This chapter introduces the core concepts that allow physiotherapists to use artificial intelligence (AI) safely in musculoskeletal assessment, gait analysis, clinical decision support and education, with emphasis on model generalizability and the clinician's role in interpreting algorithmic output.
Foundations: What AI Actually Does
AI in physiotherapy is mostly machine learning (ML), in which algorithms learn statistical patterns from data instead of following hand-written rules. Supervised learning uses labeled data (for example, videos tagged "excessive hip adduction" or not) to predict a known outcome. Unsupervised learning finds clusters in unlabeled data, such as subgroups of low back pain patients. Deep learning uses multilayer neural networks and underlies most pose-estimation and image-analysis tools. Natural language processing (NLP) extracts structured information from clinical notes.
Every model produces a probability or classification from the inputs it was trained on. It does not reason about the individual patient. This principle underlies the rest of the chapter.
The Data Pipeline: Why Inputs Define Outputs
A clinical AI tool has four stages: data capture, preprocessing, model inference and output presentation. Errors at any stage propagate to the recommendation.
- Capture: smartphone camera, inertial measurement units (IMUs), force plates, optical motion capture or electronic records.
- Preprocessing: filtering, normalization, landmark detection.
- Inference: classification (e.g., "pathological gait pattern") or regression (e.g., predicted 12-month disability).
- Output: a score, heat map or recommendation, which is a decision-support aid, not a diagnosis.
Marker-less pose estimation (e.g., DeepLabCut, OpenPose) allows kinematic analysis from ordinary video. Accuracy depends on camera position, lighting, clothing and the movement plane captured.
AI in Musculoskeletal Assessment
Consider a 45-year-old with 3–6 months of non-traumatic shoulder pain, painful overhead activity and a positive Neer test. An AI camera system can quantify scapulohumeral rhythm, peak elevation and movement variability during flexion and abduction. Normal active elevation is approximately 160–180°, with a scapulohumeral ratio of roughly 2:1 for humeral to scapular motion over the full range, and the scapula contributes about 60° of upward rotation.
The system's output must be integrated with the clinical examination:
- The Neer test has moderate sensitivity (about 70–80%) but poor specificity (about 50–60%), so a positive result is not diagnostic. It should be interpreted as part of a cluster that includes the Hawkins-Kennedy test, the painful arc, and external rotation weakness (Hegedus et al., 2012).
- A painful arc around 60–120° abduction, with full passive range of motion, points towards subacromial pain rather than adhesive capsulitis, which shows restricted passive external rotation, often below 50% of the contralateral side.
- Red flags requiring referral include night pain unrelated to position, unexplained weight loss, a history of malignancy, fever, a hot swollen joint, neurological deficit, and cardiac or visceral referral features.
An AI motion score that conflicts with the clinical picture should prompt re-examination, not automatic acceptance.
AI-Assisted Gait Analysis
Gait analysis is a leading application of AI because gait is rich in quantifiable data. Typical healthy adult reference values include:
- Comfortable walking speed: approximately 1.2–1.4 m/s (below 1.0 m/s suggests functional limitation; below 0.8 m/s is associated with increased fall and hospitalization risk).
- Cadence: approximately 100–120 steps/min.
- Stance phase: about 60% of the gait cycle; double support about 20%.
- Knee flexion at loading response: about 15–20°; peak swing about 60°.
- Ankle dorsiflexion at terminal stance: about 10°.
Clinical example (chronic low back pain). A system analyzes a 2D video of a 10-meter walk test. A single sagittal-plane camera captures stride length, cadence, trunk flexion and sagittal hip/knee angles reasonably well. It captures frontal-plane events poorly, including pelvic drop (Trendelenburg), hip adduction and trunk lateral lean, and it has difficulty with transverse-plane rotation. Parallax and perspective error also distort angles when the patient is not perpendicular to the lens. A flagged "asymmetry" should therefore be verified by observation, a 10-meter walk test timed with a stopwatch, and clinical tests such as single-leg stance (healthy adults under 50 should hold at least 30 s) and hip abductor strength (manual muscle testing, graded 0–5, or handheld dynamometry).
Clinical example (patellofemoral pain). A 40-year-old runner is assessed by a model trained on similar patients. Associated running mechanics include excessive hip adduction and internal rotation during stance, contralateral pelvic drop and increased peak knee abduction moment. A recommendation for hip strengthening is plausible, but it should be confirmed with the clinician's own findings, such as hip abductor and external rotator strength deficits, a single-leg squat assessment and pain on stairs.
AI Model Generalizability
Generalizability is the ability of a model to perform accurately on new patients, sites and equipment that differ from its training data. Failure usually arises from:
- Population mismatch: a model trained on young male recreational runners may misclassify older women, adolescents or people with obesity or neurological conditions.
- Dataset shift: differences in camera, frame rate, treadmill versus overground running, footwear or clinic environment.
- Label bias: training labels derived from one clinician's subjective ratings.
- Overfitting: excellent performance on development data but poor performance externally.
Before relying on a tool, ask the following:
- Was it externally validated in a population resembling this patient?
- What were the sensitivity, specificity, area under the curve and calibration?
- Are the demographic and clinical characteristics of the training set reported (TRIPOD+AI reporting standard)?
- Does the output include an uncertainty estimate?
- Has it been shown to change patient outcomes, rather than just agreement with experts?
Technical accuracy does not equal clinical effectiveness.
AI in Clinical Decision Support
Clinical decision support (CDS) systems combine patient data with algorithms to suggest risk estimates, classifications or management options. Consider a 55-year-old with chronic low back pain (8 months, no red flags, NPRS 6/10, Oswestry Disability Index 40%). Moderate disability on the ODI is 21–40%, and 41–60% is severe. A prognostic model could stratify risk using a tool such as the STarT Back Screening Tool (low, medium or high risk of persistent disability).
