[Orthopedic Surgery and Physical Medicine]{.chapter-title}

doi:10.5281/zenodo.18251405

Orthopedic Surgery and Physical Medicine

Orthopedic AI spans image interpretation, surgical planning, perioperative decision support, postoperative communication, and functional measurement. In 2018, FDA authorized OsteoDetect through the De Novo pathway for a narrow adjunctive use: helping clinicians identify and highlight distal-radius fractures on adult posteroanterior and lateral wrist radiographs. That product-specific authorization cannot be transferred to children, other anatomic regions, or fracture-detection software as a class (FDA DEN180005).

The specialty has more authorized products than prospective clinical evidence. A 2026 review of 70 FDA-authorized orthopedic AI devices found that 22.8% had no reported clinical testing, 68.6% relied on retrospective evaluation, and 8.6% reported prospective trials (Lee et al., 2026). Robotic surgery, navigation, rehabilitation robots, and motion-capture systems are included because they shape the same clinical decisions, but these technologies are not automatically AI.

Learning Objectives

After reading this chapter, you will be able to:

Evaluate AI systems for fracture detection on radiographs and CT
Understand AI applications in preoperative planning and surgical navigation
Assess robotic-assisted orthopedic surgery systems and their evidence base
Navigate rehabilitation robotics and gait analysis AI applications
Recognize performance limitations in pediatric and complex fracture cases
Apply evidence-based frameworks for orthopedic and PM&R AI adoption
Critically assess vendor claims for surgical planning and robotic systems

Essential for orthopedic surgeons, physiatrists, sports medicine physicians, physical therapists, and musculoskeletal care teams.

Chapter Summary (TL;DR)

The Clinical Context:

Orthopedics and physical medicine generate imaging, procedural, sensor, and functional data that can support algorithmic assistance. The evidence is uneven: one product may have an exact FDA-authorized use and reader study, while another may have only retrospective accuracy, vendor-reported performance, or no patient-outcome evaluation.

What Works Well:

Application	Systems	Evidence Level	Key Benefit
Adult distal-radius fracture assistance	OsteoDetect	FDA De Novo record plus retrospective reader study	Improved aided reader discrimination in the authorized wrist-radiograph task
Broader fracture assistance	Product-specific software	Prospective workflow and retrospective accuracy studies	May improve reader sensitivity; patient-outcome benefit is not established
Joint replacement planning and robotics	Procedure- and system-specific tools	Comparative studies with heterogeneous endpoints	Can improve selected planning or alignment measures
Rehabilitation robotics	Upper- and lower-limb devices	RCTs and systematic reviews	Can increase structured repetition; benefit depends on comparator and dose

What’s Emerging:

Application	Status	Notes
Rehabilitation robotics	Mixed clinical evidence	Stroke gait training benefit in selected patients; upper-limb and SCI outcomes depend on dose, impairment, and comparator
Gait analysis AI	Research to clinical	Objective assessment, limited adoption
Surgical navigation	Established adjacent technology	Navigation is not inherently AI; evaluate the exact system and endpoint
Sports injury prediction	Research	Not validated for clinical use

Critical Insights:

Authorization is indication-specific. OsteoDetect’s adult wrist indication does not establish a pediatric or class-wide claim.
Prospective workflow benefit is not patient-outcome benefit. Reader sensitivity, response time, and alignment are intermediate endpoints unless a study measures clinical outcomes.
Robotics and navigation are not synonymous with AI. Classify the technology before assigning AI evidence.
Most FDA-authorized orthopedic AI devices still lack strong prospective evidence. A review of 70 FDA-authorized orthopedic AI devices found 22.8% had no clinical testing, 68.6% relied on retrospective datasets, and only 8.6% reported prospective clinical trials. Clinical testing rates improved across the periods examined (62.2% untested in 2017–2019 and 19.7% in 2022–2024), but gaps remained substantial (Lee et al., 2026).
AAOS has an official AI position statement. It was adopted in February 2025 as an educational statement based on author opinion, not a systematic review or clinical practice guideline (AAOS Position Statement 1193).

The Bottom Line:

Orthopedic AI should be adopted at the level of an exact use case, not a product category. Confirm the authorized indication, evaluate performance in the local population and workflow, measure alert burden and downstream action, and distinguish technical precision from clinical benefit. Rehabilitation robotics and surgical navigation may be valuable even when their relevant mechanism is not machine learning.

Medico-Legal Considerations:

Liability is fact- and jurisdiction-specific; FDA authorization does not allocate civil liability
Document clinically material algorithmic information and the action taken under institutional policy
Confirm whether a robotic or navigation feature is actually AI and whether its use is within labeling
Do not transfer adult fracture evidence to pediatric patients
Use an informed-consent process appropriate to the procedure, material alternatives, and institutional policy

Essential Reading for Orthopedic Practitioners:

FDA’s OsteoDetect De Novo record and reader study
Lee et al. (2026) on clinical testing gaps across 70 FDA-authorized orthopedic AI devices (JAAOS Global Research & Reviews)
Oppenheimer et al. (2023) on prospective integration of fracture-detection software
Li et al. (2026) on a randomized postoperative LLM intervention and its limitations
AAOS Position Statement 1193 on artificial intelligence

Introduction: AI in Musculoskeletal Medicine

Orthopedic surgery and physical medicine have embraced AI applications across multiple domains:

Orthopedic surgery: - Fracture detection on radiographs - Preoperative planning for joint replacement - Robotic-assisted surgical systems - Spinal instrumentation planning

Physical medicine and rehabilitation: - Gait analysis and movement assessment - Rehabilitation robotics - Functional outcome prediction - Assistive technology optimization

Why orthopedics is well-suited for AI:

High-volume standardized imaging: Plain radiographs are uniform, DICOM-compliant, and abundant
Quantifiable outcomes: Joint alignment, fracture union, and functional scores are measurable
Procedural reproducibility: Joint replacement has standardized steps amenable to optimization
Objective functional data: Gait analysis, range of motion, and strength are quantifiable

Unique challenges:

Pediatric anatomy: Growth plates mimic fracture lines, requiring age-specific training data
Hardware artifacts: Orthopedic implants create imaging artifacts that confuse algorithms
Complex fractures: Comminuted fractures with multiple fragments challenge detection algorithms
Individual anatomy variation: Patient-specific bone morphology affects surgical planning accuracy

Part 1: Fracture Detection AI

The Clinical Problem

Fractures can be missed when findings are subtle, anatomy is complex, image quality is limited, or interpretation occurs under time pressure. The frequency depends on the population, modality, reference standard, reader expertise, and definition of an error, so a universal emergency-department miss rate is not defensible. Fracture-detection software targets a reader-support problem, not an autonomous replacement for the complete imaging and clinical assessment.

