This chapter introduces the core concepts of artificial intelligence (AI) in physiotherapy research: how models learn from clinical data, where they are applied in assessment, diagnosis, monitoring and prediction, and how to judge whether their outputs are valid, safe and ethical.
Foundations: AI, Machine Learning and Deep Learning
AI is the broad field of building systems that perform tasks normally requiring human intelligence. Machine learning (ML) is a subset in which algorithms learn patterns from data rather than following hand-coded rules. Deep learning is a subset of ML that uses multilayer neural networks, such as convolutional neural networks (CNNs) for images and video, and can learn features directly from raw data.
- Supervised learning trains on labelled examples (e.g., radiographs labelled by Kellgren-Lawrence grade). Typical algorithms are random forests, support vector machines (SVMs) and CNNs.
- Unsupervised learning finds structure without labels (e.g., clustering patients with chronic low back pain into subgroups).
- Reinforcement learning learns through feedback and is used experimentally to adapt exercise dosage.
Data Types and the Research Pipeline
Physiotherapy research uses three broad data types:
- Structured data: tabular records such as age, BMI, range of motion (ROM), Numeric Pain Rating Scale (NPRS) and Oswestry Disability Index (ODI) scores.
- Unstructured data: free-text clinical notes, open-ended questionnaire responses, images and video.
- Signal and sensor data: inertial measurement units (IMUs), smartphone accelerometers, force plates and electromyography (EMG).
A standard pipeline runs: define the clinical question → collect and clean data → label (establish the reference standard) → select features → split data → train → validate → test externally → monitor after deployment. Feature selection retains only informative variables (e.g., peak knee flexion in swing, trunk lean), which reduces overfitting and improves interpretability. Overfitting occurs when a model memorises noise in the training set and fails on new patients.
Representative Clinical Scenarios
The following scenarios recur throughout AI research and illustrate the concepts in this chapter:
- Chronic low back pain (45-year-old woman, radiating leg pain, failed conservative care): EHR notes are mined by NLP to classify patient subgroups, free-text questionnaires are analysed for symptom trajectories, and a smartphone app drives personalised exercise adjustments.
- Home McKenzie extension exercises: video is analysed by computer vision for repetition counting and form.
- Post-stroke gait (500 hours of video): automated detection of abnormal patterns such as circumduction, foot drop and reduced hip extension.
- Knee osteoarthritis (65-year-old man): AI-assisted radiograph grading alongside clinical diagnosis.
- Parkinson's disease (2-year history, gait instability): smartphone accelerometer gait analysis.
- Total knee arthroplasty (TKA) with persistent stiffness at 2 years: wearable-sensor gait monitoring.
Assessment, Outcome Measures and Normative Values
AI models are only as good as the measurements used as inputs and labels. Students must know the reference values.
| Measure | Key values |
|---|---|
| Comfortable gait speed (10-m walk test) | Healthy adults ~1.2–1.4 m/s. Post-stroke: <0.4 m/s household ambulator, 0.4–0.8 m/s limited community, >0.8 m/s community ambulator (Perry) |
| Normal gait parameters | Cadence ~100–120 steps/min; stance ~60% and swing ~40% of the cycle; double support ~20% |
| Timed Up and Go (TUG) | >13.5 s suggests fall risk in community-dwelling older adults; >11.5 s is associated with falls in Parkinson's disease |
| Hoehn and Yahr stage | 1 (unilateral) to 5 (wheelchair or bedbound) |
| NPRS | MCID ≈ 2 points |
| ODI | MCID ≈ 10 points (0–100 scale) |
| Knee ROM | Normal flexion ~135–140°, full extension 0°; ≥110° needed for stairs and rising from a chair |
| Kellgren-Lawrence grade | 0–4; grade 2 = definite osteophytes, possible joint-space narrowing |
| MRC muscle grade | 0–5; quadriceps ≥4/5 is a common threshold before progressing to loaded closed-chain work |
Other relevant measures include the straight-leg raise (positive for radicular pain at roughly 30–70° of hip flexion), WOMAC and KOOS for knee function, and the 6-minute walk test.
