Artificial Intelligence and Orthopaedic Prosthetic Planning: A State-of-the-Art Review and Evolving Liability Perspectives

Abstract

Background and aim: Artificial intelligence (AI) is gaining increasing relevance in orthopaedic surgery, particularly in prosthetic surgery, due to its ability to support preoperative planning through advanced imaging analysis, implant size prediction, and outcome forecasting. However, recent literature shows considerable variability in employed models, evaluated outcomes, and clinical applicability. The objective of this scoping review is to map AI applications in preoperative planning for orthopaedic arthroplasties and to assess their impact on radiographic and clinical outcomes, also discussing key ethical and medicolegal implications within both Italian and international contexts. Materials and methods: A literature review was conducted following scoping review methodology. The bibliographic search (10 September 2025) was performed in PubMed and Scopus using the query “preoperative planning WITH artificial intelligence AND prosthesis orthopaedic surgery AND outcomes”, restricted to the years 2020–2025, English-language studies, and research focused specifically on real-world AI techniques applied to preoperative planning in prosthetic surgery, reporting radiographic and/or clinical outcomes related to planning. Exclusion criteria included intra/postoperative studies, non-orthopaedic applications, robotic surgery, studies lacking clinical outcomes, case reports, and articles without full-text availability. After PRISMA screening and selection, 42 primary studies were included. Results: Of the 42 studies included, 20 focused on the hip, 19 on the knee, and 3 on the shoulder. Available evidence indicates that AI may improve templating accuracy and prosthetic component positioning, with more robust results in hip and knee arthroplasty, while applications in shoulder arthroplasty remain emerging. Nonetheless, important methodological limitations persist, including algorithm heterogeneity. Discussion: Overall, the findings suggest a promising role for AI in preoperative planning; however, the heterogeneity and variable quality of the evidence call for caution in interpretation and highlight the need for more rigorous prospective research. These considerations also carry relevant medicolegal implications, as the reliability and standardisation of AI-based tools represent essential prerequisites for their safe and conscious integration within diverse regulatory frameworks. Conclusions: AI appears to be a promising tool in the preoperative planning of orthopaedic arthroplasties, although further clinical validation and methodological standardisation are required. The evidence gathered also provides a useful foundation for addressing the associated medicolegal and regulatory implications, particularly in light of evolving Italian and European regulations and their differences from U.S. models.

1. Introduction

In recent years, artificial intelligence (AI) has gained an increasingly prominent role across numerous medical disciplines, owing to its ability to analyse large datasets and support complex decision-making processes [1]. The adoption of machine learning and deep learning algorithms is transforming clinical practice from diagnostics to workflow optimisation and personalised therapeutic strategies. Orthopaedic surgery, in particular, represents an area in which the integration of AI-based solutions is generating significant interest due to the potential for improved precision and efficiency [2,3,4,5]. Within this domain, orthopaedic prosthetic surgery stands out as a field of special relevance, with a growing number of studies exploring the application of AI in preoperative planning, imaging analysis, prediction of clinical outcomes, and assessment of potential post-surgical complications. The advancement of emerging technologies, including generative and multimodal models, is expected to redefine evaluation and treatment processes, offering new perspectives for both standardisation and personalisation of surgical interventions [6,7].

Despite growing enthusiasm and the rapid expansion of scientific literature, the maturity level of current AI-based solutions and their actual degree of integration into clinical practice remain topics of debate and active investigation. Orthopaedic prosthetic surgery, therefore, constitutes a privileged ground for the development and validation of innovative tools, with the overarching goal of improving quality of care and patient outcomes [8,9,10]. However, current literature reveals substantial heterogeneity regarding algorithmic structures, study populations, outcome measures, and the technological readiness of proposed systems. In light of this variability, the present review aims to provide an updated and critical overview of the state of the art, discussing key emerging evidence while highlighting existing gaps. This approach makes it possible to synthesise contributions from different sources and to provide a broader perspective on the current evolution of AI-assisted preoperative planning in orthopaedic prosthetic surgery.

Beyond clinical and technical parameters, the adoption of AI-based tools raises significant concerns regarding safety, decision-making transparency, data governance, and algorithmic reliability. Although such themes are not exclusive to AI-driven medicine, they acquire particular weight in the context of preoperative planning, wherein the accuracy of assessments has a direct impact on operative precision and functional patient outcomes. Building on evidence emerging from the literature published from 2020 to 2025, this work aims to discuss and deepen the main ethical and medicolegal implications inherently associated with the use of AI in orthopaedic prosthetic surgery. These considerations will be examined not only with reference to the Italian context but also within the broader international framework in order to outline shared trends, divergences, and regulatory challenges across different legal systems.

Indeed, despite the enormous potential of AI in the healthcare sector, its use introduces limitations that must be carefully evaluated. AI-based systems, beyond reshaping the labour landscape, may directly interfere with fundamental human rights, including the right to health (Art. 32 Italian Constitution), the right to a fair trial (Art. 111 Italian Constitution), and the principle of equality (Art. 3 Italian Constitution). These concerns give rise to crucial ethical–legal questions. Such issues, linked to the deployment of intelligent systems, demand an in-depth reflection on the responsible use of these technologies—particularly when they influence decisions capable of affecting health, rights, or individual freedoms. An approach that acknowledges both the opportunities and criticalities of AI aims to guide its future integration into orthopaedic clinical pathways in a manner that is at once innovative, safe, and responsible, consistent with clinical needs and with major international regulatory frameworks.

2. Materials and Methods

2.1. Search Strategy

This review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR) [11]. A literature search was performed with the aim of analysing the use of artificial intelligence (AI) in the preoperative planning of orthopaedic prosthetic surgery and evaluating its impact on radiographic and clinical outcomes. The bibliographic search was carried out on 10 September 2025 using the databases PubMed and Scopus, applying the following search string: “preoperative planning WITH artificial intelligence AND prosthesis orthopaedic surgery AND outcomes”.

For both databases, the search was restricted to the period 2020–2025, with the aim of mapping the most recent and clinically relevant evidence for current clinical practice. In Scopus, in addition to year selection, filters were applied for the Medicine subject area, publication stage (Final), language (English), and open access availability. The combined search yielded a total of 214 articles (15 from PubMed and 199 from Scopus, including duplicates).

2.2. Inclusion Criteria

In line with the objectives of this scoping review—to map the concrete use of artificial intelligence (AI) in the preoperative phase of orthopaedic prosthetic surgery and evaluate how such applications may influence planning quality and related outcomes—the following were included: systematic reviews used as supporting evidence; English-language articles published between January 2020 and 10 September 2025; studies exclusively concerning orthopaedic prosthetic surgery (hip, knee, shoulder, etc.), including procedures for specific conditions such as hip dysplasia; studies describing the actual use of AI-based techniques (machine learning, deep learning, neural networks) in the preoperative phase, with applications such as templating, implant prediction, alignment planning, or preoperative radiological assessment; studies reporting at least one clinical or radiographic outcome relevant to preoperative planning, including templating accuracy, implant size prediction, positioning precision, reduction in complications, or improvement in functional outcomes.

These criteria were defined to identify studies genuinely focused on the preoperative stage, distinguishing the contribution of AI to surgical planning from technologies applied during or after the procedure.

2.3. Exclusion Criteria

After removal of duplicates, the following were excluded: studies focusing exclusively on intraoperative or postoperative phases; articles concerning AI applied to non-orthopaedic medical fields or veterinary medicine; studies centred on surgical robotics, considered a distinct technological domain primarily related to intraoperative assistance; works lacking clinical or radiographic outcomes; case reports, editorials without supporting data, and articles published in languages other than English or not available in full text [Figure 1].

3. Results

Following the selection process (screening of titles, abstracts, and full texts), 41 studies were included and classified according to the treated anatomical district: 19 studies on hip arthroplasty, 19 on knee arthroplasty, and 3 on shoulder arthroplasty. The studies were analysed considering (i) the anatomical district involved, to clarify differences in AI applications across prosthetic types; (ii) the reported outcomes, with particular attention to templating accuracy, implant prediction, alignment, morphometric assessment, and functional parameters; and (iii) clinical aspects related to procedure-specific complications (e.g., leg length discrepancy, malpositioning, instability), for which AI appears to provide a meaningful contribution by improving planning precision, standardisation, and predictive capability.