FDA Authorization and Product Boundaries

Imagen OsteoDetect (De Novo DEN180005, 2018):

Pathway and date: FDA granted De Novo authorization on May 24, 2018, creating a class II classification for radiological computer-assisted fracture detection and diagnosis software.
Intended use: The software helps clinicians identify and highlight distal-radius fractures while reviewing adult posteroanterior and lateral wrist radiographs. It is adjunctive and is not intended to replace clinician review.
Reader study: In a retrospective multiple-reader, multiple-case study, aided readers had an average area under the receiver operating characteristic curve of 0.889 versus 0.840 without assistance. Average sensitivity was 80.3% with assistance versus 74.7% without it, and average specificity was 91.4% versus 88.9% (FDA DEN180005).
Boundary: The FDA record does not authorize pediatric use, other bones, all wrist abnormalities, or autonomous diagnosis.

Aidoc and Zebra Medical Vision/Nanox:

These names are frequently used in market-level discussions of fracture and bone-health software. They should not be treated as one product, one indication, or one evidence base. A procurement review should record the exact trade name, version, regulatory number, anatomy, modality, workflow role, and intended user for every retained product. Evidence from OsteoDetect cannot be used to establish performance for another vendor or fracture type.

Performance Data

What the evidence can show:

Evidence unit	Defensible conclusion	Important limitation
FDA OsteoDetect reader study	Assistance improved reader discrimination and average sensitivity in adult wrist radiographs	Retrospective enriched reader study; narrow authorized anatomy and population
Prospective BoneView workflow study	Resident sensitivity increased after the AI result was displayed	Single-site diagnostic workflow study; no randomized patient-outcome comparison
Other fracture products and anatomies	Performance may be evaluated for a specific product, version, anatomy, and reference standard	No class-wide sensitivity range or transferable performance claim
Pediatric fracture detection	Requires age- and anatomy-specific evidence	Adult authorization and adult-reader evidence do not establish pediatric performance

Oppenheimer and colleagues prospectively integrated BoneView into a resident reporting workflow for 1,163 examinations in 735 patients, with 367 expert-reference fractures. Resident sensitivity was 84.74% before seeing the AI result and 91.28% after assistance; 35 reports were changed, 33 changes were correct, and 25 additional fractures were identified. AI alone had 86.92% sensitivity and lower specificity than either resident workflow. This study supports an aided-reader workflow claim, not autonomous use or improved patient outcomes (Oppenheimer et al., 2023).

Limitations and failure modes:

Pediatric growth plates and developmental anatomy:
- Unfused physes and ossification centers can resemble injury or obscure subtle findings
- The direction and magnitude of error are product-, age-, anatomy-, and study-specific
- Adult device evidence must not be extrapolated to children
Comminuted fractures:
- Multiple fragment patterns may confuse algorithms
- Performance should be measured separately for complex patterns rather than inferred from simple-fracture cohorts
Image quality dependence:
- Portable technique, positioning, motion, and exposure may differ from development data
- Local validation should stratify image quality and acquisition setting
- Overlying hardware or casts obscure anatomy
Anatomic coverage gaps:
- Many systems trained on specific anatomic regions (wrist, spine)
- Out-of-scope anatomy is not a legitimate extension of an authorized or validated claim

Clinical Implementation

Workflow integration:

1. Triage mode: - AI flags suspicious studies for priority review by radiologist - May reduce queue time when triage is the product’s authorized and locally validated role - Does not replace radiologist interpretation

2. Second reader mode: - AI provides concurrent read alongside radiologist - Highlights areas of concern for review - May reduce oversight errors

3. Alert mode: - Alerts can route specified findings under an institutionally approved escalation policy - Recipients, urgency, downtime procedures, and acknowledgment requirements must be explicit

Evidence for clinical benefit:

Reader-assistance studies can measure sensitivity, specificity, and report changes
Workflow studies can measure queue position, report turnaround, notification, and downstream imaging
Neither design establishes improved patient outcomes unless complications, treatment, function, or other patient-relevant endpoints are measured
Product performance should be rechecked after software, workflow, population, or acquisition changes

When fracture AI is most valuable:

Settings with a measured diagnostic or workflow problem that matches the authorized use
Overnight/weekend shifts with limited attending coverage
Urgent care centers without on-site radiologist interpretation
Pediatric settings only when the exact product has age-appropriate authorization or validation for the intended task

When it’s less useful:

Low-volume settings where all radiographs receive timely expert review
Subspecialized orthopedic practices with experienced surgeons
Settings where MRI or CT is routinely used for fracture diagnosis

Pediatric Fracture Detection: Special Considerations

Adult fracture evidence is not pediatric evidence. Pediatric evaluation should account for:

Growth plate physiology: Unfused physis appears as lucent line
Bone density differences: Lower mineralization in children
Fracture patterns: Buckle, greenstick, and plastic deformation fractures differ from adult patterns
Variant anatomy: Accessory ossicles, normal developmental variations

Recommendation: Confirm the exact regulatory and validation population before use. If the product’s evidence is adult-only, do not represent it as pediatric-validated. A pediatric deployment needs age- and anatomy-stratified performance, defined escalation, and prospective monitoring. OsteoDetect’s FDA indication is limited to adult wrist radiographs (FDA DEN180005).

