Predicting machine's performance record using the stacked long short-term memory (LSTM) neural networks.

Area of Science:

• Medical Physics
• Machine Learning
• Radiotherapy Technology

Background:

• Daily quality control (QC) data from radiotherapy machines reveal performance patterns and can indicate potential issues.
• Predictive modeling of QC data can enable early detection of machine failures and facilitate preventive maintenance.

Purpose of the Study:

• To develop a neural network model for quantitative prediction of radiotherapy machine quality control (QC) data records and trends.
• To assess the accuracy and robustness of the developed model in forecasting machine performance.

Main Methods:

• A stacked long short-term memory (LSTM) neural network model was developed using 3 years of daily QC records (867 data points) from a radiotherapy machine.
• The LSTM model was trained to predict QC data for the subsequent 5 days.
• Performance was compared against the autoregressive integrated moving average (ARIMA) model using Mean Absolute Error (MAE), Root-Mean-Square Error (RMSE), and Coefficient of Determination (R²).
• Model robustness was validated using QC data from a second radiotherapy machine.

Main Results:

• The stacked LSTM model demonstrated superior performance over the ARIMA model in predicting QC data for 24 items (LSTM: MAE=0.013, RMSE=0.020, R²=0.853 vs. ARIMA: MAE=0.021, RMSE=0.030, R²=0.618).
• The LSTM model also outperformed ARIMA for four QC items from a different machine (LSTM: MAE=0.102, RMSE=0.151, R²=0.770 vs. ARIMA: MAE=0.162, RMSE=0.375, R²=0.550).

Conclusions:

• The stacked LSTM model accurately predicts the record and trend of quality control (QC) items for radiotherapy machines.
• The model proved robust when applied to a different radiotherapy machine, indicating its generalizability.
• Accurate prediction of future machine performance allows for timely maintenance, thereby minimizing unscheduled downtime.