Page 21 - AIH-2-2

Artificial Intelligence in Health Machine learning in arthroplasty

2.2. Outcome prediction studies have examined patients’ clinical and demographic

2.2.1. Patient-reported outcome measures (PROMs) variables and their ability to predict complications with
ML, all with varying results. 32-34 One of the earlier models
Multiple studies have examined the use of models to reported an AUROC of 0.73 for both mortality and cardiac
predict PROMs, a topic of increasing importance in complications and was then externally validated using a
recent years, as orthopedics pivot to more patient- different database. However, other models, such as Shah
34
centered metrics of success. Many of these studies, while et al.’s model, aimed to predict a more broadly defined
33
providing evidence in favor of the future applications criteria for major complications and reported an AUROC
of these models, lack external validation. 25-27 Therefore, of 0.732.
Huber et al.’s externally validated study represents the
28
most comprehensive investigation of the subject. By using ML has been used for various specific complications
35
demographic information and outcome scores, their model ranging from dislocation to loosening. Rouzrokh et al.
analyzed 65,000 total hip THA, producing an area under developed a DL model to assess the risk of hip dislocation
the receiver operating characteristic curve (AUROC) following THA from radiographic images. The final model
value (a single metric that compares true positive rate had a sensitivity of 89% and an AUROC of 76.7. In the
and false positive rate) of 0.87 for the THA Visual context of opioid usage, which poses a significant concern
Analog Scale (VAS). Similarly, Klemt et al. demonstrated in post-operative care, a predictive model was developed
7
significant outcomes in predicting various post-operative to assess prolonged post-operative opioid use in patients
PROMS following THA. Their ML models exhibited undergoing THA. This model demonstrated an AUROC
strong discrimination (AUROC > 0.83 for all PROMs), of 0.77, indicating a robust ability to predict opioid
36
and calibration analysis showed a high concordance usage outcomes. Most recently, an application detecting
between predicted and actual PROM scores, suggesting prosthetic loosening with high degrees of accuracy
the models’ reliability. Schwartz et al. investigated ML’s was able to identify loosening with 97% accuracy, an
29
37,38
ability to predict scores on post-operative pain surveys improvement from an older model with 88.3% accuracy.
achieving an AUROC of 0.79, outperforming the linear Table 1 summarizes the findings of key studies included
regression model. This finding suggests that ML models in this review, highlighting the role of AI in predicting
may have an advantage over traditional linear regression complications in patients undergoing THA.
methods in analyzing complex relationships between 2.3. Future direction
patient variables and PROMs. Table 1 presents a summary
of studies that assessed the use of AI to predict PROMs in AI, and more specifically ML, is emerging as a
patients undergoing THA. transformative tool in THA, demonstrating its early
potential in radiographic analysis, including precise implant
2.2.2. Readmissions, length of stay (LOS), and cost identification and accurate measurement acquisition.
ML has also shown rapid emerging value in predicting In addition, ML’s capability to analyze vast amounts of
crucial metrics such as readmissions, LOS, and cost in patient data and characteristics to discover correlations
THA. Studies by Ramkumar et al. have emphasized this and predict patient outcomes shows promise compared to
30
potential, with models exhibiting AUROC scores of 0.82 traditional methods such as linear regressions.
for LOS, 0.83 for inpatient charges, and 0.79 for discharge As we stand on the cusp of AI’s application and clinical
disposition. These results suggest a promising avenue for relevance in THA, it is becoming increasingly evident that
implementing cost-based reimbursement models and this field is primed for exploration and expansion by both
improving management of patient expectations. Further the academic and commercial orthopedic communities.
research by the same team, created and validated a model However, two significant challenges arise with its application
capable of predicting readmissions and LOS, achieving in medicine: limited explainability and data privacy. The
AUROCs of 0.87 and 0.71, respectively. The external ability to transparently understand the processes used
31
validation of these studies legitimizes the generalizability by AI to reach conclusions is crucial for its acceptance
and indicates the significance of the models in predicting in evidence-based medicine, especially when patient
outcomes for future use. Table 1 provides an overview of outcomes are at stake. Advancing our comprehension,
relevant studies on the use of AI in predicting readmission, documentation, and education of how these tools function
LOS, and cost to patients related to THA. is vital for their future applications. Transparency, in data
storage, security, and quality is also a growing concern for
2.2.3. Complications medical professionals, hospital executives, and patients.
Despite the continued success of THA, unforeseen One of the biggest hurdles in using large data sets is the
complications have been a target point for ML. Multiple ability to store patient health information securely. With

Volume 2 Issue 2 (2025) 15 doi: 10.36922/aih.3278