The results are presented in narrative form, organised by anatomical district, with the intention of showing how AI use in the preoperative phase can directly influence key stages of surgical planning—reducing the margin of technical error and potentially decreasing the most frequent prosthetic complications. Clinical benefits are discussed mainly in potential terms, also in light of the predominantly short-term follow-up.

The evidence collected also forms the basis for discussing the medicolegal implications, analysed both in relation to the Italian regulatory framework and in the broader international context, with particular reference to clinical liability management and the need for standardisation of AI systems.

3.1. Hip Arthroplasty

The results concerning the use of artificial intelligence in preoperative planning for total hip arthroplasty are reported in

Table 1. Hip arthroplasty results.

First Author/Year/Title	Study Design	AI Type	Objective	Outcomes Measured	Macro-Theme	Key Findings
Wu et al., 2023—Accuracy analysis of artificial intelligence-assisted three-dimensional preoperative planning in total hip replacement [12]	Retrospective single-centre cohort; AIHIP-3D planning (n = 95) vs. 2D planning (n = 66); postoperative radiographic outcomes and planning–intraoperative agreement evaluated.	AIHIP (deep learning)	Assess the accuracy of AI-assisted 3D planning in predicting implant sizing, acetabular positioning, and leg length restoration.	Acetabular/stem sizing accuracy; inclination/anteversion angles; % implants within Lewinnek/Callanan safe zones; postoperative LLD.	3D Planning & Implant Positioning	AI significantly more accurate: cup 54% vs. 38% (p = 0.048), stem 64% vs. 44% (p = 0.011). Higher safe-zone placement (Lewinnek 86.3% vs. 72.7%). Better LLD correction (2.18 mm vs. 4.42 mm, p < 0.001).
Wu et al., 2023—Short-term outcome of AI-assisted preoperative three- dimensional planning of total hip arthroplasty for developmental dysplasia of the hip compared to traditional surgery [13]	Retrospective cohort; AIHIP-3D (n = 34) vs. 2D (n = 27); follow-up ≥ 1 year.	AIHIP (deep learning, 3D reconstruction + auto-segmentation)	Evaluate the accuracy of AI planning in DDH-THA.	Cup/stem sizing accuracy; inclination/anteversion; safe-zone %; LLD; HHS; complications.	3D Planning & Positioning	AI superior: cup 56% vs. 30%, stem 68% vs. 41%. Lower LLD (1.64 mm vs. 3.53 mm). No major complications.
Zhu et al., 2025—Efficacy of an AI preoperative planning system for assisting in revision surgery after artificial total hip arthroplasty [14]	Retrospective cohort; revision THA (25 pts/26 hips); mean follow-up 25 months.	AIHIP (Transformer-UNet)	Evaluate AIHIP effectiveness in revision-THA planning.	Prosthesis matching accuracy; HHS; ROM; complications.	Revision THA + 3D Planning	Matching: cup 96.1%, stem 100%. Significant HHS and ROM improvement; planning time ~5 min; only 2 complications (hematoma + dislocation).
Li et al., 2025—Advantages and effectiveness of AI three-dimensional reconstruction technology in the preoperative planning of total hip arthroplasty [15]	Retrospective cohort; osteonecrosis; AI-3D (n = 55) vs. 2D (n = 54).	G-NET deep learning reconstruction + templating.	Compare AI-3D vs. 2D in primary THA.	Sizing accuracy; blood loss; op-time; LOS; LLD; radiographic metrics; HHS; complications	AI-3D Planning & Perioperative Outcomes	AI superior in sizing (cup 90.9% vs. 72.2%, stem 87.3% vs. 66.7%), ↓blood loss, ↓LOS, ↓LLD. Better HHS at 1–6 mo. No major complications.
Lu et al., 2025—AI-assisted 3D versus conventional 2D preoperative planning in total hip arthroplasty for Crowe type II–IV high hip dislocation: a two-year retrospective study [16]	Retrospective cohort; AI-3D (n = 49) vs. 2D (n = 43).	AI-3D deep-CNN segmentation.	Evaluate AI-3D in complex THA.	Sizing; safe-zone positioning; op-time; bleeding; LLD; HHS/WOMAC/VAS; complications.	AI-3D Complex THA	AI superior in sizing & positioning; ↓LLD, ↓bleeding, ↓op-time. No revision failures.
Zheng et al., 2025—Application of artificial intelligence-based three-dimensional digital reconstruction technology in precision treatment of complex total hip arthroplasty [17]	Prospective randomised cohort; AI (n = 29) vs. 2D (n = 27).	G-NET AI-HIP automation	Evaluate AI in complex THA planning.	Sizing; acetabular placement; LLD; offset; time; HHS; complications.	AI for Complex THA	AI superior for sizing & LLD; ↓blood loss + op-time. Better early HHS. No major complications.
Xie et al., 2024—Application and evaluation of artificial intelligence 3D preoperative planning software in developmental dysplasia of the hip [18]	Retrospective; 103 DDH Crowe I–IV.	AIHIP 3D automatic segmentation + simulation.	Compare AIHIP vs. 2D cup sizing.	Sizing accuracy ±0/±1; MAE; influencing factors.	AI-Planning in DDH	AIHIP ± 1 = 95.1% vs. 81.1% (p < 0.05). 2D accuracy collapses in Crowe II–III; AI remains stable.
Anwar et al., 2024—AI technology improves the accuracy of preoperative planning in primary total hip arthroplasty [19]	Prospective cohort 117-pt.	AIHIP deep-learning auto-segmentation.	Compare AIHIP vs. 2D.	Cup/stem accuracy; planning time; predictors of failure.	AI-Primary THA	AIHIP much more accurate; planning time halved. 2D fails in DDH & high femoral anteversion.
Yang et al., 2024—Clinical application of artificial intelligence-assisted three-dimensional planning in direct anterior approach hip arthroplasty [20]	Retrospective comparative 440-pt.	AIHIP CMGNet + Unet + LSTM.	Compare AI vs. 2D in DAA-THA.	Sizing; op-time; blood loss; LLD; FO; HHS; ICC.	AI for 3D-DAA Planning	AI improves sizing, op-time, fluoroscopy, blood loss, LLD.
Rouzrokh et al., 2024—THA-Net: A Deep Learning Solution for Next-Generation Templating and Patient-specific Surgical Execution [21]	Retrospective; 14,357 patients, 356.305 radiographs.	Diffusion-Model THA-Net.	Generate optimal post-op-like X-rays via AI.	Realism, safety-zone metrics, dislocation risk.	Generative-AI in Preoperative Simulation	96.5% images in safe-zone; high realism; biomechanically optimised implant orientation.
Zheng et al., 2024—Is AI 3D-printed PSI an accurate option for patients with developmental dysplasia of the hip undergoing THA? [22]	Prospective RCT, 60 pts.	AI-PSI (3D printed patient-specific instruments)	Evaluate implant positioning with AI-PSI vs. manual.	Anteversion/abduction; FO; LLD; accuracy; bone cuts; HHS/VAS.	AI + PSI (Guided Implant Execution)	AI-PSI significantly reduces positioning error & LLD; accuracy 90–93% vs. 60–80%.
Ding et al., 2021—Value of preoperative three-dimensional planning software (AI-HIP) in primary total hip arthroplasty: a retrospective study [23]	Retrospective 316 pts.	G-NET segmentation.	Compare AI-HIP 3D vs. manual 2D.	Sizing, positioning, LLD, offset.	3D-AI THA Planning	Cup 94% vs. 65%, stem 88% vs. 59%. ICC > 0.95 for stem, cup, and osteotomy level
Huo et al., 2021—Value of 3D preoperative planning for primary total hip arthroplasty based on artificial intelligence technology [24]	Prospective 53-pt study.	Deep-learning 3D, reinforcement-matching.	Compare AI-3D vs. 3D vs. 2D.	Sizing; planning-time.	3D-AI Efficiency	AI accuracy similar to 3D, >2D; planning time drastically shorter.
Chen et al., 2022—Development and Validation of an Artificial Intelligence Preoperative Planning System for Total Hip Arthroplasty [25]	Prospective 120 pts.	Attention-UNet + PointRend.	Clinical validation of AIHIP.	Sizing; LLD; offset; blood loss; HOOS JR; EQ-5D.	AI-3D CT Planning	AIHIP ±1 sizing = 96.7% vs. 55–65%.
Chen et al., 2022—Validation of CT-Based Three-Dimensional Preoperative Planning in Comparison with Acetate Templating for Primary Total Hip Arthroplasty [26]	Prospective 57-pt comparison.	AI segmentation + landmark extraction.	Compare 3D-AI vs. 2D acetate templating.	Sizing; MAE; ICC.	3D-CT AI Planning	Cup/stem accuracy 93% vs. 79–75%. High ICC reliability.
Jang et al., 2022—John Charnley Award: Deep Learning Prediction of Hip Joint Center on Standard Pelvis Radiographs [27]	Retrospective 3965 pts.	UNet + ResNet-34.	Predict hip centre on plain radiographs.	Mean error; % within 3–5 mm.	Landmark Automation (HJC)	91% accuracy within 5 mm; fast (0.65 s/hip).
Wang et al., 2024—Accuracy analysis of the new artificial anatomical marker positioning method (shoulder-to-shoulder) in preventing leg length discrepancy in total hip arthroplasty [28]	Retrospective; 47 THA pts.	AIHIP-3D segmentation platform.	Compare manual vs. AI-LLD correction.	LLD; cut distance; prosthesis matching.	LLD Control & Femoral Positioning	No major LLD difference; AI better in cut precision; prosthesis match >91–95%.
Zhang et al., 2023—The role of 3-dimensional preoperative planning for primary total hip arthroplasty based on artificial intelligence technology to different surgeons A retrospective cohort study [29]	Retrospective matched cohort (n = 120): senior vs. junior ± AIHIP.	AIHIP (CT-3D + neural landmark recognition)	Compare AI usefulness between junior and senior surgeons.	LLD, NSA, offset; op-time; Hb-loss; radiation; complications; sizing accuracy.	3D Planning & Surgical Performance	Major improvements only in junior surgeons: ↓LLD, ↓op-time, ↓Hb-loss, ↓radiation, ↓complications. Senior surgeons showed no significant benefit.
Karnuta et al., 2023—Artificial Intelligence for Automated Implant Identification in Total Hip Arthroplasty: A Multicenter External Validation Study Exceeding Two Million Plain Radiographs [30]	Multicentre external validation.	Inception-V3 CNN; >2 million images.	Automate femoral stem recognition.	AUC, accuracy, sensitivity/specificity; processing speed.	AI-Prosthesis Identification (Revision THA)	External accuracy 97.9%; output in 0.02 s per image.