Registry-based hip-fracture risk models such as FRACTURE-ML are not radiograph fracture detection: in a nationwide Swedish cohort (N ≈ 3.54 million), holdout 1-year AUC was 0.89 and Cox nearly matched DeepSurv, but prediction is not a care endpoint, and the study reported no external validation or implementation trial (Axelsson et al., 2026).

Part 2: Preoperative Planning and Surgical Navigation

AI in Joint Replacement Planning

Modern joint replacement planning software can integrate automated segmentation, statistical shape models, optimization, or machine learning. A digital plan, navigation system, or robotic arm should not be called AI unless the relevant function actually uses an AI method. Common functions include:

1. Implant size prediction: - Automated measurement of joint dimensions from CT or X-ray - An algorithm may propose an implant size for surgeon review - Agreement between planned and implanted size must be measured for the exact software, imaging protocol, implant system, and cohort

2. Component positioning: - Recommends angles, rotation, and depth - Patient-specific anatomic landmarks - Optimization for a defined alignment strategy or biomechanical target

3. 3D surgical planning: - Reconstruction from 2D radiographs or CT - Visualization of bone cuts and component placement - Simulation of postoperative alignment

Evidence base:

Planning tools can improve repeatability or agreement between planned and achieved component position in specific studies
Results may depend on imaging, registration, surgeon experience, implant system, and alignment philosophy
Technical accuracy must not be translated into pain, function, revision, or survival benefit without those endpoints

The accuracy-outcome gap:

Why do not more accurate plans lead to better outcomes?

Neutral mechanical alignment is not universally optimal: Some patients do better with kinematic or anatomic alignment
Soft tissue balancing also matters: Ligament tension and gap balance interact with component position
Patient factors matter: Comorbidity, anatomy, activity, pain mechanisms, rehabilitation, and expectations influence outcomes
Long-term studies needed: Implant survival at 10-15 years is the critical outcome

Robotic-Assisted Joint Replacement

Mako System (Stryker):

Mako is among the most studied robotic-arm platforms used for: - Total knee arthroplasty (TKA) - Total hip arthroplasty (THA) - Partial knee arthroplasty (PKA)

How it works:

Preoperative CT scan: Creates 3D model of patient anatomy
Surgical plan: Surgeon plans component size, position, and alignment
Intraoperative registration: Robot registers actual anatomy to preoperative plan
Haptic guidance: Robotic arm guides bone preparation with tactile feedback
Boundary constraints: Robot prevents cuts outside planned boundaries

The workflow combines preoperative imaging, a surgeon-defined plan, intraoperative registration, and constrained execution. Those functions should be described precisely rather than assuming that every component is AI.

Evidence from comparative studies and meta-analyses:

What Mako improves:

Outcome	Improvement	Evidence Level
Component alignment accuracy	Fewer alignment outliers in pooled studies	Supported for selected systems and alignment definitions
Planned vs. achieved position	More precise execution in selected studies	Technical endpoint; system- and procedure-specific
Ligament balancing	Can support quantified intraoperative planning	Protocol- and surgeon-dependent
Blood loss and length of stay	Results vary	Do not assume a class effect

What Mako does NOT clearly improve:

Outcome	Finding	Evidence Level
Patient-reported outcomes	No clinically meaningful superiority at final follow-up in a 2026 meta-analysis	Twenty comparative studies, including 13 RCTs
Implant survival	Long-term superiority not established	Follow-up and system generation remain important
Complication rates	No demonstrated class-wide reduction	Event definitions and power vary
Operating time	May increase during adoption and varies by workflow	Institution- and surgeon-specific

A 2026 systematic review and meta-analysis included 20 comparative studies involving approximately 5,800 patients, including 13 randomized trials. Pooled Knee Society Score, Oxford Knee Score, KOOS, WOMAC, Forgotten Joint Score, pain, and range-of-motion results did not demonstrate clinically meaningful superiority of robotic-assisted over conventional total knee arthroplasty at final follow-up; reported infection, revision, and reoperation rates were also similar (Al-Ghufaily et al., 2026). The evidence supports a precision claim more readily than a patient-benefit claim.

Cost considerations:

Capital, service, imaging, instrument, staffing, and opportunity costs vary by contract and platform
Per-case cost depends on disposables, throughput, scheduling, and whether equipment is shared across procedures
The learning curve should be measured locally using operative time, registration failures, conversions, and adverse events
Reimbursement and coverage must be verified for the payer, procedure, jurisdiction, and date

ROSA System (Zimmer Biomet):

Different manufacturer, workflow, imaging requirements, and evidence record
Results from Mako or another platform should not be transferred to ROSA
Procurement review should use system-specific studies and local implementation metrics

Virtual trials for implant-fixation evidence

Conventional cadaver, benchtop, and in-human studies cannot systematically sample the full range of bone quality, morphology, stem size, and alignment that affect cementless fixation. Virtual-trial and in silico implant-fixation work fills that gap between ROSA (an intraoperative robot) and spinal navigation: a different joint, a different endpoint, and a different evidence class. Maquer and colleagues described a virtual-trial framework that couples finite-element bone-implant models (the authors state these were validated against benchtop and clinical data) with a machine-learning layer, comparing two humeral stem designs (standard and anatomic) across more than 500 virtual surgeries on 45 humeri (21 male, 24 female; 27 left, 18 right) (Maquer et al., 2026). The virtual trial reconstructed micromotion as a primary-stability measure and a regional stress-shielding index (SSI), sampling short versus medium length, downsized, default, and oversized stems, and valgus, aligned, and varus alignment. This is complementary digital evidence for design comparison, planning discussion, and regulatory conversation, not a prospective patient-outcome trial. Virtual SSI and micromotion are not pain, function, loosening, or revision benefit. Six of the seven authors are Zimmer Biomet employees, and the seventh is a Zimmer Biomet consultant.

