Page 85 - The-5th-MCAIT2021-eProceeding

Prediction Model of In-hospital Mortality Post Percutaneous

Coronary Intervention (PCI) Using Machine Learning
Technique

a,b
b
a*
Rosila Rebo , Afzan Adam , Azlan Hussin
a Center fro Artificial Intelligence Technology, Faculty of Information Science and Technology. National University of Malaysia
b Clinical Research Department, National Heart Institute, 145 Jalan Tun Razak, 50400 Kuala lumpur
*Email: afzan@ukm.edu.my

Abstract

The application and development of data science and machine learning plays a vital role in bringing improvisation,
innovation and transformation in medicine domain. This empowerment catalyzes the main goal of the study in developing
in hospital mortality prediction model for post-Percutaneous Coronary Intervention (PCI) as well as to determine the
significant mortality factors. PCI has evolved for four decades in enhancing its effectiveness in treating Coronary Heart
Disease (CHD). However post-procedure mortality still haunts the reputation of the modern medical world even at a very
low rate. With the growth of PCI National Heart Institute (IJN) data over decade involving 28407 procedures and embedded
with robust technology helped achieve this goal. The prediction model has been designed and structured into three tiers of
prediction derived from a complete dataset. The advantages of this tiers concept is allows to make an efficient prediction as
early as demographic phase and gradually to the intra-procedure phase and post-procedure phase. Thus, it serves as a support
medium for clinical decision making at each phase of prognosis. This study began with data exploration and preprocessing
that required approximately 80% of effort and time in term to produce high cleaned quality data. This study implemented
the filtration Information Gain Ranking for feature selection of the significant factors and sampling SMOTE technique to
overcome the problem of extreme imbalance dataset. The dataset was split into training (70%) and testing (30%) and 10-
fold cross validation as estimator. Gradient Boosted Decision Tree (GBDT), K Nearest Neighbor (KNN) and Artificial
Neural Network (ANN) models were developed. Parameter optimization was implemented at the learning rate for the GBDT
and ANN algorithms while the K parameter for the KNN algorithm. Systolic, complications and IABP were the factors with
highest information gain for demographic, intra-procedure and post-procedure datasets. Model with SMOTE showed
significantly better performance compared to imbalance dataset and no overfitting reported. Overall, GBDT prediction
model showed the best performance across demographic, intra-procedure and post-procedure phase then followed by KNN
and ANN.

Keywords: PCI prediction; inhospital mortality prediction

1. Introduction

Coronary Heart Disease (CHD) is a non-communicable disease and the most popular type of cardiovascular
disease due to its worldwide distribution. CHD contributes the highest death statistic worldwide. Malaysia also
rd
ranked 33 in the world with 29363 (23.1%) deaths in 2014. Critically, Malaysia continues to account a portion
of 15.8% from total worldwide of CHD death along with United States, Korea, Japan and others ASEAN country
in 2016 (DOSM 2019).
Treatment of CHD via Purcutaneous Coronary Intervention (PCI) has evolved over the past four decades
thus giving new hope to CHD patients (Canfield & Totary-Jain 2018). Generally, PCI helps to clear and widen
the artery blockage and promote normal blood circulation hence comfort the symptoms dan improve the

E- Proceedings of The 5th International Multi-Conference on Artificial Intelligence Technology (MCAIT 2021) [72]
Artificial Intelligence in the 4th Industrial Revolution