A deep LSTM autoencoder-based framework for predictive maintenance of a proton radiotherapy delivery system

Unscheduled machine downtime can cause treatment interruptions and adversely impact patient treatment outcomes. Conventional Quality Assurance (QA) programs of a proton Pencil Beam Scanning (PBS) system ensure its operational performance by keeping the beam parameters within clinical tolerances but often do not reveal the underlying issues of the device prior to a machine malfunction event. In this study, we propose a Predictive Maintenance (PdM) approach that leverages an advanced analytical tool built on a deep neural network to detect treatment delivery machine issues early.Beam delivery log file data from daily QA performed at the Burr Proton Center of Massachusetts General Hospital were collected. A novel PdM framework consisting of long short-term memory-based autoencoder (LSTM-AE) modeling of the proton PBS delivery system and a Mahalanobis distance-based error metric evaluation was constructed to detect rare anomalous machine events. These included QA beam pauses, clinical operational issues, and treatment interruptions. The model was trained in an unsupervised fashion on the QA data of normal sessions so that the model learned characteristics of normal machine operation. The anomaly is quantified as the multivariate deviation between the model predicted data and the measured data of the day using Mahalanobis distance (M-Score). Two-layer and three-layer Long short-term memory-based stacked autoencoder (LSTM-SAE) models were optimized for exploring model performance improvement. Model validation was performed with two clinical datasets and was analyzed using the area under the precision-recall curve (AUPRC) and the area under the receiver operating characteristic (AUROC).LSTM-SAE models showed strong performance in predicting QA beam pauses for both clinical validation datasets. Despite severe skew in the dataset, the model achieved AUPRC of 0.60 and 0.82 and AUROC of 0.75 and 0.92 in the respective 2018 and 2020 datasets. Moreover, these amount to 2.8-fold and 10.7-fold enhancement compared to the respective baseline event rates. In addition, in terms of treatment interruption events, model prediction enabled 3.88-fold and 51.2-fold detection improvement, while the detection improvement for clinical operational issues was 1.04-fold and 1.37-fold, respectively, in the 2018 and 2020 datasets.Our novel deep LSTM-SAE-based framework allows for highly discriminative prediction of anomalous machine events and demonstrates great promise for enabling PdM for proton PBS beam delivery.