AI-Driven Myocardial Infarction Predictor
Discover our latest engineering deep dives, architectural patterns, and industry perspectives.
Overview
This retrospective cohort study was conducted at the Cardiology Unit of the National Hospital of Sri Lanka (NHSL), Colombo, in collaboration with Prime Technologies Global (Pvt) Ltd. Led by Principal Investigator Dr. Abiramy Surenthirarajah and supervised by Dr. Upul Dissanayake, Consultant Physician, the project developed and internally validated an AI-based predictive model using real-world inpatient data extracted from Bed Head Tickets (BHTs). AI model development was led by Eng. S. Kirushanth (Kiru Suren), who provided the technical expertise, model architecture, and computing resources.
The 12-month study analysed approximately 900 post-myocardial infarction (MI) patients (STEMI and NSTEMI, age ≥18 years) to identify key risk factors and accurately predict adverse inpatient outcomes, including prolonged hospital stay, major adverse cardiac events (MACE), and inpatient mortality, with the aim of supporting evidence-based clinical decision-making in Sri Lanka's resource-constrained public healthcare system.
The Challenge
Ischemic heart disease accounts for 13% of total deaths in Sri Lanka and 12–15% of deaths in government hospitals; according to the Annual Health Bulletin 2020, it was responsible for more than 52% of the 47,830 total hospital deaths recorded that year.
Noncommunicable diseases constitute 74.8% of all deaths nationally (WHO 2021), placing a severe and growing burden on public healthcare resources.
Post-MI inpatient management remains highly resource-intensive, frequently resulting in prolonged hospital stays, arrhythmias, heart failure, readmissions, and elevated healthcare costs.
Existing conventional risk scores (GRACE and TIMI) rely on limited clinical variables derived primarily from clinical trial populations, lack flexibility for real-world Sri Lankan demographics, and demonstrate reduced precision in routine clinical practice.
Our Approach
Data Collection & Preparation
Retrospective extraction of anonymized data from NHSL BHTs, targeting consecutive records from February 2025 backward until the required sample size of approximately 900 patients was reached, using strict inclusion/exclusion criteria.
Data Preprocessing & Feature Engineering
Comprehensive cleaning, imputation of missing values, normalization, one-hot encoding of categorical variables, outlier detection, and correction of class imbalance using SMOTE.
Model Development
Training of multiple machine learning algorithms (Random Forest, Gradient Boosting, Support Vector Machine, and Neural Networks) on a 70/20/10 train-validation-test split with k-fold cross-validation.
Hyperparameter Tuning & Optimization
Grid Search optimization combined with ensemble methods to enhance performance and minimize overfitting.
Model Interpretability & Validation
SHAP (SHapley Additive exPlanations) values for feature importance analysis and rigorous evaluation using AUROC, sensitivity, specificity, precision-recall curves, and calibration plots.
Use Cases
Case 01
Prolonged Hospital Stay (PLOS) Prediction
Early identification of patients likely to exceed the 75th percentile of length-of-stay distribution (≥5 days for STEMI and ≥3 days for NSTEMI) to optimize bed management and discharge planning.
Case 02
Major Adverse Cardiac Events (MACE) Risk Scoring
Real-time prediction of reinfarction, stroke, heart failure, cardiogenic shock, or urgent revascularization during hospitalization.
Case 03
Inpatient Mortality & ICU Escalation
Stratification of high-risk patients requiring immediate intensive monitoring and timely escalation of care.
Explore the Benefits
AI-Driven Complication Forecasting
Identifies high-risk post-MI patients using machine learning models validated on real-world clinical data.
Clinical Decision Support Intelligence
Surfaces explainable risk indicators for cardiology units, supporting evidence-based interventions in high-pressure environments.
Resource Allocation Optimization
Helps hospitals prioritize intensive care resources for high-risk individuals in resource-constrained public healthcare settings.
Evidence-Based Patient Care
Grounded in approximately 900 patient cases, providing a localized and highly accurate alternative to conventional risk scores.
The Impact
- 01.
Accurate prediction of prolonged hospital stay, major adverse cardiac events (MACE), and inpatient mortality.
- 02.
Identification of locally relevant predictors including age, comorbidities, admission physiology, and laboratory markers.
- 03.
Establishment of a scalable framework for integrating AI-based risk prediction into routine cardiology workflows.
- 04.
Contribution to the limited body of evidence on AI applications in cardiovascular risk modeling within South Asian and low-resource settings.
This study lays a strong foundation for prospective validation and potential national-scale deployment, demonstrating the practical value of enterprise AI data platforms in strengthening clinical decision-making and resource optimization in Sri Lanka's public healthcare system.
Project Results
Mortality Prediction AUROC
The model delivers highly accurate risk stratification validated against real-world clinical outcomes.
Clinical Record Depth
Deep analysis of locally-relevant predictors based on broad demographics and medical history.
Decision Support Utility
Real-time identification of high-risk patients optimized for resource-constrained cardiology units.
Keep Learning
Take a closer look at how we are helping to accelerate innovation and solve complex engineering challenges.