Anomaly detection of screw loosening of different implant-abutment connection designs using resonance frequency...

Screw loosening is a common mechanical complication in implant-supported restorations, yet early detection remains challenging because of the lack of noninvasive, quantitative assessment tools.

This in vitro study aimed to evaluate the ability of resonance frequency analysis (RFA) combined with machine learning techniques to detect screw loosening in 3 implant-abutment connection designs: conical, internal tri-channel, and external hexagonal.

Three connection types were tested using 6 implants inserted into mandibular jaw models. RFA was conducted using a triaxial accelerometer under 7 loosening conditions with the frequency responses in the x, y, and z directions being recorded. Supervised learning (random forest) was used for binary (intact versus abnormal) and multiclass (7 classes) classification. Unsupervised learning was used to detect anomalies.

The RFA revealed unique vibrational profiles for each connection design. The random forest classifier achieved 100% accuracy in binary classification and 83% to 92% in multiclass classification. Feature importance analysis highlighted the relevance of the triaxial vibration data. The convolutional autoencoder consistently identified screw loosening as early as 5 degrees.

RFA combined with machine learning reliably detected screw loosening across various types of implant-abutment connections. This method offers a noninvasive, quantitative tool for the early detection of mechanical complications in implant dentistry.