Advanced Clinical Data Processing: A Predictive Maintenance Model for Anesthesia Machines

Sammanfattning: The maintenance of medical devices is of great importance to ensure that the devices are stable, well-functioning, and safe to use. The current method of maintenance, which is called preventive maintenance, has its advantages but can be problematic both from an operators’ and a manufacturers’ side. Developing a model that will predict failure in anesthesia machines can be of great use for the manufacturer, the customers, and the patients. This thesis sets to examine the possibility of creating a predictive maintenance model for anesthesia machines by utilizing device data and machine learning. This thesis also investigates the influence of the data on the model performance and compare different lag sizes and future horizons to model performance. The time-series data collected came from 87 unique devices and a specific test was chosen to be the output variable of the model. A whole pipeline was created, which included pre-processing of the data, feature engineering, and model development. Feature extraction was done on the time series data, with the help of a library called tsfresh, which transformed time series characteristics into features that would enable supervised learning. Two models were developed: logistic regression and XGBoost. The logistic regression model acted as a baseline model and the result of its performance was as expected, quite poor. The XGBoost yielded an AUCPR score of 0.21 on the full dataset and 0.32 on a downsampled dataset. Although a quite low score, it was surprisingly high considering the extreme class imbalance that existed in the dataset. No clear pattern was found between the lag sizes and future horizons with the model performance. Something that could be seen was that the data imbalance had a great impact on the model performance, which was discovered when the downsampled dataset with less class imbalance yielded a higher AUCPR score.