Page 26 - AIH-2-2

Artificial Intelligence in Health Machine learning in arthroplasty

with optimal sensitivity of 95% and specificity of 90% was 4.1.2. Implant positioning
63
reported. When optimized, their DCNNs could identify ML does not only efficiently identify shoulder prosthetics
the implant’s model with an AUROC ranging from 0.86 to but it can also aid in surgical planning and positioning. In
1.0, sensitivities between 86% and 100%, and specificities TSA, following resection of the humeral head along the
of 100%. These studies demonstrate that ML in TSA can anatomical neck, the articular marginal plane becomes
63
contribute to a surgeon’s pre-surgical implant decision- a crucial area of measurement. This plane refers to the
making and post-surgical identification when a patient’s orientation and position of the humeral head implant
medical records are incomplete. Table 3 summarizes relative to the humeral shaft. 67,68 This value is manually
key studies that have explored the use of AI in implant calculated but attempts to automate this process have been
identification related to TSA. made with a regression forest-based program to improve

Table 3. A summary of key studies on the use of artificial intelligence in total shoulder arthroplasty.

Author Study population/ Performance (AUROC, accuracy, Input Function Model type
Training model sensitivity/specificity, etc.)
size
Imaging
Yi et al. 63 482 AUROC presence of implant: 1.0 Radiographic Detect implant presence, DCNN
AUROC differentiate rTSA/aTSA: 0.97 imaging differentiate between
AUROC identify model: 0.86 – 1.0 rTSA/aTSA, identify 5
different models
Urban et al. 64 597 VGG-16 accuracy: 75.2% Radiographic Classify implant by VGG-16, VGG-19,
VGG-19 accuracy: 76.2% imaging manufacturer ResNet-50, ResNet-152,
ResNet-50 accuracy: 75.2% DenseNet, NASNet
ResNet-152 accuracy: 74.5%
DenseNet accuracy: 78.9%
NASNet accuracy: 78.8%
VGG-16 AUROC: 0.93
VGG-19 AUROC: 0.93
ResNet-50 AUROC: 0.92
ResNet-152 AUROC: 0.91
DenseNet AUROC: 0.93
NASNet AUROC: 0.93
VGG-16 precision: 0.74
VGG-19 precision: 0.75
ResNet-50 precision: 0.77
ResNet-152 precision: 0.71
DenseNet precision: 0.79
NASNet precision: 0.78
Sultan et al. 65 597 Accuracy: 85.92% Radiographic Identify implant of 16 Dense residual ensemble
F1 score: 84.69% imaging prostheses from four network
Precision: 85.33% different manufacturers
Recall: 84.11%
Kunze et al. 66 3,060 AUROC: 0.994 – 1 Radiographic Identify 22 implants from ResNet-34
Accuracy: 97.1% imaging eight manufacturers
Sensitivity: 0.8-1
Tschannen 72 Mean localization error: 2.40 mm Radiographic Positioning implant Random regression
et al. 69 Mean angular error: 6.51° imaging forest
PROMS
McLendon 472 Probability prediction with model ASES scores, Predict improvement Machine learning model
et al. 72 including all variable data: 0.94-0.95 radiographic with ASES scores
Probability prediction with model excluding imaging
morphological variable data: 0.73-0.93
Probability prediction with model
excluding ASES variable data: 0.71-0.77
(Cont'd...)

Volume 2 Issue 2 (2025) 20 doi: 10.36922/aih.3278