Takino K; Gifu Heart Center, Gifu. Japan
Kuze Y; Nagai T; Hori M; Nakazawa Y; et al

Archives of Physical Medicine & Rehabilitation.
107(11):3257-3264, 2026 Nov.

OBJECTIVE: To develop and validate a machine learning (ML) model for
predicting reduced exercise capacity at 3 months postoperatively using
clinical parameters available at hospital discharge in patients undergoing
cardiovascular surgery, as the prediction of patient-centered functional
outcomes remains unexplored in patients with cardiac disease.

DESIGN: Retrospective cohort study conducted in accordance with the
Strengthening the Reporting of Observational Studies in Epidemiology
statement.

SETTING: A single cardiac center.

PARTICIPANTS: Patients (N=590) aged <=75 years who underwent
cardiovascular surgery between April 2015 and January 2025 and completed
cardiopulmonary exercise testing at 3 months postoperatively.

INTERVENTIONS: Not applicable.

MAIN OUTCOME MEASURES: Reduced exercise tolerance was defined as a peak
oxygen uptake <14 mL/min/kg at 3 months postoperatively. The model
performance was assessed using the area under the receiver operating
characteristic curve, sensitivity, and specificity.

RESULTS: Of the 590 patients analyzed, 139 (23.6 %) demonstrated reduced
exercise tolerance (peak oxygen uptake <14 mL/min/kg) at 3 months
postoperatively. Following multicollinearity assessment, forward feature
selection identified 9 predictors from the 43 candidate variables. Among
the 5 ML algorithms evaluated using 5-fold cross-validation, the random
forest model achieved the highest area under the receiver operating
characteristic curve (0.855), with a sensitivity of 0.643 and specificity
of 0.867. The SHapley Additive exPlanations analysis identified the
percentage of isometric knee extension strength, ratio of early mitral
inflow velocity to early diastolic mitral annular velocity, and
high-density lipoprotein cholesterol level as the primary predictive
factors.

CONCLUSIONS: An ML model using clinical parameters at hospital discharge
can effectively predict postoperative exercise intolerance during the
recovery phase. This model enables early risk stratification and might
facilitate targeted rehabilitation interventions in patients undergoing
cardiovascular surgery.