Page 22 - AIH-2-2

Artificial Intelligence in Health Machine learning in arthroplasty

the rapid evolution of AI technology, the processes of data 100% in identifying implants from plain radiographs. 12,40,41
storage may lag, creating vulnerabilities that jeopardize the This may better aid surgeons in identifying knee implants
security of society’s most private information. during pre-operative planning should there need to be a
The integration of AI in medicine is further complicated revision. 13,41 Beyond solely identifying existing implants,
42
by the issues of overfitting, bias, and large resource allocation, Kunze et al. found that ML could preoperatively predict
which are not specific to any particular application of this implant size within ±1 sizes with an accuracy of up to 90%,
technology. For smaller studies, inputting patient data and within ±2 sizes with an accuracy of up to 97.7%. In
addition, other studies have shown promising applications
can cause the models to memorize and regurgitate the of ML that extend beyond implant selection for TKA. 42-44
inputted data, instead of creating new predictions based on Table 2 summarizes key studies that have explored the use
new data. This issue, commonly referred to as overfitting, of AI in implant identification related to TKA.
may complicate predictive models like the ones discussed
above in predicting LOS and readmission following THA. 3.1.2. Implant positioning
Although this issue is more common in smaller data sets, it
is a reminder that cross-validation techniques are needed The use of ML may also aid surgeons in identifying implant
to improve output quality and generalizability. 39 positioning and osseointegration. In most models, ML was
able to determine optimal positioning in the posterior tibial
Contrastingly, the large data sets required of some slope, femoral flexion, and tibio-femoral alignment as well as
orthopedic predictive models may experience challenges targeting the implant in the sagittal plane. 45,46 In addition to
with models trained on biased data. As medical research identifying the optimal positioning of knee implants, ML also
continues to focus on eliminating health disparities, enables the identification of malposition or loose implants.
biased predictive outputs may decrease public trust in the Similar tools comparable to the highly accurate DCNN
technology. Lastly, a significant challenge in processing models created by Loppini et al. can be effectively applied
37
large data sets is the requirement for robust computational to TKA for detecting implant loosening using plain films. 38,47
resources, which limits the ability of certain institutions and Table 2 presents a summary of relevant studies investigating
academic centers to utilize this technology. As a result, the the use of AI in implant positioning related to TKA.
application of AI in orthopedics may be applied unevenly
across different settings. Therefore, it is crucial that the 3.1.3. Disease identification
future directions of AI research in orthopedics account for One of the most frequent indications for TKA is OA of
these restraints and work to mitigate the challenges posed the knee. Therefore, ML has also been investigated for its
by transparency, security, overtraining, bias, and resource ability to identify OA. Mehta et al. found that ML can
48
allocation. distinguish between OA and rheumatoid arthritis (RA) in
Despite these challenges, AI opens up numerous 82% of cases, with AUROC reaching as high as 0.92±0.06
promising avenues within THA, that are still in the based on histology alone. Other models have found that
early stages of exploration. AI offers great potential to ML was able to predict OA based on radiographs, with
alleviate the workload and time burden on surgeons AUROC between 0.79 and 0.87. 49-51 These diagnostic aids
in planning, templating, and interpreting radiography. are, therefore, well-positioned to identify patients who
Furthermore, as a diagnostic assistant, AI has the potential may benefit from TKA. Table 2 provides an overview
to enhance efficiency and diagnostic accuracy significantly. of the relevant studies examining AI’s role in disease
Nevertheless, at its current stage, AI serves best as a identification for TKA.
supplementary tool to surgeons. Prioritizing efforts to 3.2. Outcome prediction
enhance transparency and perform real-world validations
will be critical to building trust and actualizing the full 3.2.1. PROMs
potential of this exciting direction in orthopedics. Various PROMs are used to evaluate knee function and
improvement. Similarly, numerous ML models designed
3. Knee arthroplasty have shown promise in predicting outcomes post-TKA.
3.1. Radiographic imaging DCNNs and DL systems have indicated potential in
predicting patient satisfaction following TKA with an
3.1.1. Implant identification
AUROC of 0.81 and a C-statistic between 0.74 and 0.77,
Several developed ML models have shown efficacy in indicating reasonable to strong prediction. 52,53 Huber et al.
28
identifying total knee arthroplasty (TKA) implants based on tracked VAS pain scores and developed an extreme gradient
plain films. Similarly to THA predictions, studies reported boosting model with an AUROC of 0.86 for post-operative
the accuracy of varying ML models to be between 92% and pain. Other common PROMs include the knee injury and

Volume 2 Issue 2 (2025) 16 doi: 10.36922/aih.3278