Predicting Breakdown Types of Biomedical Equipment in Public Hospital Using Machine Learning

Abstract

Predictive maintenance has emerged as a vital methodology in biomedical engineering to ensure the reliability of life-support equipment, particularly ventilators. This study presents a machine learning-based approach for predicting ventilator breakdown types using unstructured historical maintenance data. The dataset, consisting of 10,314 unscheduled maintenance records, was obtained from the Asset and Services Information System (ASIS), managed by the Ministry of Health Malaysia (KKM). The primary objective is to classify technician-written, free-text fault descriptions into structured breakdown categories to enable proactive maintenance decision-making. Four machine learning models, Support Vector Machine (SVM), Artificial Neural Network (ANN), Decision Tree (DT), and Random Forest (RF) were developed and evaluated under two workflows: one using basic label encoding, and the other incorporating Term Frequency–Inverse Document Frequency (TF-IDF) for semantic feature extraction. The results demonstrate that TF-IDF significantly enhanced classification performance across all models. SVM achieved the highest accuracy of 94.70%, followed by DT (94.56%), RF (94.14%), and ANN (92.19%). In contrast, the same models trained without TF-IDF performed poorly, with accuracies falling below 36% for SVM and ANN. These findings emphasize the importance of textual feature engineering in predictive maintenance applications and validate the effectiveness of AI-based text classification for unstructured biomedical data. By operationalizing real-world maintenance notes from clinical engineering workflows, this study demonstrates the practical viability of machine learning in improving equipment uptime, optimizing service scheduling, and enhancing patient safety in healthcare settings.