Spinal Navigation and Robotics

Applications:

Pedicle screw placement accuracy
Spinal deformity correction planning
Minimally invasive spine surgery guidance

Evidence:

Comparative studies frequently emphasize screw-placement accuracy, breach grade, revision, operative time, and radiation exposure
Radiation effects can differ for the patient, surgeon, and operating-room staff and depend on the comparator workflow
Learning varies with system, procedure, team, and surgeon experience
Radiographic screw accuracy is not itself evidence of improved pain, neurologic function, complications, or reoperation

Systems:

Mazor X (Medtronic)
ExcelsiusGPS (Globus Medical)
ROSA Spine (Zimmer Biomet)

These examples preserve the technology landscape, but any regulatory or comparative claim requires the exact model, software version, procedure, and source. Navigation and robotic guidance may use deterministic registration and constraint systems rather than machine learning.

Part 3: Rehabilitation Robotics and Gait Analysis

The Clinical Need in Physical Medicine

Physical medicine and rehabilitation face challenges in:

Subjective assessments: Manual muscle testing, visual gait observation have poor inter-rater reliability
Limited therapy intensity: Insurance constraints limit in-person PT sessions
Adherence monitoring: Home exercise programs poorly tracked
Outcome measurement: Patient-reported outcomes subject to placebo and recall bias

AI offers potential for: - Objective, quantitative functional assessment - High-intensity repetitive practice via robotics - Home-based monitoring and feedback - Personalized rehabilitation progression

Rehabilitation Robotics

Post-Stroke Upper Extremity:

InMotion ARM/HAND (Bionik Laboratories): - Robotic device for arm and hand rehabilitation - Adaptive difficulty based on patient performance - Used in stroke and neurological rehabilitation studies - Evidence: Small average improvements in impairment measures, but functional gains depend on therapy dose and comparator

Armeo Power (Hocoma): - Exoskeleton for upper extremity rehabilitation - 6 degrees of freedom, adaptive support - Gaming interface for engagement - Used in inpatient and outpatient settings

Evidence for upper extremity robotics:

The strongest reading is nuanced, not promotional. A 2018 Cochrane review of 45 trials and 1619 participants found electromechanical and robot-assisted arm training improved activities of daily living (SMD 0.31), arm function (SMD 0.32), and arm muscle strength (SMD 0.46), but the authors cautioned that trials varied by dose, device, participant characteristics, and outcome measures (Mehrholz et al., 2018). The pragmatic RATULS multicenter RCT randomized 770 participants and found that robot-assisted training did not improve upper-limb functional recovery compared with enhanced upper-limb therapy or usual care, although impairment scores improved versus usual care (Rodgers et al., 2019). The clinical takeaway is that upper-extremity robotics can add structured repetition but should not be presented as superior to equal-intensity functional therapy.

Lower Extremity and Gait Training:

Lokomat (Hocoma): - Robotic-assisted treadmill training - Body weight support + robotic leg orthoses - Widely used for stroke, SCI, cerebral palsy

Ekso GT (Ekso Bionics): - Wearable exoskeleton for gait training - Variable assistance levels - FDA 510(k) K143690, Class II powered lower extremity exoskeleton, indicated for ambulatory functions in rehabilitation institutions under trained physical therapist supervision for selected adults with stroke hemiplegia or SCI (FDA K143690)

Evidence for gait robotics:

A 2025 Cochrane update included 101 randomized studies and 4224 adults after stroke. Electromechanical or robotic gait devices plus physiotherapy probably helped more patients walk independently at the end of treatment, with about one additional independent walker for every nine treated, but probably did not improve average walking velocity or 6-minute walking distance compared with physiotherapy or usual care (Mehrholz et al., 2025). Benefit appears most plausible when the device increases total supervised walking practice, especially early after stroke. Cost-effectiveness remains uncertain because capital, maintenance, staffing, setup time, and throughput vary by institution.

Spinal Cord Injury Rehabilitation:

Robotics for SCI has mixed evidence. A 2025 meta-analysis of 15 randomized trials and 579 participants found robotic exoskeleton gait training improved walking balance, WISCI-II, lower-extremity motor scores, and FEV1 compared with conventional physical gait training, but did not improve 10-meter walking speed or 6-minute walking distance (Liu et al., 2025). Earlier safety-focused reviews support feasibility of powered exoskeleton walking in selected SCI patients, but sample sizes remain small and functional independence gains are not consistent (Miller et al., 2016).

Clinical bottom line: Rehabilitation robots can increase dose, repetition, and measurement precision. They should be evaluated as adjuncts to skilled rehabilitation, not as replacements for therapy intensity, task-specific practice, or clinician judgment.

AI-Enhanced Gait Analysis

Traditional gait analysis: - Requires dedicated motion capture laboratory - Multiple cameras, force plates, EMG sensors - Requires specialized equipment, setup, calibration, and interpretation - Limited to research and specialized centers

AI innovations:

1. Computer vision gait analysis: - Single video camera captures walking - AI extracts joint positions, angles, temporal parameters - Can be compared with laboratory motion capture for specific spatiotemporal or kinematic variables - Agreement must be reported variable by variable; one correlation range is not an accuracy claim

2. Wearable sensor integration: - IMUs (inertial measurement units) on key body segments - Machine learning classifies gait patterns - Real-time feedback for gait retraining - Home and community monitoring

3. Smartphone-based assessment: - Accelerometer and gyroscope data during walking tests - AI predicts fall risk, gait speed, stride variability - Scalable to large populations - Limited accuracy compared to dedicated systems

Clinical applications:

Use Case	Technology	Evidence	Limitations
Parkinson’s gait monitoring	Wearable sensors	Condition- and device-specific studies	Does not replace clinical assessment
Postoperative TKA gait	Computer vision	Emerging validation	Clinical action thresholds are not standardized
Fall-risk estimation	Smartphone or wearable sensors	Heterogeneous evidence	Predictive values depend on population and intervention
Stroke gait asymmetry	Wearable sensors	Measurement and rehabilitation studies	Requires interpretation in the patient’s functional context

Stenum and colleagues compared video-based pose estimation with three-dimensional motion capture and found that agreement varied across gait parameters rather than supporting one global accuracy percentage (Stenum et al., 2021). A measurement tool becomes clinically useful only when its error, repeatability, reference range, and decision consequence are understood.