Interpretation requires knowing the minimal clinically important change (MCIC). For chronic low back pain, this is approximately 2 points (or 30%) on the NPRS and 10 points (or 30%) on the ODI (Ostelo et al., 2008). If AI predicts a good prognosis but the patient scores high on psychosocial yellow flags (fear-avoidance, catastrophizing, low self-efficacy), the clinician should weigh the person's context. Guideline-concordant care for non-specific chronic low back pain includes education, graded activity or exercise, and manual therapy as an adjunct, with imaging avoided without red flags (Foster et al., 2018).
Key safeguards include automation bias (over-trusting a recommendation), algorithmic opacity (not knowing why a recommendation was made) and accountability, because the treating clinician remains legally and ethically responsible.
Remote and Smartphone-Based Monitoring
A 30-year-old athlete two weeks after a grade II lateral ankle sprain has limited dorsiflexion and begins rehabilitation. Smartphone inclinometer or AI-based angle apps can monitor active range of motion daily.
- Normal ankle dorsiflexion is about 20° (non-weight-bearing); the weight-bearing lunge test typically yields 35–45° tibial inclination, or roughly 10–12 cm toe-to-wall distance, with side-to-side differences above 1.5–2 cm or 5° considered clinically relevant.
- Validated smartphone tilt-meter methods show good reliability for the weight-bearing lunge (Williams et al., 2013), but measurement error is typically 2–4°. Use the same method, time of day, shoe condition and surface, and track trends.
- Protocol example: PRICE/POLICE principles for days 0–3, early weight-bearing as tolerated, and range-of-motion exercise 3 × 10 repetitions, 2–3 times daily, progressing to resisted band work (3 × 15), balance training (single-leg stance 30–60 s, progressing to unstable surfaces) and sport-specific loading.
- Escalate if there is persistent swelling, inability to bear weight for four steps, bony tenderness (Ottawa ankle rules), or signs of syndesmotic injury.
AI in Physiotherapy Education
Adaptive learning platforms adjust content based on learner performance. A student studying gait analysis who answers correctly may be advanced automatically. This supports mastery-based progression, but risks include false mastery when the questions test recall rather than application, narrow content coverage, and a lack of educator oversight. Educators should confirm that adaptive thresholds are explicit, that content includes clinical reasoning, and that the platform provides transparent analytics.
Adaptive platforms can advance a student who answers correctly, but questions that test recall rather than application risk false mastery.
Other educational applications include virtual patient simulators using NLP and branching logic, and automated OSCE video scoring, which can improve consistency but must be checked for bias and validity against expert examiners.
Curriculum Analysis with AI
An instructor reviewing three years of student performance data may find that cohorts consistently struggle with proprioceptive and neuromuscular concepts. AI can reveal patterns across assessment items, but correlation is not causation. Before changing teaching, the instructor should check item quality, assessment alignment with learning outcomes, prerequisite knowledge (neuroanatomy, motor control), cohort differences and sample size. The appropriate response is a targeted intervention, such as more case-based labs on joint position sense testing, balance training and motor learning, followed by re-evaluation of outcomes. Student data must be handled under privacy and consent rules.
AI can reveal patterns in cohort performance, but correlation is not causation: check item quality, alignment and cohort differences before changing teaching.
Ethical, Legal and Practical Considerations
- Privacy: video of patients is identifiable health data and needs consent, secure storage and regulatory compliance (e.g., HIPAA, GDPR).
- Equity: underrepresented groups in training data face higher error rates.
- Transparency: patients should be told when AI contributes to decisions.
- Contraindications to relying on AI alone: acute red flags, atypical presentations, people with amputations or neurological conditions that fall outside training data, and poor-quality data.
Summary and Clinical Pearls
- AI outputs are probabilistic aids, and the clinician integrates them with history, examination, patient preference and evidence.
- Sagittal 2D video is limited for frontal and transverse-plane gait deviations.
- Always ask whether a model has been externally validated in a population like your patient.
- Use validated outcome measures (NPRS, ODI, weight-bearing lunge, 10-meter walk test) with their MCIC values to interpret change.
- Be alert to automation bias and dataset shift.
- In education, AI should support mastery and curriculum review, with educator judgment retained.
References
- Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med. 2019;25(1):44-56.
- Collins GS, Moons KGM, Dhiman P, et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 2024;385:e078378.
- Mathis A, Mamidanna P, Cury KM, et al. DeepLabCut: markerless pose estimation of user-defined body parts with deep learning. Nat Neurosci. 2018;21(9):1281-1289.
- Kidziński Ł, Yang B, Hicks JL, et al. Deep neural networks enable quantitative movement analysis using single-camera videos. Nat Commun. 2020;11:4054.
- Hegedus EJ, Goode AP, Cook CE, et al. Which physical examination tests provide clinicians with the most value when examining the shoulder? Br J Sports Med. 2012;46(14):964-978.
- Ostelo RWJG, Deyo RA, Stratford P, et al. Interpreting change scores for pain and functional status in low back pain. Spine. 2008;33(1):90-94.
- Foster NE, Anema JR, Cherkin D, et al. Prevention and treatment of low back pain: evidence, challenges, and promising directions. Lancet. 2018;391(10137):2368-2383.
- Williams CM, Caserta AJ, Haines TP. The TiltMeter app is a novel and accurate measurement tool for the weight bearing lunge test. J Sci Med Sport. 2013;16(5):392-395.