.

Model creation and segmentation

In the included studies, artificial intelligence was primarily applied to the creation of patient-specific anatomical models through CT-based three-dimensional reconstruction and automatic segmentation, with some approaches extending to radiograph-based landmark extraction or generative image synthesis [12,13,14,15,16,17,18,19,20,21,23,24,25,26,27,29,30]. These AI-based model generation techniques enabled a consistent representation of complex hip anatomies, including hip dysplasia, and demonstrated greater agreement between preoperative planning and postoperative radiographic assessment compared with traditional two-dimensional model creation [12,13,14,15,16,17,18,19,23,24,25,26]. Automated landmark identification, including localisation of the hip joint centre, further contributed to standardisation and reduction in inter-observer variability [27]. Among emerging approaches, THA-Net introduced a generative framework to simulate postoperative-like radiographic configurations from preoperative data [21].

Implant sizing

Implant sizing represented one of the most consistently reported applications of AI-assisted preoperative planning. Across multiple cohorts, AI-based systems demonstrated significantly higher accuracy in acetabular and femoral component sizing compared with conventional two-dimensional templating, with stable performance in primary total hip arthroplasty, complex anatomies, and revision settings [12,13,14,15,16,17,18,19,23,24,25,26]. Accuracy remained high even in scenarios where two-dimensional templating showed a marked decline in performance, such as Crowe II–III dysplasia or high femoral anteversion [18,19]. These findings primarily reflect improvements in technical planning precision rather than direct evidence of improved long-term clinical outcomes.

Component alignment and biomechanical reconstruction

Several studies reported improved component alignment and biomechanical reconstruction with AI-assisted planning. This included more accurate acetabular positioning within predefined safe zones, improved restoration of the hip rotation centre, reduced leg length discrepancy, and more consistent femoral offset reconstruction [12,13,15,16,17,18,20,22,23,25,28,29]. Reductions in leg length discrepancy were particularly evident in anatomically complex cases and in cohorts managed by less-experienced surgeons [13,16,17,29]. Although these parameters represent clinically relevant biomechanical targets, the reported improvements predominantly reflect geometric and radiographic accuracy rather than validated long-term clinical benefit.

Prediction and decision-support functions

Alongside geometric planning activities, AI-assisted approaches were also applied to predictive and decision-support functions. Several studies reported associations with reduced operative time, blood loss, fluoroscopy use, and improved workflow efficiency, particularly in complex cases and in direct anterior approach total hip arthroplasty [15,16,17,20,24,29]. Benefits were more pronounced among junior surgeons, suggesting a potential role of artificial intelligence in standardisation and mitigation of operator-dependent variability [29]. In addition, AI-based systems were developed for automated implant identification in revision settings and for the prediction of biomechanically favourable implant configurations, supporting preoperative decision-making rather than demonstrating definitive improvement in clinical outcomes [21,30]. Overall, clinical benefits were generally discussed as potential and exploratory, often in the context of short-term follow-up.

3.2. Knee Arthroplasty

The results concerning the use of artificial intelligence in preoperative planning for total knee arthroplasty are reported in

Table 2. Knee arthroplasty results.

