A Fully Integrated Orthodontic Aligner With Force Sensing Ability for Machine Learning-Assisted Diagnosis

Abstract

Currently, the diagnosis of malocclusion is a highly demanding process involving complicated examinations of the dental occlusion, which increases the demand for innovative tools for occlusal data monitoring. Nevertheless, continuous wireless monitoring within the oral cavity is challenging due to limitations in sampling and device size. In this study, by embedding high-performance piezoelectric sensors into the occlusal surfaces using flexible printed circuits, a fully integrated, flexible, and self-contained transparent aligner is developed. This aligner exhibits excellent sensitivity for occlusal force detection, with a broad detection threshold and continuous pressure monitoring ability at eight distinct sites. Integrated with machine learning algorithm, this fully integrated aligner can also identify and track adverse oral habits that can cause/exacerbate malocclusion, such as lip biting, thumb sucking, and teeth grinding. This system achieved 95% accuracy in determining malocclusion types by analyzing occlusal data from over 1400 malocclusion models. This fully-integrated sensing system, with wireless monitoring and machine learning processing, marks a significant advancement in the development of intraoral wearable sensors. Moreover, it can also facilitate remote orthodontic monitoring and evaluation, offering a new avenue for effective orthodontic care.

Keywords: electrospinning; force sensing; malocclusion diagnosis; piezoelectric nanogenerator; wearable devices.

Figure 1

Overview of the fully integrated ARIA system a) Schematic diagram of ARIA featuring the piezoelectric sensor array on the occlusal surface of the teeth, printed circuit board, and battery. The exploded view of the board (bottom left) shows the analog front‐end, analog‐to‐digital converter, and microcontroller unit components. b) Structural scheme of a 6‐layer composite flexible piezoelectric sensor. c) A block diagram showing the workflow of ARIA, including signal processing, control, communication, and display. d) Closed‐mouth and smiling photographs of a subject wearing ARIA. The inset displays an intraoral photo of the maxillary teeth fitted with ARIA.

Figure 2

Design and characterization of high‐performance piezoelectric sensors a) Schematic representation of the preparation of the hPTM fibers and its TEM image showing MXene nanosheets embedded in the PVDF matrix (scale bar: 2 µm). b) The fluorine atom concentration distribution of PVDF‐TrFE at 3 ns from MD simulations is depicted, with the red curve representing the results after the introduction of PVP. The inset shows the distribution of PVDF‐TrFE molecular chains at this time, with the red regions indicating PVDF‐TrFE, and PVP located above these regions. c) Results of FE analysis of the deformation of PVDF‐TrFE and hPTM fibers under the pressure applied by two parallel plates. d) Comparison of piezoelectric fibrous membranes across six performance metrics: elongation at break, flexibility, output voltage, output current, volume charge coefficient, and piezoelectric power density. The red sections represent the performance of hPTM‐ENMs. e,f) Open‐circuit voltage of hPTM‐ENMs at different frequencies e) and loads f). g) Charge voltage diagrams of electronics for 0.1 and 4.7 µF capacitors. h) Comparison of the piezoelectric output of hPTM‐ENMs with previously developed PENGs. i) Long‐term stability testing of hPTM‐ENMs. j) Power variations of hPTM‐ENMs at increasing load resistance.

Figure 3

Evaluating ARIA for monitoring occlusal force and bad oral habits during daily activities. a) Schematic diagram of the open and closed mouth fitted with ARIA. b) Schematic diagram of the four stages of charge transfer during the tooth contact‐separation cycle of an occlusal bite. c) The ARIA signals can be transferred to PC terminals and smartphones via Bluetooth, which then process and display them on an interactive interface. d) Photographs of the subjects wearing ARIA and performing CO, LB, TS, and GT, and the corresponding voltage curve generated during these activities. e) Control experiment conducted using five healthy subjects to assess the reliability of ARIA in monitoring bad oral habits. f) Correlation heatmap of data from five subjects testing CO, LB, TS, and GT occlusion conditions. g) Cluster analysis of data obtained from five subjects testing CO, LB, TS, and GT after PCA.

Figure 4

Utilizing ARIA to collect occlusal data from various malocclusions in a large clinical sample a) Overview of employing ARIA for the collection of 1467 clinical samples and extraction of comprehensive data. b,c) Sensor signals were acquired from five different malocclusion models wearing ARIA on the maxilla b) and mandible c). d,e) t‐SNE plots showing mandibular d) and maxillary e) feature separation in a 3D space. f) Feature vector matrices of the maxillary and mandibular structures simultaneously integrated using the t‐SNE algorithm.

Figure 5

Classification of malocclusions based on ML a) Confusion matrix showing the accuracy of the XGBoost ML model in predicting the malocclusion in the test set. b) ROC curve of the XGBoost ML model for predicting five types of malocclusions. c) Correlation heatmap between 16 features of maxilla and mandible. d) Chord diagram showing the relative correlation between different teeth. e) Sankey diagram of SHAP analysis depicting the relative contribution of different teeth to malocclusion classification. f) SHAP decision plot of the ML model predicting different malocclusions using maxillary and mandibular data. g) Stacked bar plot of feature importance showing the contribution of each tooth to each malocclusion type. h) SHAP summary plot of the XGBoost ML model based on the dataset collected by ARIA.