Improving validation of Instadose dosimeters using machine learning algorithms

Dosimetry represents a key aspect in the radioprotection field for nuclear related activities: people working with ionising radiation need to monitor their exposure to be able to optimise their doses and to keep them below the dose limits. Such dosimeters are provided by approved dosimetry services, like SCK CEN. These dosimetry services need to prove that the doses they measure are correct. In this process, the validation of the personal dosimeter results is of major importance. The validation process ensures a reliable measurement of the personal dose equivalent and guarantees the correct operation of the dosimeter. In this project the focus is on Instadose dosimeters, which is a novel type of hybrid personal dosimeter. SCK nuclear research centre provides currently ~ 2000 Instadose dosimeters to different costumers, such as hospitals and companies. These Instadose dosimeters transfer the measured doses automatically to the dosimetry services once per week. Once per month a validation process is performed manually for each dosimeter by means of Microsoft Excel sheets. The latter contain information on the status of the device, details on the last measurements, and graphs with dose trends during the last months. In particular, these graphs permit an easy control of measurements that deviate from expected values or that present anomalies. Thanks to this modality, it is straightforward to classify these dosimeters in 10 different classes, based on the actions that have to be adopted for the dosimeter. This process is remarkably time consuming and the need to fasten it up is required since an increase of the number of these dosimeters is foreseen in the incoming years. In order to allow a faster validation, Machine Learning algorithms have been developed and tested in this project. The main idea is to identify all dosimeter results that are in line with the expected performances, for which no actions need to be taken. Although, the remaining results will still require manual evaluation, the automatic validation would reduce the workload significantly. The implementation process starts from a careful cleaning and rearranging of the data containing all the measurements of the previous year. The latter has been later used to create a unique database for training and testing of the models. Information collected inside this database are based on the validation process performed manually. Four different supervised classification models have been implemented: Decision Trees, Random Forest, K-Nearest Neighbours and Neural Networks. This first implementation was based on a binary classification approach to distinguish well-performing dosimeters from faulty ones. For each model, many combinations of hyperparameters have been implemented for the tuning of the algorithms. This approach has been followed in order to find a subset of combinations that can be easily applied on new incoming data, providing confidence intervals for the dosimeters classification. Although Decision Tree model for classification is rather straightforward, with respect to other developed models, it provides slightly better results concerning precision for binary classification. It is difficult to declare an overall best model based on the results obtained: small changes can be observed and no explicit correlations are always present. Therefore, a set of hyperparameter combinations has been identified for each model, in order to be implemented on new unseen data for monthly validations.