Author/Year/Title	Study Design	Type of AI	Objective	Evaluated Outcomes	Macro-Theme	Key Findings
Min et al., 2025—Comparison of traditional template measurements and artificial intelligence preoperative planning in total knee arthroplasty [31]	Prospective randomised trial (48 patients, primary TKA)	AI-KNEE (Changmugu Medical Technology): automatic segmentation, 3D reconstruction, prosthetic planning	Assess the accuracy and clinical performance of 3D AI-assisted planning vs. traditional 2D templating	Accuracy of femoral/tibial/liner sizing; operative time; blood loss; drainage; HKA alignment; VAS (pain); AKS (function)	Planning accuracy; Alignment; Perioperative efficiency	AI significantly superior in sizing (femur 92% vs. 67%; tibia 87.5% vs. 62.5%), shorter operative time (~68 vs. 84 min), lower blood loss, improved postoperative alignment, lower VAS in first 2 weeks, higher AKS at 3 months. No postoperative complications reported.
Pan et al., 2025—Comparison of 3D Printing Technology and Artificial Intelligence Assisted in Total Knee Arthroplasty [32]	Single-centre double-blind RCT	AI-based 3D surgical planning (Changmugu software)	Compare patient-specific guides derived from 3D printing vs. AI-based planning in TKA	Surgical time, bleeding, drainage, length of stay, alignment accuracy (HKA, FFC, FTC), VAS, HSS	Planning accuracy; Alignment; Perioperative efficiency	AI group showed significantly reduced operative time and length of stay (p < 0.05). 3D guides yielded less pain and bleeding. Alignment accuracy similar (<3° deviation). No significant VAS/HSS differences at 3 months.
Lan et al., 2024—Assessment of preoperative planning and intraoperative accuracy of the AIKNEE system for total knee arthroplasty [33]	Retrospective observational cohort (64 patients)	AIKNEE—3D planning with automatic axis recognition and prosthesis simulation	Evaluate the accuracy of sizing, alignment, and correlation with postoperative outcomes	Prosthesis size prediction, deviations in mFTA/LDFA/MPTA, number of insert trials, ROM, VAS, KSS	Alignment accuracy + clinical outcomes	Sizing accuracy: 48% femur, 73% tibia. Alignment within 3° for mFTA (88%), LDFA (92%), MPTA (95%). Significant improvements in pain, ROM, and KSS (p < 0.001). Larger deviations correlated with more pain and lower KSS.
Liao et al., 2024—Efficiency assessment of intelligent patient-specific instrumentation in total knee arthroplasty: a prospective randomized controlled trial [34]	Prospective double-blind RCT (102 patients, 107 knees)	AI-KNEE (3D-UNet + HRNet + CNN) for CT-based planning & PSI fabrication	Evaluate resection accuracy, postop alignment, and perioperative metrics of i-PSI vs. conventional instruments	Resection accuracy (CT), HKA/FCA/JLCA/FSA alignment, surgical time, bleeding, complications	Osteotomy accuracy + alignment	i-PSI improved resection precision of the distal femur (p = 0.032–0.035), produced more neutral HKA/FCA/JLCA (p < 0.05), better FSA (p = 0.005). No bleeding/complication differences; surgical time slightly longer (p = 0.027).
Liu et al., 2024—Development and clinical validation of a deep learning-based knee CT image segmentation method for robotic-assisted TKA [35]	Development + clinical validation study (538 CT scans training/testing; 20 robotic-TKA patients)	DDA-Transformer (CNN + Transformer dual-path double-attention)	Develop a DL CT-segmentation model and evaluate its impact on robotic planning	Segmentation accuracy; sizing; resection precision; alignment (HKA, MPTA, PTS, FPPA)	CT segmentation + robotic planning	High segmentation accuracy; sizing and resection error < 0.5 mm and <0.7°; robotic alignment significantly more accurate vs. manual TKA (p < 0.05).
Vidhani et al., 2024—Automating Linear and Angular Measurements for the Hip and Knee After CT [36]	Technical development + validation on 100 CT scans	Three-stage pipeline: VGG16 + XGBoost (classification); U-Net3+/Attention U-Net/2D TransUNet (segmentation); OpenCV measurements	Automate CT-based classification, segmentation, and pre-op measurements	Classification accuracy; Dice/IoU; comparison vs. manual measurement; mean error; processing time	Automated CT-based measurement	Classification 90.8% (hip)/87.8% (knee); segmentation Dice/IoU > 0.95. Mean time 2.58 ± 1.92 min/case. No significant difference vs. manual (p > 0.05). Mean errors: FV 3.72°, sulcus 2.44°, TT–TG 2.34 mm, PCA 0.70°, AA 2.01°.
Burge T., 2022—Performance and Sensitivity Analysis of an Automated X-Ray Based Total Knee Replacement Mass-Customization Pipeline [37]	Computational validation on X-ray (78 pts) + DRR (147)	U-Net segmentation + PDM/SSM + CAD	Evaluate the accuracy and sensitivity of AI-based custom implant design	RMSE reconstruction; component RMSE; over/underhang (≥3 mm); sensitivity to alignment/scale	AI-based custom implant planning from X-ray	Accurate pipeline (~1 mm RMSE), OUH < 3 mm in most cases; robust across subjects but lower in under-represented ethnicities; sensitive to X-ray alignment/scale.
Factor et al., 2024—Validating a Novel 2D-to-3D Knee Reconstruction Method on Preoperative TKA Patient Anatomies [38]	Technical validation; 18 OA patients (CT + calibrated AP/LAT X-ray)	Neural network 2D→3D reconstruction	Validate 3D reconstruction from 2D radiographs for pre-op planning	RMSE global; landmark accuracy; osteotomy contour RMSE; axis deviation vs. human	3D reconstruction from X-ray	Global RMSE: 0.93 ± 0.25 mm (femur), 0.88 ± 0.14 mm (tibia). Landmark RMSE ~0.5 mm. Osteotomy RMSE ~0.7 mm. Anatomical axis deviation: TEA 1.89°, PCA 1.78°, MLTA 2.82°—comparable to inter/intra-observer variability.
Fernandes et al., 2023—Accuracy, Reliability, and Repeatability of a Novel AI Algorithm Converting 2D Radiographs to 3D Bone Models for TKA [39]	Preclinical validation on 5 cadaveric knees (AP/LAT X-ray + CT + manual measures)	AI 2D→3D reconstruction	Assess accuracy, reliability & repeatability vs. CT/manual	MAE 2D→3D vs. CT; MAE vs. manual; ICC inter/intra-observer	3D reconstruction for planning	Excellent accuracy: MAE < 2 mm in 9/12 anatomical parameters. All ICC > 0.90—high reliability and reproducibility.
Wang et al., 2023—Automatic extraction of medical feature points using PointNet++ for robot-assisted knee arthroplasty [40]	Technical development/validation; 20 CT point clouds (10 patients)	Modified PointNet++ (Point_RegNet)	Automate landmark extraction for robotic pre-op registration	Mean error of 3 feature points; processing time; comparison vs. other networks	Automatic anatomical landmark extraction	Mean error < 5 mm, 1 mm lower than manual marking; SD < 1 mm. Extraction < 3 s per point—much faster than manual.
Yi et al., 2020—Automated detection & classification of knee arthroplasty using deep learning [41]	Retrospective balanced sets (native vs. TKA; TKA vs. UKA; 2 TKA models)	ResNet-18/ResNet-152 transfer learning + CAM	Detect prosthesis; classify TKA vs. UKA; differentiate implant models	AUC, sensitivity, specificity, PPV, NPV, CAM maps	Automatic implant detection for revision strategies	AUC = 1 with 100% sensitivity & specificity in all tasks; CAM accurately localised prosthetic components.
Tiwari et al., 2022—Application of deep learning algorithm in automated identification of knee arthroplasty implants from plain radiographs using transfer learning models: Are algorithms better than humans? [42]	Retrospective 521 AP/LAT X-rays; 7 DL models compared	Transfer learning (VGG-16, MobileNet, ResNet50, etc.)	Identify TKA implant model vs. experts	Accuracy, precision, recall, loss, human comparison	Preoperative implant identification in revision	VGG-16 accuracy 95.5%, precision 98.4%; five models > 90%. Human specialists ~78%.
Karnuta et al., 2021—Artificial Intelligence to Identify Arthroplasty Implants From Radiographs of the Knee [43]	Multicentre retrospective (682 AP X-rays, 424 pts)	InceptionV3 CNN (1000 epochs)	Automatic manufacturer/model recognition	AUC; accuracy; sensitivity; specificity; PPV; NPV	Implant recognition for revision TKA	AUC 0.992; accuracy 98.9%; SE 94.6%; SP 99.4%. Many implant categories reached 100% accuracy.
Burge T., 2022—A computational tool for automatic selection of total knee replacement implant size using X-ray images [44]	Computational comparison (AP + LL X-rays, 78 pts) vs. MRI ground truth	U-Net + 2D→3D pipeline + automatic sizing	Develop an automatic predictor for femoral/tibial size	RMSE; over/underhang; correct prediction rate; ±1 size accuracy	X-ray-based pre-op sizing	Femur: 77.95% RMSE accuracy; ±1 size 99.7%. Tibia: 80.51% RMSE; ±1 size 99.7%. RMSE ~1 mm.
Kunze et al., 2022—Machine learning algorithms predict within one size of the final implant ultimately used in total knee arthroplasty with good-to-excellent accuracy [45]	Multicentre retrospective (11,777 TKA, 2012–2020)	ML (SVM, ENPLR, RF, XGB, SGB)	Predict final implant size using demographics only	Exact accuracy; ±1 size; MAE; RMSE	Pre-op sizing from demographic data	Best results: Femur exact 42.2%, ±1 88.3%, MAE 0.73. Tibia exact 43.8%, ±1 90.0%, MAE 0.70. Good ±1 performance using demographics alone.
Yue et al., 2022—Prediction of knee prosthesis using patient gender and BMI with non-marked X-ray by deep learning [46]	Retrospective 308 pts (AP/LAT X-ray + anthropometrics)	ResNet-18 + ECOC + transfer learning	Predict femoral/tibial size pre-operatively	Accuracy per component; baseline vs. optimised vs. ECOC	Preop sizing from X-ray + physical data	ECOC best: 88.23% femur, 86.27% tibia—superior to baseline/optimised, comparable or superior to surgeons.
Kunze et al., 2021—Validation and performance of a machine-learning derived prediction guide for total knee arthroplasty component sizing [47]	Retrospective 17,283 pts, 80/20 split, 5 model comparison	SGB, RF, SVM, XGB, Elastic-Net Logistic Regression	Predict femoral/tibial size using demographics only	Accuracy (±4 mm/exact/±1), MAE, RMSE, R2; variable importance	Demographic-based sizing	SGB best: femur ±1 95.0%, tibia ±1 97.8%; ±4 mm ≈83%. Sex strongest predictor.
Yu et al., 2024—Development of an artificial intelligence model for predicting implant size in total knee arthroplasty using simple X-ray images [48]	Retrospective cohort (714 patients, 1412 AP + LL X-rays)	ResNet-101 CNN	Develop AI to predict femoral/tibial implant sizing using X-rays only (no demographics)	Sizing accuracy; micro-F1 for exact size and ±1 size	Preoperative sizing	Exact-match prediction: micro-F1 = 0.91 (femur) and 0.87 (tibia). ±1 size accuracy: 0.99 (femur) and 0.98 (tibia). Outperformed traditional templating and demographic-based models.
Park et al., 2024—Clinical validation of a deep learning-based approach for preoperative decision-making in implant size for total knee arthroplasty [49]	Retrospective cohort (234 patients)	YOLO-v4 + CNN (detection + classification) on AP X-rays	Validate DL model for automatic femoral and tibial size prediction	Accuracy, match with implanted component, Spearman rho, ±1 accuracy	Preoperative sizing	Model significantly superior to manual templating (femur 89.32% vs. 61.54%; tibia 90.60% vs. 68.38%). ±1 accuracy ~98%. High concordance (rho femur 0.91; tibia 0.94).

