Page 28 - AIH-2-2

Artificial Intelligence in Health Machine learning in arthroplasty

efficiency in surgical planning. The attempted model to several established clinical measurement tools: ASES
69
displayed a position error of 2.4 mm and an angular error scores, Constant, University of California Los Angeles,
of 6.51° compared to manually computed measurements, Simple Shoulder Test, and Shoulder Pain and Disability
indicating the feasibility of applying such ML models to Index. Ultimately, the six-question score provided an
TSA. Table 3 displays the summaries of relevant studies equivalent or better validity, responsiveness, and clinical
69
exploring the use of AI in implant positioning related to interpretability value compared to previously developed
TSA. tools, paving the way for ML applications in creating
PROM assessments. Table 3 presents the summary of key
4.2. Outcome prediction studies assessing the role of AI in predicting PROMs in
4.2.1. PROMs patients undergoing TSA.
Shoulder procedures are no different from other areas of 4.2.2. Readmissions, LOS, cost
orthopedics, given that patient perspectives on outcomes Hospital stay length and patient-accrued costs can directly
and quality of care have become increasingly important
in orthopedics. Therefore, it is expected that AI has affect patient satisfaction and the perceived value of
moved beyond lower extremity procedures, in an attempt shoulder operations. The ability to anticipate deviations
to anticipate the factors that go into more challenging from normal care can allow patients to anticipate
prolonged hospital stays and increased costs. Biron
shoulder operations. Specific to shoulder procedures, et al. employed a TSA-specific ML model to predict
74
Polce et al. developed an ML algorithm to predict patient patients who would experience a hospital stay for less
70
satisfaction 2 years post-operation following TSA, further than a day. Their model produced an AOC of 0.77 with
evaluating predictive factors that could affect patient moderate reliability. In addition, Karnuta et al. used
75
74
satisfaction. Their most successful algorithm displayed artificial neural networks (ANN) to predict LOS, discharge
70
an AUROC value of 0.80 and a Brier score of 0.14. A Brier disposition, and inpatient costs in patients undergoing
score measures the accuracy of a prediction, where zero shoulder arthroplasty secondary to chronic degenerative
is the best possible score. The five most important inputs changes and traumatic injury. In patients with a chronic
to the algorithm were baseline Single Assessment Numeric degenerative disease, respective accuracies for predicting
Evaluation score, exercise and activity, insurance status, LOS, discharge disposition, and cost were 91.8%, 73.1%,
diagnosis, and pre-operative duration of symptoms. A and 76.5. In patients with traumatic presentations, the
70
75
more recent use of ML is to use natural language processing accuracy was slightly diminished for predicting LOS,
models to automatize and streamline PROMs. Menendez discharge disposition, and cost, which were 79.1%, 72%,
et al. aimed to enhance the efficiency of categorizing and 70.3%, respectively. 75
71
negative patient comments following TSA. Their model
76
demonstrated a 98.5% accuracy, reliably classifying Readmissions were examined by Arvind et al.
patient comments. These ML applications mark a wave through ML employment to predict 30-day readmission
of autonomization in research, making it possible for following primary TSA. The maximum derived C-statistic
orthopedic surgeons to conduct research despite a possible (equivalent to AUROC) was 0.74, and the peak f1-score
lack of personnel. (combines precision and sensitivity to quantify accuracy)
was 0.18 . Gowd et al. achieved a prediction for
77
76
Moreover, ML algorithms have been employed to unplanned readmission following TSA with an AUROC
predict American Shoulder and Elbow Surgeon (ASES) of 0.66, indicating slightly less reliability. However,
scores based on pre-operative data, including demographic they also added a model that predicted inpatient costs,
information, computed tomography imaging, pre- demonstrating a higher AUROC of 0.87. Table 3 outlines
77
operative VAS pain score, and pre-operative ASES scores. the relevant studies investigating the use of AI in predicting
The model predicted patient-reported improvement with readmission, LOS, and cost to patients related to TSA.
sensitivities between 0.84 and 0.95. 72
Beyond predictive measures, ML has been used to 4.2.3. Complications
develop a novel shoulder outcomes measure called the Complications related to TSA can have multitudes of effects
Shoulder Arthroplasty Smart (SAS) score. Using data from on patient mortality, functional outcomes, and satisfaction.
3,667 patients receiving anatomic or reverse shoulder ML algorithms have displayed capabilities in potentially
arthroplasty with a minimum 2-year follow-up, the ML planning for these unanticipated complications. Gowd
algorithm chose six outcome measures, assessed by three et al. developed ML algorithms to anticipate any operative
78
objective questions regarding motion input and three adverse events, anemia necessitating blood transfusion,
subjective patient measures. The SAS was then compared and extended hospital stay. Accuracies for predicting all
73

Volume 2 Issue 2 (2025) 22 doi: 10.36922/aih.3278