Software-based concordance analysis of a predictive surgical planning model for target gap achievement in robotic-assisted total knee arthroplasty

Background In robotic-assisted total knee arthroplasty (TKA), achieving soft-tissue balance requires repeated intraoperative plan adjustment and remains experience-dependent. Predictive models may support planning, but their clinical value depends on whether model-generated plans approximate predefined target gaps within the robotic workflow. This study assessed the software-based concordance between model-generated plans and predefined target gaps in robotic-assisted TKA. The predefined target gaps were planning set points established within the robotic software rather than measured intraoperative gaps. Methods This retrospective software-based concordance study included 209 complete-case knees undergoing robotic-assisted TKA. A machine-learning-based predictive model generated a bone-resection plan, which was imported into the robotic planning software to calculate software-derived medial and lateral gaps in extension and flexion. These software-derived gaps were compared with predefined target gaps, which were planning set points rather than measured values and were 18 mm in 819 of 836 gap-level observations (98.0%). The primary analysis focused on target gap deviation, defined as the software-derived gap minus the predefined target gap. As a benchmark, the deviation of the preadjustment gaps from the same targets was used as a preadjustment reference. Results A total of 836 gap-level observations were analyzed. The overall mean target gap deviation was −0.142 mm (SD, 1.411 mm), with a mean absolute error of 1.016 mm. Overall, 77.6% of software-derived gaps were within ±1 mm of target, and 93.4% were within ±2 mm. Bland–Altman analysis showed a small overall bias of −0.142 mm, with 95% limits of agreement from −2.91 to +2.62 mm. At the knee level, only 43.1% of knees had all software-derived gaps within ±1 mm of target. Gap-level deviations ranged from −7 mm to +5 mm, and 16.3% of knees (34/209) had at least one gap with an absolute deviation > 2 mm. Compared with the preadjustment reference (MAE 2.002 mm, 49.4% within ±1 mm), the model substantially reduced deviation (MAE 1.016 mm, 77.6% within ±1 mm, p < 0.001). Conclusions In this software-based concordance analysis, the predictive planning model generated surgical plans whose robotic software-derived gaps approximated predefined target gaps with a mean absolute deviation of approximately 1 mm. The model-generated plan reduced target-gap deviation relative to the initial preadjustment plan within this target-conditioned workflow. However, because the targets were near-uniform (98.0% at 18 mm) and all measurements were software-derived, compartment-specific bias, moderate whole-knee consistency (43.1%), and occasional outlier deviations indicate that the model should be used as a surgeon-supervised decision support tool rather than as an autonomous planner. Prospective validation against measured intraoperative and postoperative gaps is required before clinical effectiveness can be claimed.