Page 123 - Demo

7A multiclass classification model for tooth removal procedurest121use for modern learning techniques. Further data collection is necessary to allow for the use of clinical metadata and to further increase the models’ performance and generalizability.The feature design and selection processes are an essential part of building a classification model. The evaluation of which features are most relevant for the algorithm to classify an experiment, is an important first step to improve our fundamental understanding of these complex procedures. Whilst a detailed discussion on the relevance of each feature falls outside the scope of this article, a few key findings are highlighted here. In terms of force and torque data, in each group of features the sum of forces and torques on all three axes combined were selected. This means that the sum of all forces and torques in an experiment is descriptive for classification purposes, rather than forces in individual directions. When looking at rotational and velocity data, features containing rotation around the Y-axis (buccoversion and/or palato/linguoversion) and around the axis of the tooth (Z-axis) were selected most frequently. This is in contrast to rotation around the X-axis (mesial and/or distal angulation) which was selected only once. These findings seem to correlate well with our clinical experience and seem in accordance with the limited available textbook instructions that mostly focus on rotations or movements around the longitudinal axis and buccolingual axis of a tooth [2]. Some of the selected features, on the other hand, are less well understood. For example, the selection of an average torque feature (mesial angulation) that does not match with an unselected rotational feature in the same direction. It might have to do with the position of the teeth, for example, a more mesial angulation is expected in dorsally located teeth, especially if a neighboring (mesial) tooth is absent. The latter has not been taken into account and these findings need additional analysis in future work.Due to the pioneering character of this study, no direct comparison is possible with previous literature. The available scientific literature on tooth removal procedures is surprisingly scarce and limited to the evaluation of exerted forces using a variety of methodologies and heterogeneous outcomes [3, 5, 6, 19, 20]. When this project started, many uncertainties in terms of achievability existed [7]. One of the most important downsides to our dataset and therefore our model, is that the data was collected ex vivo. Whilst the participating, experienced, oral- and maxillofacial surgeons considered the fresh frozen material as clinically representative, it is unknown in what way the freezing process influences the biomechanical properties of tooth removal. This Tom van Riet.indd 121 26-10-2023 11:59