Machine-learning prediction of 90-day readmission after primary total hip Arthroplasty: Analysis of 1,340 cases from the Michigan Arthroplasty Registry (MARCQI).

BACKGROUND

Ninety-day readmission after total hip arthroplasty (THA) drives cost and signals sub-optimal recovery, yet existing risk-stratification tools are imprecise. We aimed to develop and validate a machine-learning model to predict 90-day readmissions and to identify modifiable risk factors.

METHODS

The Michigan Arthroplasty Registry Collaborative Quality Initiative (MARCQI) was queried for all primary THAs performed between 2012 and 2023 at a single institution. All surgeries were performed by fellowship-trained adult reconstruction surgeons. Demographics, comorbidities, peri-operative variables, and discharge dispositions were extracted. Univariate analyses compared patients readmitted within 90 days with those not readmitted. A multilayer perceptron neural network (MPNN) was trained on 70% of the cohort and tested on the remaining 30%. Model discrimination was assessed with area under the receiver-operating-characteristic curve (AUC), and variable importance was calculated.

RESULTS

Of 1,340 THA patients, 69 (5.1%) were readmitted within 90 days, with rates climbing from 0% in ASA I to 24% in ASA IV (p < .001). Spearman correlations pinpointed length of stay (LOS) as the strongest readmission predictor (midnights ρ = 0.130; hours ρ = 0.123; both p < .001), followed by discharge to post-acute care (ρ = -0.074; p = .007), smoking (ρ = 0.084; p = .002), and alcohol use (ρ = -0.072; p = .008). No other demographic or comorbidity variables reached significance.An MPNN model achieved 94.7 % training accuracy, 95.2% testing accuracy, and an AUC of 0.71, ranking length of stay, ASA score, and bleeding disorders as its top three predictors.

CONCLUSION

Prolonged hospital stays and higher ASA status are key drivers of 90-day readmission after THA. Integrating machine-learning risk stratification with strategies to shorten LOS, enhance preoperative optimization, and refine discharge planning may reduce readmission rates.

LEVEL OF EVIDENCE: Prognostic Level III.