.

Model creation and segmentation

Consistently with the findings reported for hip arthroplasty, across the included studies artificial intelligence was applied to the creation of patient-specific anatomical models through automated segmentation and three-dimensional reconstruction, primarily using CT-based workflows, with emerging approaches enabling 3D reconstruction from standard radiographs [31,32,33,34,35,36,37,38,39,40]. These AI-based pipelines demonstrated high geometric accuracy and reproducibility when compared with manual measurements or CT reference standards, with errors generally within the range of inter- and intra-observer variability [36,38,39]. Automated extraction of anatomical landmarks and axes further supported standardisation of preoperative planning and registration processes, including applications integrated into robotic-assisted workflows [35,40]. Studies based on radiograph-derived 2D-to-3D reconstruction provided proof-of-concept evidence for reducing reliance on CT imaging, although validation remained primarily technical or preclinical [37,38,39].

Component sizing

Component sizing represented one of the most extensively investigated applications of AI in TKA planning. CT-based AI systems demonstrated improved accuracy in predicting femoral and tibial component sizes compared with conventional templating, with consistent performance across different study designs [31,33,34,35]. Parallel lines of research explored radiograph-based deep learning models, achieving high accuracy for exact or ±1 size prediction using plain AP and lateral radiographs [48,49]. Additional approaches relied on demographic or anthropometric data, showing moderate exact-size accuracy but good performance within ±1 size ranges, thereby providing probabilistic preoperative guidance rather than anatomically driven planning [45,46,47]. Collectively, these findings reflect improvements in technical sizing accuracy, with varying levels of anatomical specificity depending on the input data used.

Alignment and biomechanical planning

Several studies evaluated the role of AI-assisted planning in achieving accurate limb alignment and biomechanical reconstruction. Reported benefits included improved control of mechanical alignment parameters such as HKA, mFTA, LDFA, and MPTA, as well as more precise osteotomy execution and joint line restoration [31,32,33,34,35]. AI-based patient-specific instrumentation and segmentation-driven planning demonstrated higher resection accuracy and alignment consistency compared with conventional techniques, while integration with robotic systems further enhanced geometric precision [34,35]. Although some studies reported correlations between alignment accuracy and early clinical scores, these observations primarily reflect radiographic and biomechanical optimisation rather than validated long-term clinical benefit.

Prediction and decision-support functions

Beyond geometric planning tasks, AI-assisted approaches were applied to predictive and decision-support functions across the preoperative workflow. Several clinical studies reported associations with reduced operative time, blood loss, or perioperative resource use, although findings were not uniform across all techniques and study designs [31,32,34]. Additional applications included automated identification and classification of knee arthroplasty implants from radiographs, achieving high diagnostic accuracy and supporting preoperative planning in revision settings [41,42,43]. Machine learning models were also developed to provide probabilistic guidance on implant selection based on imaging or demographic inputs, facilitating preoperative decision-making rather than demonstrating direct improvement in clinical outcomes [45,46,47]. Overall, these applications were primarily discussed as supportive or exploratory tools, frequently evaluated without long-term follow-up.

3.3. Shoulder Arthroplasty

The results concerning the use of artificial intelligence in preoperative planning for shoulder arthroplasty are reported in

Table 3. Shoulder arthroplasty results.

First Author/Title	Study Type	Study Design	AI Objective	Outcomes Assessed	Main Results
Rajabzadeh-Oghaz et al., 2024— Impact of Deltoid Computer Tomography Image Data on the Accuracy of Machine Learning Predictions of Clinical Outcomes after Anatomic and Reverse Total Shoulder Arthroplasty [50]	Multicentre prospective cohort study	aTSA + rTSA; Preoperative CT image analysis of 1057 patients undergoing shoulder arthroplasty with a single-platform implant: 799 primary rTSA and 258 primary aTSA. The deltoid muscle was segmented to extract 15 three-dimensional features used in machine-learning models.	To evaluate the impact of three-dimensional deltoid characteristics on clinical outcomes after aTSA and rTSA, and to determine whether their integration improves ML postoperative outcome prediction capabilities.	Clinical outcomes (1–5 years): Active ROM: abduction, forward elevation, external rotation, internal rotation (IR score); Pain VAS; Global Shoulder Function; Constant score; ASES; SAS score. Statistical outcomes: MAE of models.	Integrating deltoid image data into ML models improved prediction accuracy compared with ML without imaging, particularly for abduction and forward elevation prediction after aTSA and rTSA.
Caprili et al., 2025—Assessing the accuracy of a machine learning prediction for 2 different shoulder prostheses: an external validation study [51].	Retrospective monocentric cohort study with external validation of a machine-learning model	rTSA; 90 patients undergoing rTSA, divided into two groups based on implant type, applying a machine-learning algorithm (Predict+) to preoperative data (19 variables).	To validate a predictive platform based on machine learning (Predict+), apply it to preoperative data, and compare predicted outcomes with postoperative clinical outcomes.	Clinical outcomes (3–6 months, 1 year, 2 years): Pain VAS; active forward elevation; active abduction; active external rotation; functional internal rotation. Analytical outcomes: MCID verification (VAS, FE, AB, ER); MAE of the two groups vs. MAE internal validation.	Predict+ showed good accuracy in predicting VAS and forward elevation in both groups, with minimal differences between expected and observed values, and with valid results also for external and internal rotation (despite the absence of an MCID for the latter). In all assessed outcomes, MAE values were better or similar to internal validation, confirming predictive reliability up to 2 years, even for an implant different from the training one.
Crutcher et al., 2025—An accelerated deep learning model can accurately identify clinically important humeral and scapular landmarks on plain radiographs obtained before and after anatomic arthroplasty [52].	Retrospective cohort validation study of a deep-learning model	aTSA; 120 patients undergoing anatomical shoulder hemiarthroplasty. Manual annotation of 240 AP radiographs (pre- and postoperative) with 11 bony landmarks.	To evaluate the accuracy of a deep-learning model in identifying scapular and humeral landmarks and calculating 14 anatomical measurements, comparing them with expert surgeon annotations.	Deviation between AI-identified landmarks (DLM) and surgeon-identified landmarks (SI). Accuracy in 14 scapular, humeral and gleno-humeral measurements. Analysis of differences between cortical vs. non-cortical and scapular vs. humeral points.	The model achieved a mean deviation of 1.9 ± 1.9 mm versus the surgeon. Scapular landmarks were more accurate than humeral (1.5 vs. 2.1 mm). Anatomical measurements derived from the DLM showed a mean deviation of 2.9 ± 2.7 mm. Despite a limited training dataset, the model demonstrated high efficiency, reduced observer bias, and potential for large-scale radiographic analysis.