Barriers to adoption:

Reimbursement: Limited CPT codes for AI-based gait analysis
Workflow integration: Separate assessment outside usual PT workflow
Interpretation training: PTs need training to use quantitative gait data
Validation: Many systems lack large-scale clinical validation

Patient-Facing AI for Post-Operative Recovery

RecovryAI Virtual Care Assistants (VCAs):

In March 2026, RecovryAI announced that FDA had granted Breakthrough Device Designation to a physician-prescribed, patient-facing postoperative support product (RecovryAI press release, March 2026; Palmer, STAT News, March 2026). A public FDA device authorization record was not identified for this chapter update.

The company describes virtual assistants that provide recovery information, compare patient-reported symptoms with clinical protocols, and escalate selected concerns
Statements about a Class II pathway, multicenter study, or pilot performance remain company-reported unless supported by a public FDA record or peer-reviewed publication
Breakthrough Device Designation is not marketing authorization and does not establish safety, effectiveness, or clinical benefit

Clinical relevance for orthopedic teams: Post-discharge recovery after total joint arthroplasty involves patient education, symptom monitoring, rehabilitation questions, and care coordination. A patient-facing system therefore needs defined content governance, urgent escalation, audit logs, accessibility testing, clinician response coverage, and a safe downtime pathway.

Randomized evidence with important limitations: Li and colleagues randomized 300 adults after sports-medicine or joint-replacement procedures to a GPT-4-based WeChat agent or physician messaging. The final analysis included 140 of 150 AI participants and 121 of 150 physician-messaging participants. The AI responded faster but had lower response accuracy (93.9% versus 98.1%) and a 6.3% audited hallucination rate. Functional, physical-health, and satisfaction differences favored the AI group at 1 and 3 months but were not significant at 6 months. Enrollment occurred from December 2023 through June 2024, while trial registration was recorded on April 23, 2025, after enrollment; attrition was also differential (Li et al., 2026). This trial supports cautious study of supervised postoperative communication, not replacement of clinicians or a durable outcome benefit.

Consent and patient education evidence: A 2025 single-blind pilot RCT randomized 60 total knee arthroplasty patients to ChatGPT-assisted versus traditional surgeon-led informed consent. Among 55 completers, the ChatGPT-assisted group had lower post-consent anxiety scores and higher satisfaction, with no significant differences in depression, pain, or knee function (Gan et al., 2025). The study is small and physician-mediated, but it supports a narrow use case: standardized answers to common perioperative questions under surgeon supervision.

AI Clinical Decision Support for Perioperative Hyponatremia

Hyponatremia (serum sodium <135 mmol/L) is a recognized complication after total joint arthroplasty and has been associated with prolonged length of stay and other postoperative complications. Kunze and colleagues at Hospital for Special Surgery published an AI-based clinical decision support tool aimed at reducing post-arthroplasty hyponatremia in NEJM Catalyst Innovations in Care Delivery (Kunze et al., NEJM Catalyst, 2026).

This report illustrates an orthopedic use case beyond imaging: linking perioperative risk estimates to a defined care pathway. It should not be treated as class-wide evidence for AI electrolyte management.

Limitations: Independent validation outside the originating institution has not been published, and performance metrics from this single-center report should not be assumed to generalize. Readers should consult the primary publication for study design, cohort characteristics, and outcome measures.

Part 4: Sports Medicine and Injury Prevention

Injury Risk Prediction

Theoretical applications:

ACL injury risk from movement screening
Stress fracture prediction from training load
Shoulder injury risk in overhead athletes
Return-to-play decision support

Reality check:

Most sports medicine AI is in research phase:

ACL injury prediction: - Multiple studies using video analysis + ML to predict injury risk - Reported discrimination and classification performance vary by cohort, feature set, outcome definition, and validation design - A model is not a screening program: Predictive values, preventable risk, intervention capacity, and harms from restriction or false reassurance determine clinical value - No evidence that prediction leads to effective prevention

Training load monitoring: - Wearable sensors track volume, intensity, recovery - Machine learning identifies injury risk patterns - Used by professional teams, limited evidence for injury reduction - Confounding factors: Genetics, nutrition, sleep, psychosocial stress

Current status:

Useful research tool for understanding injury mechanisms
Not yet validated for individual athlete screening
May have role in population-level injury surveillance

Return-to-Play Decision Support

Objective functional testing:

AI can enhance return-to-play assessment by: - Comparing injured to uninjured limb symmetry - Tracking recovery trajectory vs. population norms - Integrating multiple test results (strength, ROM, balance, sport-specific tasks)

Evidence:

Can standardize data display and longitudinal comparison
A reduction in reinjury cannot be claimed without a comparative intervention study that measures reinjury
Does not replace clinical judgment about psychological readiness, sport demands

Part 5: Implementation Challenges in Orthopedic AI

Integration with Radiology Workflow

Who interprets fracture AI results?

Confusion about roles creates implementation barriers:

Emergency department model: - Radiologist interprets radiograph, orthopedist treats fracture - AI alerts both radiologist and ED physician - Risk of alert fatigue if both groups receive redundant notifications

Urgent care model: - No radiologist on-site, images read remotely - AI provides provisional read pending radiologist review - Orthopedist or ED doc makes initial treatment decision

Liability considerations:

If AI flags fracture but radiologist reads as normal, who is responsible?
If AI misses fracture (false negative), is this a system failure or radiologist oversight?
Responsibility depends on the facts, jurisdiction, roles, institutional policy, product labeling, and applicable standard of care

Best practices:

Use the product only for its authorized and locally approved role
Do not infer a fracture diagnosis from an alert without the required clinical and imaging review
Define where alerts and clinician responses are documented
Define discordance escalation instead of relying on informal responsibility assumptions

Robotic Surgery Learning Curve

Adoption challenges:

Capital and operating cost barrier: - Acquisition or lease terms differ by platform and contract - Maintenance, imaging, disposables, staffing, training, and operating-room time must all be included - Opportunity costs and shared use across service lines affect the economic model - Volume alone is not sufficient; the analysis must connect measured benefit to the institution’s actual bottleneck

Learning curve: - Measure setup time, registration failures, procedure time, conversions, technical errors, and clinical outcomes case by case - Stratify by surgeon, procedure, operating-room team, and platform - Do not use a universal case count as proof of proficiency

Surgeon resistance: - Experienced surgeons may question whether the measured benefit justifies workflow and cost - Concern about de-skilling: Reliance on robot reduces manual surgical skills - Marketing vs. evidence: Patients request “robot surgery” without understanding benefits

Evidence-based approach:

Review published RCT data for your specific procedure
If considering adoption, plan realistic volume to justify cost
Invest in structured training (manufacturer courses, proctored cases)
Track outcomes prospectively to validate benefit in your practice
Informed consent should include lack of proven outcome superiority

Data Privacy in Rehabilitation Monitoring

Home-based rehabilitation robotics and monitoring:

Wearable sensors collect continuous movement data
Video-based gait analysis records patient in home environment
Data transmitted to cloud platforms for AI processing

Privacy concerns:

HIPAA compliance: Is PHI adequately protected?
Video data: Home videos may capture family members, living conditions
Third-party access: Device manufacturers, AI vendors may access data
Data ownership: Who owns gait data, movement data?

Best practices:

Verify the applicable privacy and security obligations, contract terms, and business-associate arrangements rather than accepting a generic “HIPAA-compliant” label
Obtain explicit consent for video recording
Minimize data collection to clinically necessary information
Ensure data deletion policies align with medical record retention requirements

Part 6: Professional Society Guidelines and Position Statements

American Academy of Orthopaedic Surgeons (AAOS)

AAOS adopted Position Statement 1193, “Artificial Intelligence,” in February 2025. AAOS explicitly describes it as an educational tool based on the authors’ opinion and not the product of a systematic review. It is an official position statement, not an AI clinical practice guideline (AAOS Position Statement 1193).

AAOS Position on AI (Educational Focus)

Principles in the AAOS statement:

Transparency and explainability: Developers and users should understand system purpose, data sources, limitations, and the basis of outputs.
Safety, privacy, and security: AI use should protect patients and data across development and deployment.
Fairness and equity: Evaluation should identify bias and avoid worsening disparities.
Human responsibility: AI should support, not displace, professional judgment and accountability.
Patient and community engagement: Development and implementation should include affected users and populations.

What AAOS has NOT published: - An evidence-graded clinical practice guideline for fracture-detection AI - Product-specific credentialing standards for robotic or AI-assisted procedures - Universal performance or outcome thresholds for orthopedic AI

Why guidelines are limited: - Evidence base still evolving - Heterogeneity of AI applications across orthopedic subspecialties - Lack of consensus on outcome metrics for AI benefit

Rehabilitation and Arthroplasty Society Context

The following material is retained as a practical professional framework. It should not be attributed to AAPM&R, APTA, or AAHKS as formal AI guidance without a direct society statement or guideline.

Rehabilitation medicine focus:

Key focus areas:

Rehabilitation robotics integration: Education on appropriate patient selection
Objective outcome measurement: AI for quantifying functional gains
Telerehabilitation: AI-enhanced remote monitoring and coaching
Assistive technology: Machine learning for device optimization (prosthetics, orthotics)

Position themes:

AI should augment clinician assessment, not replace it
Patient-centered outcomes (independence, quality of life) matter more than isolated metrics (gait speed)
Equity considerations: Expensive robotic therapy may exacerbate healthcare disparities
Interdisciplinary collaboration: AI tools require PT, OT, physician, and engineer input

Physical therapy focus:

Core principles:

Clinical decision-making authority: PT retains responsibility for all clinical decisions
Evidence-based adoption: AI tools should have published validation for intended use
Patient autonomy: Patients should be informed when AI is used in their care
Professional development: PTs need training to interpret AI outputs

Specific AI applications addressed:

Wearable sensor data interpretation
Telehealth AI assistants
Gait analysis algorithms
Rehabilitation robotics dosing

Hip and knee arthroplasty focus:

Guidance themes:

Technology is a tool: Robotic systems do not replace surgical skill and judgment
Evidence-based adoption: Review RCT data, not just marketing materials
Cost-effectiveness: Institutions should perform cost-benefit analysis
Surgeon training: Manufacturer training is necessary but not sufficient; ongoing education essential

Chapter Appraisal Principles for Orthopedic AI

The AAOS statement and the chapter’s evidence-appraisal framework support the following implementation questions. These are not represented as a multisociety consensus statement:

Patient Safety: - Verify whether the product is a regulated medical device and, if so, its exact authorization and labeling - Local validation recommended before clinical deployment - Maintain human oversight for all clinical decisions

Evidence Requirements: - Peer-reviewed publication of performance data - External validation beyond development site - Outcomes beyond radiographic accuracy (functional outcomes, patient satisfaction)

Professional Responsibility: - Professional and organizational responsibilities remain fact- and jurisdiction-specific - Document AI assistance in clinical notes - Maintain competence in non-AI techniques (do not become dependent)

Equity and Access: - AI should not exacerbate existing healthcare disparities - Cost-effectiveness considerations important given resource constraints - Pierson and colleagues developed a model trained to predict patient-reported pain from knee radiographs; in the study population, its predictions explained more of the racial disparity in pain than conventional radiographic severity measures. This supports investigation of measurement bias, not a conclusion that deployed AI has reduced disparities (Pierson et al., 2021)

Transparency: - Patients should be informed when AI is used in their diagnosis or treatment - Vendor algorithm transparency desirable but often limited by proprietary concerns

Check Your Understanding

The four cases below are explicitly hypothetical teaching exercises. Their patients, institutions, products, contracts, costs, workflows, and outcomes are not reports of real events. Any numerical assumption is labeled as an input to be replaced with local measured data.

