A novel concept of an acoustic ultrasound wearable for early detection of implant failure

Abstract

Mechanical failure of medical implants, especially in orthopedic poses a significant burden to the patients and healthcare system. The majority of the implant failures are diagnosed at very late stages and are of mechanical causes. This makes the diagnosis and screening of implant failure very challenging. There have been several attempts for development of new implants and screening methods to address this issue; however, the majority of these methods focus on development of new implants or material and cannot satisfy the needs of the patients that have already been operated on. In this work we are introducing a novel screening method and investigate the feasibility of using low-intensity, low-frequency ultrasound acoustic waves for understanding of interfacial implant defects through computational simulation. In this method, we simultaneously apply and sense acoustic waves. COMSOL simulations proved the correlation between implant health condition, severity, and location of defects with measured acoustic signal. Moreover, we show that machine learning not only can detect and classify failure types, it can also assess the severity of the defects. We believe that this work can be used as a proof of concept to rationalize the development of non-invasive screening acoustic wearables for early detection of implant failure in patients with orthopedic implants.

Fig. 1

(A) The wearable acoustic cuffs tightened around the target joint. (B) the cuff signals are recorded and conditioned to form dataset through data augmentation. (C) The processed data train predictive model to predict existence, severity, and location of defects for the clinician.

Fig. 2

(A) Simulated displacement field when the transducer #1 is actuated at 300 kHz for healthy (A₁), cracked (A₂), and loose (A₃) interfaces. (B) Magnified displacement field distribution at the interface of the implant. (C) Simulated displacement spectrum of transducer #5 when the transducer #1 is actuated for a healthy interface, with fracture and, with loosening scenarios.

Fig. 3

(A₁- A₇) schematics of simulated fractures with different sizes, (B₁-B₇) Simulated amplitude signature images of fractures with different sizes, (C₁₋C₇) Simulated phase signature images of fractures with different sizes.

Fig. 4

Data augmentation based on superposition principle: (A, B) Simulated displacement field distribution when transducers # 1 and 7 are actuated by 300 kHz waves. (C) Calculated displacement field distribution when transducers # 1 and 7 are actuated by 300 kHz wave at the same time using superposition principles. (D) Schematics of machine learning algorithms architecture and performance of each stage. (E) Confusion matrix of bilayer neural network classifier. (F) Predicted fracture diameters distribution using bilayer neural network regressor. (G1, G2) Predicted fracture and loosening locations distribution using bilayer neural network regressor.

Fig. 5

(A) Effect of noise on accuracy of classification system. (B) variance of the dataset stored in each principal component. (C) Comparison of classifier accuracy with different noise to signal ratio with and without PCA. (D) Effect of frequency on accuracy of classifier.

Fig. 6

Numerical model: (A) The domain is modeled as a circle covered by skin and fat layers. A titanium implant is placed at the center covered by two rings of bone marrow and compact bone. (B) Selected mesh with 84,314 domain elements and 3003 boundary elements. (C) An example of a displacement field under 300 kHz excitation by transducer #1 showing attenuation, reflections, and refraction of the generated wave. (D) Geometry parameters and location of modeled crack/fracture. (E) Geometry parameters and location of modeled loosening.