• Which routinely collected clinical parameters are most strongly associated with the need for catheterization within 24h after admission?
• Can early emergency department data be used to predict admission to an intensive care unit within 24h?
• Is it possible to derive a simple and interpretable predictive score that provides clinically meaningful decision support?

Background

Cardiovascular emergencies require rapid decision-making under time pressure and limited information. Early identification of high-risk patients is crucial for timely intervention and efficient resource allocation. Emergency departments manage heterogeneous cardiological populations, while traditional clinical scores often rely on fixed thresholds and predefined variables. Machine learning approaches may complement established scores by integrating routinely collected parameters and modeling complex, non-linear relationships to provide outcome-specific risk estimates early in care.[1]

Methods

Retrospective, monocentric observational study based on anonymized routine clinical data from 422 adult patients presenting with suspected cardiological emergencies.

Predictor variables were restricted to information available at hospital admission (demographics, vital signs, clinical presentation, ECG findings, laboratory parameters).

Two outcome-specific Random Forest models were developed to predict
• ICU admission within 24 hours
• Cardiac catheterization within 24 hours

Model performance was assessed using standard classification metrics and feature importance analysis.

Results and Discussion

The ICU admission model achieved an AUC of 0.94 (sensitivity 0.83, specificity 1.00), correctly identifying 5 of 6 ICU admissions. Results should be interpreted cautiously due to the low event rate. The cardiac catheterization model showed excellent performance (AUC 0.98, accuracy 0.96, sensitivity 0.86, specificity 0.997), correctly identifying 86 of 100 patients requiring early invasive management. Key predictors were clinically plausible. Troponin, acute coronary syndrome, and NT-proBNP dominated catheterization prediction, while age, oxygen saturation, hemodynamic instability, and arrhythmias were most relevant for ICU admission. Overall, the outcome-specific Random Forest models demonstrated higher apparent discriminatory performance than commonly reported clinical scores.[2]

Conclusion

Routinely available emergency department data can be used to develop interpretable machine learning–based models for early cardiological risk stratification. The outcome-specific Random Forest models demonstrated good to excellent discriminatory performance, particularly for predicting early cardiac catheterization, highlighting the potential of data-driven decision support tools to complement clinical judgment in acute cardiology.

References

[1] Christodoulou E, et al. BMJ. 2019. [2] Poldervaart JM, et al. Int J Cardiol. 2017.