. The three studies analysed collectively explore the growing role of artificial intelligence in preoperative planning for shoulder arthroplasty, offering increasingly accurate predictive tools that may prove valuable in preventing complications.

Model creation and anatomical feature extraction

In shoulder arthroplasty, artificial intelligence was primarily applied to the extraction and analysis of patient-specific anatomical features from preoperative imaging. Rajabzadeh-Oghaz et al. (2024) [50] demonstrated that three-dimensional deltoid characteristics derived from preoperative CT scans—including normalised muscle volume, fatty infiltration, sphericity, and flatness—represent highly informative variables for outcome prediction after both anatomic and reverse total shoulder arthroplasty. These AI-derived anatomical features were shown to be more predictive than traditional demographic or clinical variables, supporting the role of advanced imaging-based model creation in capturing anatomical risk factors that are otherwise difficult to quantify.

Prediction of functional outcomes and risk stratification

Predictive modelling constituted a central application of AI in the included shoulder studies. Caprili et al. (2025) [51], through external validation of the Predict+ tool, showed that machine learning-based models can provide reliable predictions of functional outcomes after reverse total shoulder arthroplasty, maintaining consistent accuracy across different follow-up time points and across implants not included in the training dataset. These findings suggest that AI-based predictive algorithms may enable early identification of patient-specific functional risk profiles, supporting personalised strategies aimed at mitigating unfavourable postoperative recovery patterns or biomechanical compensations.

Measurement standardisation and planning precision

A complementary application of AI was observed in the automation and standardisation of preoperative measurements. Crutcher et al. (2025) [52] investigated deep learning-based automatic landmark recognition and reported a substantial reduction in inter-observer variability compared with manual assessment. The model achieved landmark recognition with a mean deviation of approximately 1.9 mm relative to expert surgeons, with statistically significant improvements in most comparisons. This supports the role of AI in increasing the reliability and reproducibility of preoperative evaluation, improving planning precision, and potentially reducing errors associated with component malpositioning.

Overall interpretation

Taken together, these studies indicate that AI applications in shoulder arthroplasty extend beyond measurement automation or isolated predictive tasks, contributing to a more objective and reproducible paradigm of preoperative planning. Although none of the included studies directly evaluated complication rates, the identified elements—functional outcome prediction, anatomical risk factor detection, and measurement standardisation—represent potential mechanisms for anticipating biomechanical challenges, planning complexity, and variability in postoperative functional response [50,51,52]. However, it must be noted that the available evidence remains limited, both in terms of the number of studies and methodological heterogeneity across clinical contexts.

Limitations of the evidence

This scoping review has several limitations that should be considered when interpreting the mapped evidence. The included studies were highly heterogeneous in terms of design, populations, imaging modalities, AI architectures, and evaluated endpoints, and were predominantly retrospective and monocentric, often based on small or selected cohorts and limited numbers of surgeons or implant models, which may restrict generalisability. Clinical and functional outcomes were not consistently defined as primary endpoints across studies and, when reported, were frequently secondary, short-term, or assessed using non-uniform measures, limiting comparative interpretation. The follow-up duration was commonly short or variable, precluding robust evaluation of long-term outcomes or complications. Methodological constraints related to imaging acquisition, input standardisation, and segmentation procedures, as well as the use of restricted or unbalanced datasets and limited external or multicentric validation, further contribute to variability across studies. These aspects reflect the exploratory and developmental nature of the current evidence base and should be taken into account when interpreting the scope and applicability of the reported findings.

4. Discussion

Within this context of evolving but still methodologically constrained evidence, broader ethical and regulatory considerations become particularly relevant. In light of the findings presented and of what has been discussed so far regarding the implementation of orthopaedic techniques through supervised use of artificial intelligence, the need for a clear and effective regulatory framework becomes evident—one that enables the use of such technologies in a targeted and functional manner, amplifying their purpose and efficiency.

Indeed, given the already widespread global use of AI and the issues arising as a consequence, it is necessary to establish rules that ensure the responsible use of these resources, especially in view of their impact on individuals’ lives and health. Applications of artificial intelligence in healthcare have the potential to reshape the physician–patient relationship, affecting the principle of informed consent. It becomes necessary, therefore, to determine in which circumstances informed consent principles should be applied when clinical procedures involve artificial intelligence, and to what extent clinicians have the responsibility to explain to patients the complexities of AI and its role in medical decision-making.

In this context, recent studies have highlighted how AI models explicitly designed to promote fairness, transparency, and robustness—such as demographic parity-based approaches for the multi-quantification of maxillofacial traits and conceptual frameworks integrating confidence, adaptability, stability, explainability, and security (CASES)—represent concrete efforts to address the ethical, regulatory, and clinical challenges associated with the adoption of artificial intelligence in medicine [53,54].

From an ethical standpoint, the opacity of many machine learning algorithms—so-called black-box models, especially deep neural networks—raises concerns regarding transparency and the ability to reconstruct the decision-making process leading to a given output. These systems may provide predictive results without clarifying the underlying criteria, and thus risk generating suboptimal clinical care. In such cases, a key question arises: to what extent must a physician disclose to the patient that they may not fully understand the reasoning behind an AI-generated diagnosis or treatment recommendation?

This issue is closely linked to personal data protection, as established by Article 22 of the GDPR, which grants individuals the right not to be subjected solely to automated decision-making—including profiling—when such decisions produce legal effects or significantly affect the person. In orthopaedic applications, existing regulatory frameworks such as the U.S. HIPAA and the European GDPR provide fundamental protection but struggle to fully address the nuanced privacy challenges introduced by multimodal AI [55].

To mitigate these risks, researchers have developed advanced technical solutions, including differential privacy, federated learning, homomorphic encryption, and swarm learning. These approaches enable collaborative model training while preserving data confidentiality—using encrypted data obfuscation, decentralised learning protocols, and blockchain-secured computation. Emerging technologies such as edge computing also offer promising perspectives by processing sensitive information closer to its source, reducing transmission risks, and maintaining patient privacy [56].

From a medico-legal perspective, where every assessment must be traceable and justifiable, algorithmic opacity represents a critical limitation, as it compromises the ability to explain decisions in forensic, judicial, or insurance settings. Explainable AI (XAI) has become a major research priority to address this challenge, yet it remains far from full clinical maturity. Moreover, AI-based techniques are not widely understood by most orthopaedic surgeons and are not taught with the same depth of intuition and critical reasoning as traditional statistical methodology. Conversely, a greater understanding of the science underlying AI could increase trust among surgeons and patients, thereby expanding its clinical applicability and amplifying its impact on healthcare value and quality [57].

From an ethical and governance perspective, embedding XAI within clinical workflows is essential to sustain trust and professional accountability in AI-assisted care. By providing interpretable, context-aware explanations at the point of use, XAI can support real-time clinical reasoning, enabling surgeons to critically evaluate algorithmic outputs rather than passively rely on them. This integration preserves meaningful human oversight, reinforces the clinician’s role as the final decision-maker, and mitigates the risk of uncritical trust in opaque systems. When appropriately designed, XAI strengthens the physician–patient relationship and establishes a necessary bridge between algorithmic assistance and medico-legal responsibility.

From a legal standpoint, the issue of liability emerges. Although patients are often those most exposed to the risks associated with machine learning tools, clinicians likewise face intrinsic medico-legal vulnerability. Current machine learning tools in surgery remain assistive—not substitutive—providing outcome and complication prediction to support shared decision-making. However, future scenarios in which algorithms provide real-time intraoperative recommendations will raise questions regarding whether responsibility should fall on the surgeon or on the developer of the AI system [58]. If a surgeon chooses to disregard an AI-generated recommendation, how should such a decision be documented and justified? Discussions surrounding autonomous robotic surgery further highlight concerns for injury or death directly resulting from algorithm-driven surgical error [59,60].