The knee OA scenario fits the ACR clinical criteria: age >50, morning stiffness <30 minutes, crepitus, bony enlargement and decreased ROM. Here, AI imaging tools support rather than replace clinical diagnosis.
AI in Gait Analysis
Gait generates high-dimensional data. In stroke research, ML models (random forests, SVMs) classify gait phases (stance vs swing) and detect deviations from IMU or video data. In fall prediction, supervised models use spatiotemporal parameters (stride length, cadence, double-support time, stride variability) to estimate risk.
For the researcher with 500 hours of post-stroke video, markerless pose estimation (OpenPose, MediaPipe, DeepLabCut) extracts joint coordinates, from which features are computed and classified. Key safeguards:
- Validate joint angles against marker-based 3D motion capture, the reference standard. Sagittal-plane errors of roughly 5–10° are common and may exceed clinically meaningful change.
- Use subject-wise cross-validation (e.g., leave-one-subject-out). Splitting by stride allows strides from the same patient into both training and test sets (data leakage) and inflates accuracy.
- Expert physiotherapist labelling needs documented inter-rater reliability (e.g., kappa).
For the Parkinson's patient, smartphone accelerometry (sampled at ≥50 Hz) can estimate step time variability, asymmetry and freezing episodes. Phone placement and walking surface must be standardised, and outputs should be validated against an instrumented walkway. For the TKA patient, wearable sensors can track gait symmetry over time, but a persistent flexion deficit still requires clinical assessment for infection (CRP, ESR), component malposition, arthrofibrosis or complex regional pain syndrome.
Computer Vision in Exercise Monitoring
Computer vision estimates body landmarks from video and compares movement with reference models. For home McKenzie extension exercises (typically 10 repetitions every 2–3 hours, assessing centralisation of symptoms), a system can count repetitions, estimate lumbar extension and flag poor compliance.
Limitations include camera angle, lighting, clothing, occlusion, and training sets that under-represent older adults, larger body types or people with disability. Computer vision cannot assess symptom response, so a patient reporting peripheralisation, new neurological symptoms or saddle numbness needs clinician review. Video also raises privacy concerns: it contains identifiable faces and home environments, so storage and consent must be addressed explicitly.
AI in Diagnosis and Diagnostic Imaging
CNNs can grade knee radiographs by Kellgren-Lawrence score, detect fractures and segment MRI structures with performance approaching that of radiologists in selected datasets. Chapter points:
- Performance is often inflated by single-centre, retrospective datasets. External validation on different scanners and populations is essential.
- Imaging findings correlate weakly with pain. Many asymptomatic adults show disc degeneration or osteophytes, so AI output must be interpreted alongside the clinical picture.
- In low back pain, red flags (cauda equina features, progressive motor deficit, fever, history of cancer, unexplained weight loss, trauma) override any algorithmic output and trigger urgent referral.
Natural Language Processing in Clinical Documentation
NLP converts unstructured text into structured data. Core tasks include named-entity recognition (pain scores, ROM values, functional scores), negation detection ("no numbness"), classification of notes into patient subgroups, and sentiment or topic modelling of free-text questionnaires. Large language models can summarise notes but may produce plausible but inaccurate text ("hallucination").
Before mining EHR notes: get ethics approval, de-identify names, dates and addresses, build a double-annotated reference set, and test on notes from other clinicians or sites.
For the researcher using EHR notes to classify patients with radiating low back pain:
- Obtain ethics approval and the appropriate legal basis or consent waiver.
- De-identify text (names, dates, addresses) before model development.
- Establish a labelled reference set with double annotation.
- Evaluate on notes from other clinicians or sites, since documentation style varies.
