Comparison of three large language models in postoperative rehabilitation question answering after anterior cruciate ligament reconstruction based on expert ratings Authors: Tiange Xia, Yijie Zhang, Shaoshuo Li, Yi Zhou, Wenyu Tian, Zhiwei Jiang, Yang Shao, Jianwei Wang Published: 2026 ABSTRACT AbstractBackgroundLarge language models (LLMs) are increasingly used by patients to obtain health information. Postoperative rehabilitation after anterior cruciate ligament reconstruction (ACLR) has distinct phase boundaries and safety considerations. Therefore, responses should be not only clear and understandable, but also medically accurate, safe, and stage-fit. This study compared the performance of three publicly accessible LLMs in standardized post-ACLR rehabilitation question answering.MethodsThis was a standardized, blinded, expert-rated comparative evaluation study. On a single prespecified data collection day in March 2026, 30 English-language rehabilitation questions were submitted separately to GPT-5.4, Doubao, and MiniMax-M2.7. The questions covered five postoperative rehabilitation phases. Responses were anonymized and randomly reordered before blinded rating by five orthopaedic clinicians across five domains: Accuracy, Safety, Stage-fit, Completeness, and Understandability. Paired non-parametric tests, effect size analyses, intraclass correlation coefficients, and linear mixed-effects modelling were used for statistical analysis.ResultsA total of 90 model-generated responses and 450 expert rating records were included. Overall scores differed significantly among the three models (Friedman χ² = 46.067, P < 0.001; Kendall’s W = 0.768). GPT-5.4 achieved the highest overall score (4.61 ± 0.13), followed by MiniMax-M2.7 (4.53 ± 0.19), whereas Doubao had the lowest score (3.86 ± 0.29). GPT-5.4 performed best in Accuracy, Safety, and Stage-fit; MiniMax-M2.7 achieved the highest score for Completeness; and Doubao achieved the highest mean score for Understandability. Inter-rater agreement was good [ICC(3,k) = 0.893], and sensitivity analysis supported the primary findings.ConclusionsThe three models showed distinct rating profiles in standardized single-turn post-ACLR rehabilitation question answering. Evaluation of patient-facing rehabilitation information should not rely solely on linguistic fluency, but should prioritize medical accuracy, safety, and Stage-fit. These findings provide preliminary benchmark evidence in a phase-sensitive rehabilitation setting, but they should not be interpreted as evidence supporting clinical implementation, clinician substitution, or patient benefit.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12891-026-10051-4.Keywords: Anterior cruciate ligament reconstruction, Postoperative rehabilitation, Large language models, Expert evaluation, Patient education Introduction Anterior cruciate ligament (ACL) injury is a common injury in sports medicine and can substantially impair knee stability and activity capacity. For patients seeking to return to a high level of activity, anterior cruciate ligament reconstruction (ACLR) is an important treatment option [1, 2]. However, surgery does not mark the end of recovery; postoperative rehabilitation after ACLR must be continuously adjusted according to tissue healing, symptom response, functional status, and objective assessment results, with training goals, progression criteria, and safety boundaries varying across rehabilitation phases [3–5]. During this process, patients often need supplementary information beyond outpatient visits or formal rehabilitation sessions [6]. Traditionally, such information has been obtained from written educational materials, internet searches, videos, or other multimedia patient education resources. With the increasing availability of large language models (LLMs), patients can now directly ask specific rehabilitation-related questions in a conversational manner and receive immediate, coherent responses that may appear individualized [7–9]. This interactive format makes LLMs potentially attractive in patient education settings [10]. Because rehabilitation decisions after ACLR are phase dependent, post-ACLR questions differ from general health information queries. Relevant advice cannot be based only on postoperative time, symptom descriptions, or patients’ subjective perceptions. Patients often do not provide surgical details, rehabilitation progress, functional recovery, or objective assessment results, and public models cannot perform physical examination or functional testing [11–14]. As a result, a fluent and well-structured response may still be inadequate if it is poorly matched to the rehabilitation phase or omits safety boundaries. In recent years, LLMs have been widely used in medical question answering, patient consultation, and health information evaluation, and related research in orthopaedics has also increased [15–18]. Existing orthopaedic studies have addressed patient education after total knee arthroplasty, common patient questions about osteoarthritis, patient consultation regarding reverse shoulder arthroplasty, and the readability of orthopaedic patient education materials [17, 19–21]. In contrast, post-ACLR rehabilitation question answering has more explicit phase dependence and progression-related decision requirements. Therefore, it requires dedicated evaluation using a unified question set and a blinded expert rating framework. Accordingly, this study compared the performance of three publicly accessible LLMs in standardized post-ACLR rehabilitation question answering. Responses were rated by blinded experts across five domains: Accuracy, Safety, Stage-fit, Completeness, and Understandability. The aim was to provide benchmark evidence for model performance in a phase-sensitive rehabilitation setting, not to evaluate real-world clinical implementation or patient outcomes. Methods Study design This was a blinded expert comparative evaluation study under standardized single-turn question-answering conditions. Three publicly accessible LLMs were compared. Each standardized rehabilitation question was submitted separately