Abstract
Accurately predicting mortality following acute myocardial infarction (AMI) is essential for optimizing patient care. Traditional risk-scoring systems often require manual calculation and specific clinical variables that may not be readily available in electronic health records (EHR). This study leverages machine learning to automate one-year mortality prediction for ICU patients using the MIMIC-III database. By evaluating various shallow algorithms and a deep feedforward neural network across segmented feature sets, we demonstrate that a combined clinical data approach achieves a peak AUC of 0.901. Our findings highlight the potential for EHR-integrated models to provide rapid, high-accuracy risk assessments in clinical settings.
Introduction
Acute myocardial infarction (AMI) remains a leading cause of global mortality and hospital readmission. For clinicians, identifying high-risk patients early is vital for tailoring interventions and improving long-term outcomes. While established risk scores like TIMI and GRACE are widely used, they often necessitate manual data entry and rely on variables—such as Killip class—that are not consistently captured in structured EHR formats.
This research addresses the need for automated, data-driven prognostic tools. By utilizing the rich longitudinal data within EHRs, we aim to build computational models capable of predicting one-year mortality without the need for manual clinician intervention.
Methodology
The study frames mortality prediction as a binary classification problem: predicting whether a patient survives or dies within one year of ICU admission.
Algorithms
We compared two primary approaches:
- Shallow Learning: Utilizing the WEKA workbench, we tested algorithms including Logistic Model Trees (LMT), Simple Logistic, Random Forest, and J48 decision trees.
- Deep Learning: A Deep Feedforward Neural Network (Deep FNN) was implemented in TensorFlow, featuring two hidden layers, ReLU activations, and softmax cross-entropy optimization.
Feature Engineering
To identify the most predictive clinical indicators, we segmented 79 extracted variables into five distinct datasets: Admission Information, Demographics, Treatment Information, Diagnostic Information, and Lab & Chart Values. A sixth "Combined" dataset aggregated all features to test holistic predictive power. All models were validated using 10-fold cross-validation on a cohort of 5,436 ICU admissions from the MIMIC-III database.
Results
The performance of the models varied significantly based on the feature set used. The "Combined" dataset consistently outperformed individual segments, underscoring the multi-factorial nature of AMI prognosis.
The shallow models, specifically LMT and Simple Logistic, achieved the highest overall accuracy (85.12%) and AUC (0.901) when applied to the combined dataset. Interestingly, admission information alone proved to be a strong predictor (AUC 0.833), while demographics and comorbidities were less effective in isolation.
The Accuracy-Recall Trade-off
A critical finding was the performance difference between shallow and deep models. While LMT provided superior overall accuracy, it suffered from low recall (0.594), failing to identify a significant portion of the patients who died. In contrast, the Deep FNN achieved a much higher recall of 0.820, albeit with slightly lower overall accuracy (82.02%). In a clinical context, the higher sensitivity of the deep learning model may be preferable to avoid missing high-risk patients.
Discussion and Conclusion
Our results demonstrate that machine learning models can effectively predict post-AMI mortality using standard EHR data, outperforming several previously published benchmarks. Key insights include the critical importance of rapid triage; patients whose initial ER diagnosis correctly identified AMI showed significantly higher survival rates. Furthermore, demographic analysis revealed higher mortality rates among female patients, suggesting a need for further research into gender-specific symptom presentation and treatment delays.
While limited by class imbalance and missing data typical of real-world EHRs, this study provides a foundation for automated risk stratification. Future work will explore advanced architectures like LSTMs to better capture the temporal dynamics of patient health and integrate unstructured clinical notes to further refine predictive power.
Summary updated