Pre-Visit AI Symptom-Checking and Shared Decision-Making in Spine Physical Therapy

Background and Rationale AI-based symptom-checking tools, including large language models, are increasingly used by patients as a first point of clinical interpretation before seeking care. Available evidence suggests these tools show moderate and variable diagnostic accuracy and can influence patients' diagnostic beliefs, perceived symptom severity, and healthcare-seeking behavior, raising concerns about pre-diagnostic anchoring. In spine and musculoskeletal care, early presentation, accurate classification, and patient engagement are important determinants of outcomes, and current practice relies on history, physical examination, and shared decision-making under the assumption that patients present without strongly preformed diagnostic expectations. The growing use of AI may alter this dynamic. Evidence on the real-world impact of pre-visit AI use on the physical therapy encounter is currently limited.

Objective The primary objective is to determine whether prior use of AI symptom-checking tools influences shared decision-making during the initial physical therapy evaluation in patients presenting with spine-related musculoskeletal conditions. Secondary objectives examine associations between prior AI use and stage of presentation, agreement between patient-expected and clinician diagnosis, baseline clinical status, functional performance, health-seeking behavior, and short-term clinical outcomes.

Design and Setting This is a prospective observational cohort study conducted in outpatient physical therapy clinics. No study-specific interventions are introduced; all participants receive standard physical therapy care determined by their treating clinician.

Participants and Grouping Consecutive adults (≥18 years) presenting for a new evaluation for a spine-related musculoskeletal complaint (neck, thoracic, or low back pain, with or without radicular symptoms) who can provide informed consent and complete study questionnaires in English are eligible. Patients are excluded for recent spinal surgery within the past 3 months, serious spinal pathology under active medical management (e.g., malignancy, spinal infection, acute fracture), cognitive impairment limiting consent or reliable completion of measures, or current enrollment in another study that may influence clinical decision-making or spine-related outcomes. Participants are classified into two groups based on whether they used an AI symptom-checking tool for their current complaint within the prior 30 days: AI users and non-AI users.

Study Instruments and Measures AI exposure is captured at baseline via a questionnaire assessing AI use (yes/no), tool type, frequency, degree of personalization, and reported influence on care-seeking. Clinical measures include symptom duration, pain intensity (Numeric Pain Rating Scale, 0-10), region-specific disability (Oswestry Disability Index for low back pain or Neck Disability Index for neck pain), and the Five Times Sit-to-Stand test. Diagnostic expectation is assessed by comparing the patient-expected diagnosis with the clinician's classification, categorized as match, partial match, or mismatch. Shared decision-making is measured with the SDM-Q-9 (score range 0-45; higher scores indicate greater patient involvement) immediately after the initial evaluation. Follow-up outcomes at 2 and 6 weeks include pain, disability (ODI/NDI), Global Rating of Change, self-reported adherence, and healthcare utilization.

Procedures At baseline (pre-visit), AI exposure, symptom duration, pain, and disability are collected. During the initial evaluation, the clinician performs a standard physical therapy assessment and classification, and patient expectations are recorded. Immediately post-visit, participants complete the SDM-Q-9. At 2-week and 6-week follow-up, clinical outcomes and adherence are assessed and AI use is re-evaluated. Data are recorded on standardized forms and stored securely using coded identifiers.

Sample Size The primary outcome is SDM-Q-9 following the initial evaluation. Using G*Power (v3.1.9.7), a sample of 128 participants (64 per group) provides 80% power to detect a moderate between-group difference (Cohen's d = 0.5) at a two-sided alpha of 0.05. To account for potential group imbalance and an anticipated attrition of 15-20%, approximately 200 consecutive patients will be recruited.

Statistical Analysis Continuous variables are summarized as mean ± standard deviation and categorical variables as frequencies and percentages. AI users and non-AI users are compared using independent-samples t-tests for continuous variables and chi-square tests for categorical variables. Statistical significance is set at p ≤ 0.05.