Machine learning–based models may accurately predict improvement in American Shoulder and Elbow Surgeons (ASES) scores after total shoulder arthroplasty (TSA) for glenohumeral osteoarthritis (GHOA), according to a poster study on display today and tomorrow in Academy Hall.
“We have observed that patient satisfaction is not always directly correlated with measurable outcomes achieved after TSA,” presenting author Paul McLendon, MD, shoulder arthroplasty specialist, told AAOS Now Daily Edition. “We suspect that incongruent patient expectations are largely responsible for this discrepancy. We undertook this study to determine whether it would be possible to develop machine learning algorithms to predict level of improvement after TSA.” With this information, he said, orthopaedic surgeons could set appropriate expectations for patients.
The researchers retrospectively assessed data from 472 shoulders diagnosed with primary GHOA and treated with either anatomical TSA (n = 431) or reverse TSA (n = 41). The patient cohort was 55 percent male, and the average age was 68 years. Patients were classified according to preoperative glenoid and rotator cuff morphology assessed via CT scan. Improvement was measured based on changes in pre- and postoperative ASES scores. Patients were stratified into three groups according to ASES score changes (class A, < 28 points; class B, 28–55 points; class C, >55 points).
Machine learning was utilized to assess patterns of morphology and ASES scores related to outcomes (Fig. 1). Three predictive modeling approaches were compared:
- Model 1 used all baseline variables.
- Model 2 omitted morphological variables.
- Model 3 omitted ASES variables.
The investigators found that Model 1, which included all available preoperative data and morphology, gave the most accurate prediction of outcomes. Probability values for Model 1 were 0.94, 0.95, and 0.94 for classes A, B, and C, respectively. For Models 2 and 3, probability values were 0.93, 0.80, and 0.73, and 0.77, 0.72, 0.71, for classes A, B, and C, respectively.
Notably, no significant relationship was found between preoperative structural pathology and outcomes. “We were surprised to observe that the presence of more advanced structural pathology preoperatively did not necessarily correlate with inferior postoperative outcomes,” Dr. McLendon said. He also noted that the study group contained an unexpectedly high percentage of patients with GHOA who demonstrated fatty infiltration and/or atrophy of the rotator cuff.
“The major takeaway is that machine learning can accurately predict the level of improvement—with up to 95 percent accuracy—after shoulder arthroplasty for primary glenohumeral arthritis,” Dr. McLendon concluded.
The study is limited by its retrospective nature, and Dr. McLendon added that the applicability of the findings may be limited by the fact that “machine learning algorithms are not freely available to the general public at the present time.”
Despite the restrictions, Dr. McLendon added: “This study opens the possibility for further prospective studies to assess the effects accurate preoperative expectations may have on overall patient satisfaction after shoulder arthroplasty.”
The study will be on display as P0810 today and tomorrow in Academy Hall, Sails Pavilion, from 7 a.m. to 5 p.m.
Dr. McLendon’s coauthors of “Machine Learning Can Predict Level of Improvement in Shoulder Arthroplasty” are Kaitlyn N. Christmas BS, CCRC; Peter Simon, PhD; Otho R. Plummer, PhD; Audrey Lee Hunt, BS; Adil Ahmed, MD; Mark A Mighell, MD; and Mark A. Frankle, MD.
Rebecca Araujo is the associate editor of AAOS Now.