Thus, when harm arises from a decision informed or suggested by an artificial intelligence system—whether fully or partially algorithm-mediated—it remains unclear who should be held accountable: the clinician, the consultant, the software developer, or the institution that adopted the tool. Current legislation does not yet provide univocal solutions, making revision of existing regulatory frameworks necessary, with specific attention to AI-assisted decision-making. In medical and forensic contexts—where the tolerance for error must be minimal—defining certification standards and algorithmic validation protocols is a priority. In this sense, both national and international scientific communities have recently begun to engage proactively in this field.

4.1. European Context

In May 2016, a first draft report was issued by the JURI Committee of the European Parliament following a hearing involving several experts in fields related to artificial intelligence. The aim was to develop a legislative framework for the regulation of AI. Within this context, the centrality of civil liability for damages caused by robots (the term robot referring broadly to AI systems as well) was emphasised, calling for European Union-wide analysis to ensure a consistent level of efficiency, transparency, and legal certainty for citizens, consumers, and businesses alike. That resolution already urged the Commission to present a legislative proposal concerning the legal issues surrounding the development of AI, accompanied by non-legislative instruments such as guidelines and codes of conduct [61].

The European legal basis resides in Article 114 of the Treaty on the Functioning of the European Union (TFEU), which establishes provisions concerning products and services employing AI technologies and autonomous AI systems. Some Member States have begun to draft national legislation ensuring that AI is developed and applied safely and in accordance with fundamental rights obligations [62].

These regulatory strategies culminated in the proposal published by the European Commission on 21 April 2021—the Artificial Intelligence Act—which introduces a crucial risk-based approach to AI regulation. The proposal classifies AI systems into categories with distinct regulatory requirements, namely prohibited, high-risk, and other AI systems, with the latter including both limited-risk and minimal-risk systems. Classification may evolve over time in order to balance user and developer rights while avoiding harm to scientific progress. High-risk systems explicitly include medical and legal applications, as these operate autonomously and carry the potential to cause significant harm to one or more individuals. This potential is determined by the seriousness of the possible damage, the degree of automated decision-making, the likelihood of risk materialisation, and the operational context in which the AI system is deployed.

Under the proposal, high-risk systems must undergo risk assessment, rely on high-quality datasets, and ensure output traceability. They must comply with technical documentation, data storage, and transparency requirements, as well as provide human oversight and a high level of robustness, safety, and accuracy.

From a civil-liability standpoint, the proposed regime is centred on the assumption that the operator is the party responsible for damage or harm caused by an AI system. The European Parliament considers that, given the inherent complexity and connectivity of AI systems, the operator is often the first point of reference for affected individuals. If multiple operators are involved, they are jointly liable. In such circumstances, the allocation of responsibility shares must reflect each operator’s degree of control over the operational risk, with internal recourse proceedings used to recover proportional contributions [63].

Regarding the mental (fault) element, high-risk AI systems fall under strict liability—meaning operators cannot evade responsibility by asserting due diligence or by claiming that harm was caused by the autonomous behaviour of the AI system. Operators are exempt only in the case of force majeure.

Taken together, this legislative trajectory illustrates only part of the evolving process through which the European Union is shaping AI regulation. The proposed European Parliament Regulation on Civil Liability for AI Systems and the Commission’s Artificial Intelligence Act represent just two initial pillars, complemented by ongoing work on intellectual property and criminal liability in relation to AI technologies. What emerges clearly, however, is the pervasive impact that civil-liability regulation for AI systems will exert across multiple domains of public and private life.

4.2. Italian Context

Until the adoption of the current Italian law regulating AI, physicians were protected from the consequences of professional liability so long as they adhered to the standard of care—the “safest” way to use medical artificial intelligence. However, this approach tended to discourage clinicians from fully exploiting the potential of AI systems, favouring conservative use and limiting technological innovation [64].

Following the European momentum generated by the AI Act, Law No. 132/2025, enacted on 10 October 2025, entered into force with the objective of integrating the European AI Regulation and establishing the first national legal framework governing artificial intelligence and its main fields of application while upholding fundamental rights and constitutional principles. The law is based on a human-centric, transparent, and safe use of AI, with particular emphasis on innovation, cybersecurity, accessibility, and data protection. It intervenes organically across multiple sectors that may benefit from AI, establishing guarantees of traceability, human accountability, and the centrality of final decision-making by a natural person.

Regarding governance, the law designates the National Cybersecurity Agency (ACN) and the Agency for Digital Italy (AgID) as competent national authorities: ACN oversees system adequacy and safety—with inspection powers—while AgID handles notifications and promotes safe use cases for citizens and businesses within a stable inter-institutional coordination framework. The legislation also introduces a strategic planning mechanism: the National AI Strategy will be drafted and updated biennially by the Department for Digital Transformation of the Presidency of the Council of Ministers, supported by ACN and AgID and with the involvement of sectoral authorities. Transparency is reinforced through annual reporting to Parliament.

To accelerate competitiveness and adoption, the law establishes a €1 billion investment programme for startups and SMEs operating in AI, cybersecurity, and emerging technologies, supporting technological transfer and strategic industrial development [65].

4.3. United States Context

While the European Union has adopted the AI Act to regulate artificial intelligence and protect individual rights, the regulatory landscape in the United States remains in development, shaped by multiple federal and state-level proposals. The U.S. approach is more flexible and innovation-driven, aiming to foster technological growth with fewer binding constraints compared to the European model [66].

Among the existing federal laws addressing AI is the National AI Initiative Act of 2020 (updated in 2023), designed to expand research and development in AI and establish the National Artificial Intelligence Initiative Office, responsible for oversight and implementation of the national AI strategy.

A major federal intervention is the White House Executive Order on Artificial Intelligence, issued on 30 October 2023 [67], addressing multiple sectors and built upon the principle that harnessing AI for public benefit requires risk mitigation. The Executive Order primarily targets federal agencies and developers (industry), reflecting a business-centred approach—distinct from the EU’s AI Act, which places human rights and individual protection at its core. The Order outlines eight guiding principles and requires AI system developers to collaborate with federal authorities to ensure safety, reliability, and data protection.

Despite a more liberal, innovation-oriented regulatory philosophy, recent federal activity demonstrates a clear intent to regulate AI. A notable example is the U.S. Senate Hearing on AI (September 2023), which explored possible regulatory strategies, including mandatory licensing and the creation of a federal agency dedicated to AI oversight.

Several federal legislative proposals have emerged in recent years. Prominent bills (not yet formally enacted) include the following:

• SAFE Innovation AI Framework—guidelines for developers, companies, and policymakers. While not law, it provides foundational direction for future federal AI legislation, balancing innovation and rights protection.
• REAL Political Advertisements Act—regulates generative AI use in political campaigns.
• Stop Spying Bosses Act—limits employee surveillance using ML/AI-based monitoring tools.
• NO FAKES Act—bipartisan bill that restricts the creation and use of generative AI replicas of unconsented faces, voices, and identities, targeting the widespread issue of deep-fake misuse.
• AI Research, Innovation, and Accountability Act—promotes transparency, responsibility, and safety in high-risk AI through testing requirements and mandatory corporate transparency reports.

Given this regulatory trajectory, the future likely involves a decentralised oversight ecosystem composed of federal agencies with delegated authority or state-level AI statutes, reflecting the U.S. legal tradition of distributed risk management rather than a singular codified framework.

In healthcare, regulations already exist to protect sensitive patient data—such as Institutional Review Boards (IRBs) and the Health Insurance Portability and Accountability Act (HIPAA). However, the increasing complexity of ML/AI requires legislative updates designed specifically for clinical algorithmic tools. This may require ongoing public engagement on data use, including how ML is applied to patient care or foreseeable limitations. As AI becomes embedded in routine medical decision-making and ambient intelligence systems, maintaining transparency may prove increasingly challenging [68].