Predictive Modelling and Analytics in Rehabilitation
Predictive modelling estimates an outcome (e.g., ODI at 12 weeks, risk of falling, risk of chronicity) from baseline variables. It is distinct from diagnostic modelling (what is present now) and prognostic modelling (what will happen). Predictive analytics in rehabilitation might flag patients unlikely to meet ROM targets after TKA (e.g., flexion ≥90° by 2 weeks and ≥110° by 6–12 weeks) so that therapy can be intensified earlier.
Adaptive exercise systems that use smartphone activity data should operate in graded stages:
- Baseline: 1–2 weeks of passive monitoring (step count, activity minutes).
- Recommendation: the system suggests changes within pre-set limits (e.g., ≤10–20% weekly load progression).
- Clinician review: adjustments outside limits, flare-ups, or any red flag are escalated to a physiotherapist.
- Evaluation: outcomes are re-measured with validated tools (NPRS, ODI).
Patients should be told what data are collected, who can see them and how recommendations are generated, and should be able to decline or override them.
Evaluating Model Performance
- Sensitivity and specificity: true positive and true negative rates.
- Positive predictive value: depends on prevalence, so a "95% accurate" model can still perform poorly when the condition is rare.
- AUC (area under the ROC curve): 0.5 is chance and 1.0 is perfect discrimination.
- Precision, recall and F1: preferred over accuracy with imbalanced classes.
- Calibration: whether predicted risks match observed risks.
- Regression models: mean absolute error or RMSE, interpreted against the MCID.
Reporting should follow TRIPOD (and TRIPOD+AI), and risk of bias can be appraised with PROBAST. Techniques such as SHAP values improve explainability.
Ethics in Physiotherapy AI Research
- Privacy and consent: health, video and location data are highly sensitive. Follow GDPR, HIPAA or local equivalents, use data minimisation and secure storage, and obtain informed consent that covers secondary use.
- Bias and equity: under-representation of women, older adults, people of different ethnicities or people with disability produces unequal accuracy. Obermeyer et al. showed that a widely used health algorithm systematically disadvantaged Black patients because it used cost as a proxy for need.
- Transparency and accountability: the clinician remains responsible for decisions. Avoid automation bias (over-trusting the output).
- Autonomy: patients may withdraw from AI-driven recommendations without penalty.
- Governance: ethics committee approval, data protection impact assessments and post-deployment monitoring for data drift.
Health, video and location data are highly sensitive. Informed consent should cover secondary use, and patients may withdraw from AI-driven recommendations without penalty.
Clinical Pearls
- Model accuracy on training data is not evidence of clinical usefulness. Look for external validation and, ideally, randomised evaluation.
- Always split by patient, not by stride, image or note.
- Ground truth is only as reliable as its labelling. Check inter-rater agreement.
- A statistically significant change in a sensor-derived measure is not necessarily above the MCID.
- Red flags and clinical reasoning always override algorithmic suggestions.
- Document data sources, preprocessing and limitations so results are reproducible.
Summary
AI and ML offer physiotherapy researchers tools for gait analysis, exercise monitoring, imaging interpretation, NLP of clinical text and outcome prediction. Each depends on high-quality, representative data, validated outcome measures, rigorous subject-wise validation and transparent reporting. Ethical practice requires privacy protection, informed consent, fairness checks and retained clinician oversight.
References
- Rajkomar A, Dean J, Kohane I. Machine learning in medicine. N Engl J Med. 2019;380(14):1347–1358.
- Collins GS, Reitsma JB, Altman DG, Moons KGM. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement. BMJ. 2015;350:g7594.
- 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.
- Moons KGM, Wolff RF, Riley RD, et al. PROBAST: a tool to assess risk of bias and applicability of prediction model studies. Ann Intern Med. 2019;170(1):W1–W33.
- 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.
- Obermeyer Z, Powers B, Vogeli C, Mullainathan S. Dissecting racial bias in an algorithm used to manage the health of populations. Science. 2019;366(6464):447–453.
- World Health Organization. Ethics and Governance of Artificial Intelligence for Health: WHO Guidance. Geneva: WHO; 2021.
- 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.