Clinical Scenario 1: The Missed Pediatric Fracture

Hypothetical clinical situation: An 8-year-old falls from playground equipment and presents to urgent care with wrist pain. A fictional fracture-alert system flags a possible distal-radius fracture. The system’s local deployment record does not include pediatric validation. A radiologist interprets the films as showing an unfused distal radial growth plate with no fracture. The urgent care physician is unsure how to proceed.

Question: How should the clinician reconcile the discrepancy between AI alert and radiologist interpretation?

Answer: Do not let an out-of-scope alert substitute for pediatric imaging expertise, clinical assessment, or an established discordance pathway.

Reasoning:

Pediatric evidence boundary: An adult-authorized or adult-validated product cannot be presumed valid in a child. Unfused physes and developmental variants require age-appropriate evidence.
Imaging expertise: A pediatric-trained or experienced radiologist interprets the complete examination in its anatomic and clinical context. If the finding and examination remain discordant, the institution’s escalation or second-review process should apply.
Clinical correlation: Physical exam matters. Is there focal bony tenderness over the fracture site vs. physis? Is there deformity? Can the child bear weight (for lower extremity) or use the hand (for wrist)?
When in doubt: If high clinical suspicion despite negative radiologist read:
- Consider splinting and orthopedic follow-up in 5-7 days
- Repeat radiographs at follow-up may show subtle fracture line or periosteal reaction
- MRI not typically needed for simple wrist injuries

What to document: - “Radiographs obtained. AI alert for possible fracture. Radiology interpretation: No fracture, normal growth plate. Physical exam shows tenderness over distal radius but no deformity. Discussed with family: placed in volar splint for comfort, orthopedic follow-up arranged in 1 week. If pain worsens or new symptoms develop, return for re-evaluation.”

Lesson: An out-of-scope pediatric alert is not confirmatory evidence. Clinical assessment, radiologist review, appropriate immobilization or follow-up, and safety-net instructions should follow the actual presentation and local protocol.

Clinical Scenario 2: Counseling a Patient About Robotic Knee Replacement

Hypothetical clinical situation: A 68-year-old with end-stage osteoarthritis is considering total knee arthroplasty. She has researched “robot surgery” online and asks whether the institution offers Mako robotic-assisted TKA. She believes robotic surgery guarantees less pain, faster recovery, and longer implant survival.

Question: How do you counsel this patient about robotic vs. conventional TKA?

Answer: Explain the difference between more precise execution and proven patient benefit, using evidence for the exact platform and procedure.

What to say:

“Robotic-assisted knee replacement is available. The system helps the surgeon plan and execute bone preparation within defined boundaries. The evidence should be divided into what is measured technically and what patients experience:

What robotic surgery improves: - Selected studies show more precise component positioning and fewer alignment outliers. - The technology can provide a repeatable planning and execution workflow. - Results depend on the platform, alignment strategy, registration, surgeon, and procedure.

What’s the same between robotic and conventional: - Pooled comparative evidence has not shown clinically meaningful superiority in patient-reported pain, function, or satisfaction at final follow-up. - Long-term implant-survival superiority is not established. - Infection, revision, and reoperation should not be represented as improved unless the exact evidence supports those endpoints. - Recovery and rehabilitation remain individualized.

Important points: - The robot is a tool used under surgeon control; it does not independently perform the operation. - Whether I use the robot or conventional technique, the goal is the same: a well-aligned, balanced, durable knee replacement. - Success depends more on patient factors (your health, bone quality, commitment to rehab) than on robotic vs. conventional technique.

My recommendation: The choice should reflect the patient’s anatomy and goals, the surgeon’s experience with each technique, the exact system, and institutional results. Technology should not be offered as a guarantee.”

What NOT to say: - “Robotic is better” (not supported by outcome data) - “The robot does the surgery” (surgeon still controls all decisions) - “You’ll recover faster” (not established as a class-wide promise)

Lesson: Patients often overestimate the benefits of surgical technology based on marketing. Evidence-based counseling builds trust and manages expectations appropriately.

Clinical Scenario 3: Implementing Fracture Detection AI in Your ED

Hypothetical clinical situation: An emergency medicine director is considering a fictional fracture-detection product. The emergency department has a defined radiography volume and radiology coverage model. The vendor has proposed a yearly contract, but the institution has not yet measured its baseline diagnostic discrepancy rate, alert burden, or downstream cost.

Question: How do you evaluate whether fracture detection AI is worth implementing at your institution?

Answer: Conduct a systematic cost-benefit analysis and assess workflow integration.

Key questions:

1. What is the current diagnostic and workflow baseline? - Review a prespecified sample of reports, addenda, callbacks, delayed diagnoses, safety events, and patient complaints - If missed fracture rate is already very low, marginal benefit is small - Separate fractures detectable on the product’s supported anatomy and views from out-of-scope cases

2. What is the current radiology workflow? - If all radiographs receive real-time attending radiologist interpretation, benefit is limited to “second reader” - If residents or mid-levels provide preliminary reads, AI can reduce errors - If no radiologist coverage overnight, AI provides valuable triage

3. What is the alert workflow? - How will ED physicians be notified of AI alerts? - Will alerts go to radiologists, ED docs, or both? - What is the expected false positive rate and resulting alert burden?

4. Build a transparent local sensitivity analysis:

Input	Local source	Low/base/high range
Eligible examinations	PACS and product labeling	Measured, not total radiograph volume
Actionable baseline discrepancies	Blinded local audit	Include uncertainty
Incremental detections with assistance	Silent prospective evaluation	Do not substitute standalone sensitivity
False and duplicate alerts	Silent prospective evaluation	Translate into staff time and downstream tests
Patient harm avoided	Clinical review	Do not equate every detection with prevented harm
Contract and implementation cost	Executed proposal plus internal labor	Include interfaces, support, monitoring, and downtime

The former example’s fixed settlement value, assumed sensitivity, and favorable net-benefit conclusion were invented teaching inputs, not measured evidence. A defensible business case replaces every input with local data and shows how the conclusion changes across uncertainty ranges.