The legal debate in the U.S. concerns both AI design and deployment, including how design choices themselves may create direct or indirect safety and usability issues [69]. Issues of liability and accountability remain a primary barrier to AI integration in surgery. Surgical AI systems must be federally certified and regulated prior to widespread clinical use. The FDA approved its first autonomous diagnostic AI in 2018 for diabetic retinopathy screening [70] and has since authorised several hundred additional AI/ML-enabled devices. However, continuously adaptive ML models introduce new regulatory challenges, requiring anticipation of algorithm evolution over time.

In 2018, the U.S. DoD funded the FORwARD Program (Foundational Research for Autonomous, Unmanned, and Robotics Development of Medical Technologies), boosting momentum in robotic surgery. Sullivan et al. [59] propose a regulatory model for autonomous surgery inspired by guidelines for self-driving vehicles, advocating a surgeon-in-the-loop framework mirroring the human-in-the-loop standard established by the Society of Automotive Engineers. When AI operates as an assistive tool rather than a replacement, medico-legal responsibility remains more manageable. As long as clinicians retain oversight, malpractice liability continues to fall primarily on human supervisors—although debate persists regarding the shared responsibility of engineers and developers [71].

Ethical challenges regarding autonomy, responsibility, and informed consent must be managed as AI becomes further integrated into surgical workflows. There is also concern that AI may exacerbate healthcare disparities through algorithmic bias, underscoring the need for equitable regulation and vigilant oversight. Central to this effort is the inclusion of patients, whose perspectives are critical to the development of AI systems that are not only advanced but also ethically aligned. Going forward, a collaborative, multidisciplinary approach—involving ethicists, clinicians, researchers, regulators, and patient communities—will be essential. Establishing transparent decision-making processes and open dialogue will ensure that technological progress remains aligned with core medical values: human dignity, autonomy, and trust [72].

Given historic and ongoing disparities in healthcare access, AI adoption must be evaluated carefully to avoid amplifying structural inequality [60]. Since ML/AI performs best when trained on large, heterogeneous datasets drawn from multiple healthcare systems, aggregating such data from diverse EHR sources raises issues of privacy, security, and patient interest protection. When private corporations drive AI development, with profit tied to data acquisition, concerns arise around the potential exploitation of sensitive patient information [73].

Existing U.S. regulations, such as IRBs and HIPAA, already provide baseline protection, but technological evolution demands dedicated legislative refinement. Public engagement regarding data usage, consent, and expected limitations may become increasingly necessary—though challenging—as AI becomes pervasive in clinical environments and ambient-care models [68]. Finally, equitable AI deployment requires ensuring that benefits are not restricted by socioeconomic status or geography. Greater efforts toward inclusivity and dataset diversity are needed to mitigate geographic and demographic bias in model training and implementation [74].

4.4. Chinese Context

In contrast to the European rights-based regulatory model and the decentralised, innovation-driven approach of the United States, the Chinese framework for artificial intelligence is characterised by a state-centric, sector-based, and security-oriented regulatory strategy. China has not adopted a single comprehensive AI statute comparable to the European Artificial Intelligence Act; instead, it relies on a layered system combining national strategic plans, administrative regulations, ethical guidelines, and sector-specific legislation, including for healthcare and medical devices.

The foundational policy document is the New Generation Artificial Intelligence Development Plan (2017), which identifies healthcare as a strategic priority for AI deployment while emphasising controllability, safety, and alignment with social governance objectives. In the medical field, AI systems—particularly those used in diagnostics, robotic surgery, and clinical decision support—are regulated primarily through existing medical device legislation under the supervision of the National Medical Products Administration (NMPA). AI-enabled medical software and surgical robots, including applications relevant to orthopaedic surgery, are classified and authorised as medical devices according to their risk level, with requirements for clinical evaluation, safety, and efficacy testing [75].

From a liability perspective, Chinese law treats AI systems as technical tools rather than autonomous legal actors. Consequently, responsibility for harm caused by medical AI generally rests with healthcare institutions and clinicians, or with manufacturers in cases of product defects, under existing tort and product-liability rules. Unlike the EU’s strict-liability regime for high-risk AI, China has not introduced a dedicated civil-liability framework specific to AI; instead, liability is allocated through traditional fault-based or product-liability mechanisms, with a strong emphasis on institutional accountability and regulatory compliance.

Data governance plays a central role in the Chinese model. The Cybersecurity Law, Data Security Law, and Personal Information Protection Law impose stringent requirements on the processing, storage, and localisation of sensitive health data, directly affecting the development and deployment of medical AI systems. These rules prioritise national security, public interest, and state oversight, often limiting cross-border data flows and shaping how machine learning models can be trained and updated. In clinical practice, AI is formally positioned as an assistive technology under human supervision, and autonomous medical decision-making without physician oversight is not permitted [76].

Overall, the Chinese approach reflects a regulatory philosophy that prioritises technological advancement under strong governmental control, public safety, and data sovereignty. While this model enables rapid large-scale adoption of AI in healthcare, including robotic and AI-assisted surgery, it offers less explicit protection for individual rights and less transparency in liability allocation when compared to the European framework (

Table 4. International comparison of regulations governing the use of artificial intelligence (AI).

Dimension	European Union	Italy	United States	China
Regulatory model	Centralised, rights-based, risk-based (AI Act)	National framework aligned with the EU AI Act	Decentralised, sector-based, innovation-driven	Centralised, state-centric, sector-based
Core legal instrument	Artificial Intelligence Act (2021 proposal; risk-based)	Law No. 132/2025 integrating the EU AI Act	Executive Orders, federal initiatives, sectoral laws	National AI plans + medical device and data laws
Approach to risk	Explicit classification (prohibited, high-risk, limited/minimal risk)	Mirrors the EU risk classification	No unified risk taxonomy	Risk assessed via sectoral regulation (e.g., medical devices)
Healthcare/surgical AI	Explicitly classified as high-risk	High-risk; strong human-oversight requirement	Regulated via FDA approval pathways	Regulated as medical devices by NMPA
Human oversight	Mandatory for high-risk AI	Mandatory; final decision by a human	Strongly encouraged; surgeon-in-the-loop model	Mandatory; AI as an assistive tool only
Civil liability regime	Operator-centred strict liability for high-risk AI	Aligned with the EU approach	Predominantly fault-based malpractice liability	Traditional tort/product liability; no AI-specific regime
Role of developers/manufacturers	Regulated via conformity, transparency, and documentation duties	Subject to EU-derived obligations	Limited direct liability; growing debate	Product-liability exposure under existing law
Data protection	GDPR-based, rights-oriented	GDPR + national safeguards	HIPAA, IRBs, fragmented data protection	Strong data localisation and state oversight
Regulatory philosophy	Protection of fundamental rights and legal certainty	Human-centric and innovation-supportive	Innovation-first with gradual regulation	Technological development under state control
Transparency & accountability	High (documentation, traceability, reporting)	High; parliamentary reporting	Variable; sector-dependent	Lower public transparency; strong administrative control

).

5. Conclusions

The integration of artificial intelligence into preoperative planning for joint arthroplasty represents a major evolutionary step, with the potential to improve accuracy, personalisation, and clinical predictive capacity. However, despite its promise, large-scale adoption requires a structured pathway involving clinical validation, shared performance metrics, and model standardisation in order to ensure algorithmic reliability, reproducibility, and transparency [77].

At the same time, current technical and scientific evidence provides a foundation for addressing the medico-legal and regulatory implications arising from AI-assisted prosthetic planning. The development of European legislation, particularly the AI Act and the updated regulatory framework for medical devices, necessitates careful consideration of the responsibilities borne by clinicians, manufacturers, and AI developers. These regulatory perspectives differ substantially from those in the United States, where risk governance and legislative strategy follow partially divergent principles.

In this rapidly evolving landscape, continuous dialogue between scientific communities, legal experts, regulatory authorities, and industry stakeholders is essential to ensure that the potential of AI in arthroplasty planning translates into real improvements in patient care while safeguarding patient safety, clarifying accountability, and ensuring regulatory compliance. The ultimate goal is to integrate technological innovation into a model of care that remains sustainable, ethical, and legally sound.