Medical Device Failure Predictions Through AI-Driven Analysis of Multimodal Maintenance Records

Abstract

Medical device failure and maintenance records are essential information, as some nations lack dedicated systems for capturing this valuable data. In addition to making healthcare more intelligent and individualized, machine learning has the potential to transform the conventional healthcare system. Optimizing AI models in decision-making could mitigate the effects of three research issues: malfunctioning medical devices, high maintenance costs, and the lack of a strategic maintenance framework. This study proposes a data-driven machine-learning model for predicting medical device failure. The proposed predictive model is developed using multimodal data of structured maintenance and unstructured text narrative of maintenance reports to predict the failure of 8,294 critical medical devices. In developing the model, 44 varieties of essential medical devices from 15 healthcare institutions in Malaysia are utilized. A classification problem is addressed by classifying failure into three prediction classes: (i) class 1, unlikely to fail within the first three years, (ii) class 2, likely to fail within three years; and (iii) class 3, likely to fail after three years from the date of commissioning. The topic modelling and synthesis strategy: Latent Dirichlet Allocation is applied to unstructured data in order to uncover concealed patterns in maintenance notes captured during failures. In addition, sensitivity analysis is performed to select only the most significant parameters affecting the failure performance of the medical device. Then, four machine learning algorithms and three deep learning networks are evaluated to determine the best predictive model. Based on the performance evaluation, the Ensemble Classifier is further optimized and demonstrates improved accuracy of 88.80%, specificity of 94.41%, recall of 88.82%, precision of 88.46%, and F1 Score of 88.84%. The study proves a reduction in intervention from 18 to 8 features and a reduction in training time from 1660.5 to 901.66 seconds for comprehensive model development.