Decision rule: Proceed only if the exact product addresses a measured local problem, the silent evaluation meets prespecified safety and workflow criteria, and the benefit remains plausible across the sensitivity analysis. A favorable vendor estimate alone is insufficient.

Pilot protocol: 1. Conduct a time-bounded silent evaluation under institutional approval, with no clinical reliance on unvalidated output 2. Measure: Sensitivity, specificity, false positive rate, time to flagging 3. Compare performance with prespecified local criteria, not only vendor claims 4. Measure: ED physician satisfaction, radiology workflow impact, missed fracture rate 5. Continue monitoring after updates, workflow changes, and population shifts

Lesson: AI implementation requires systematic assessment of local needs, workflows, costs, and safety. A silent phase can characterize performance without directing care, but it still requires governance, privacy review, and a plan for unexpected safety findings.

Clinical Scenario 4: Rehabilitation Robotics for Stroke Recovery

Hypothetical clinical situation: A physiatrist is treating a 62-year-old man 4 weeks after ischemic stroke with left hemiparesis. His coverage permits a finite number of therapy sessions, and the facility has an Armeo Power upper-extremity robotic system. The family asks whether “robot therapy” is better than conventional therapy.

Question: How do you approach the decision to use rehabilitation robotics vs. conventional therapy?

Answer: Robotics can be one component of a comprehensive program but is not superior to equal-intensity conventional therapy.

Assessment considerations:

1. Patient factors favoring robotics: - Motivated, cognitively intact (can follow gaming interface instructions) - Sufficient active participation or device-assisted range for the selected protocol - Recovery phase and goals that match the evidence and device program - Tolerates intensive repetitive exercise

2. Patient factors against robotics: - Impairment or range that the selected device cannot safely accommodate - Goals better served by task-specific functional practice - Cognitive impairment: Cannot engage with gaming interface - Shoulder pain: Robotic movements may exacerbate

3. Resource considerations: - Is conventional therapy slot available at same time, or does robotic access allow more therapy time? - Therapist supervision required: Not autonomous therapy - Session duration and supervision should follow the patient, device protocol, and treatment plan

Your recommendation:

“The Armeo robotic system can be helpful as part of your therapy program. Research shows it’s about as effective as the same amount of conventional therapy, not better. The benefit is that it allows high-intensity repetitive practice with adaptive difficulty.

My recommendation: Use the robot only as a planned component of a comprehensive program that also includes: - Conventional occupational therapy for functional tasks (dressing, eating, grooming) - Task-specific training (reaching, grasping real objects) - Home exercise program

The robot is a tool, not a magic bullet. Your recovery depends more on total therapy intensity, your effort, and neuroplasticity than on whether we use robotics vs. conventional techniques.

Let’s plan to reassess in 4 weeks. If you’re progressing well, we’ll continue. If plateau, we’ll adjust the program.”

What NOT to say: - “Robotics is the most advanced treatment” (implies superiority not supported by evidence) - “The robot will help you recover faster” (no evidence for accelerated recovery) - “This is better than what therapists can do” (undervalues skilled therapist intervention)

Lesson: Rehabilitation robotics is one tool in a comprehensive program. Benefits are from intensity and repetition, not from robotics per se. Patient-centered goal-setting and therapist expertise remain central to rehabilitation success.

Key Takeaways

Clinical Bottom Line for Orthopedic and PM&R AI

Fracture-detection evidence is product- and indication-specific. OsteoDetect has a narrow adult-wrist De Novo authorization and retrospective reader study. It does not establish a pediatric or class-wide claim.
Robotic joint replacement can improve selected technical measures. A 2026 meta-analysis did not show clinically meaningful superiority in pooled patient-reported outcomes at final follow-up.
Preoperative planning software may improve repeatability. Confirm which function uses AI, and do not transfer one system’s evidence to another.
Rehabilitation robotics can increase structured repetition. Upper-limb and gait evidence depends on dose, comparator, patient selection, and endpoint. Robotics should remain part of skilled rehabilitation.
Gait analysis is parameter-specific. Agreement for one variable cannot be converted into a global accuracy percentage or a clinical-action threshold.
Sports medicine prediction remains research-stage. Prediction alone does not show that an intervention prevents injury or safely changes return-to-play decisions.
AAOS has an official AI position statement, not an AI clinical practice guideline. Do not attribute chapter-created principles to AAPM&R, APTA, or AAHKS as formal guidance.
Pediatric applications require separate evidence. Adult authorization cannot be extrapolated across age, anatomy, acquisition, or fracture pattern.
Implementation requires workflow and safety integration. Define recipients, documentation, discordance, escalation, downtime, monitoring, and update review.
Economic conclusions require local measured inputs. Contract cost, eligible volume, incremental action, alert burden, staffing, downstream testing, and patient-relevant benefit belong in a sensitivity analysis.

Questions for Orthopedic AI Adoption

What does the FDA authorization for OsteoDetect cover?

FDA authorized OsteoDetect through the De Novo pathway in 2018 as an adjunct for clinicians reviewing adult posteroanterior and lateral wrist radiographs for distal-radius fractures. That authorization does not establish performance for children, other anatomic sites, or fracture AI as a class (FDA DEN180005).

Does robotic surgery improve joint replacement outcomes?

Robotic and navigation systems can improve the precision of component positioning in selected studies, but radiographic precision is not the same endpoint as pain, function, complications, revision, or implant survival. The specific system, procedure, comparator, follow-up, and patient-relevant endpoint determine the claim.

What are the limitations of pediatric fracture AI?

Adult fracture-device authorization and validation cannot be transferred to children. Pediatric anatomy, developmental variants, and fracture patterns require age-appropriate evidence for the intended anatomy and workflow.

Does orthopedic AI improve patient outcomes?

Some studies show improved reader sensitivity, response time, alignment precision, or short-term patient experience. These are not interchangeable with reduced complications, improved function, fewer revisions, or durable clinical benefit. Outcome claims require comparative studies that measure those endpoints.