to the three models. The generated responses were anonymized and randomly reordered before expert rating by orthopaedic clinicians. The main unit of analysis was the question-by-model response, with the mean score across the five experts used as the item-level analysis value. The study workflow included question development, model response generation, anonymization and randomization, blinded expert rating, and statistical analysis. Model selection This study evaluated GPT-5.4, Doubao, and MiniMax-M2.7. During the study period, all three models were accessible to general users through public web interfaces. In this study, the terms “GPT-5.4,” “Doubao,” and “MiniMax-M2.7” refer to the model or product names displayed in their respective publicly accessible web interfaces at the time of data collection. Stable backend model identifiers were not available through these public interfaces. GPT-5.4 was selected as an internationally used general-purpose model; Doubao was selected as a model familiar to users in mainland China; and MiniMax-M2.7 was selected as another publicly accessible multilingual model. Model selection was based on public accessibility, patient-facing availability, and different platform backgrounds. Other publicly available models during the same period, such as Claude, Gemini, and DeepSeek, were not included; therefore, this study does not represent a comprehensive comparison of all available models. Question development The research team developed 30 formal study questions with reference to ACLR rehabilitation guidelines, related literature, and common postoperative consultation topics encountered in clinical practice [1–4, 22–25]. Candidate questions were discussed by the research team, and questions that were repetitive, too broad, too narrow, or unsuitable for a single-turn question-answering format were removed or revised. The final question set covered five rehabilitation phases: postoperative 0–2 weeks, 2–6 weeks, 6–12 weeks, 3–6 months, and beyond 6 months, with six questions in each phase. Equal allocation was used to ensure balanced weighting of different phases in model comparison and was not intended to reflect the real-world frequency distribution of patient questions. The number of 30 questions was based primarily on balanced coverage of the five phases, feasibility of expert rating, and the number of questions used in previous comparable LLM evaluation studies. No formal power calculation was performed because this study was designed as an exploratory benchmark evaluation rather than a confirmatory clinical trial. These questions simulated single-turn rehabilitation questions that patients might independently enter into a public chatbot. Examples included: “My knee is quite swollen on day 3 after ACL reconstruction. Is this normal, and when should I be concerned?”, “On day 5 after surgery, the wound looks red, warm, and has a little drainage. Should I go back to the hospital right away?”, “My calf is swollen, tight, and painful after surgery. Is this a normal postoperative reaction or a sign of a blood clot?”, “Can I start running 3 months after ACL reconstruction?”, “Can I start shuttle running and cutting drills 4 months after surgery?” All 30 English-language questions and their corresponding rehabilitation phases are provided in Supplementary Table S1. Supplementary Table S1 also labels the main content theme of each question to improve transparency and interpretability of the question set. Model response generation All prompts were written in English. Each question was entered into a separate new conversation for each model. No additional clinical background was provided, no follow-up questions were asked, and no model output was manually edited. Responses were generated through public web interfaces rather than a controlled application programming interface (API) environment. All 90 responses were generated consecutively on the same prespecified data collection day in March 2026 to reduce the influence of model version updates. For each question-by-model combination, the first complete response returned by the public interface was retained. Because the study used public web interfaces, specific model parameters and settings, including temperature, token limits, reasoning-related modes, backend system prompts, and stable version identifiers, could not be fully controlled or consistently obtained. The researchers did not intentionally enable web browsing, file upload, image input, or other external retrieval functions. Rating domains and operational definitions Response quality was evaluated across five domains: Accuracy, Safety, Stage-fit, Completeness, and Understandability. These domains addressed whether the response was medically accurate, avoided potentially harmful advice, matched the postoperative rehabilitation phase, covered key information, and was easy for patients to understand. Accuracy, Safety, and Stage-fit are partly related, since advice that is poorly matched to the rehabilitation phase may also be inaccurate or unsafe. We retained them as separate domains because they reflect different clinical concerns: factual correctness, avoidance of harm, and phase appropriateness. Detailed operational definitions are provided in Table 1. Table 1Operational definitions of the five rating domainsRating domainOperational definitionAccuracyFactual correctness of the response in medical terms; whether it is consistent with authoritative guidelines, expert consensus, or current evidence, and free of clearly misleading information.SafetyWhether the response avoids potentially harmful recommendations, clearly communicates precautions, and appropriately identifies warning signs or circumstances requiring reassessment or urgent medical attention.Stage-fitWhether the response matches the stated postoperative rehabilitation phase, including timing, permitted activity level, progression criteria, and risk-control requirements; and whether it avoids premature or delayed recommendations that conflict with the context.CompletenessWhether the response covers the main clinically relevant elements required to answer the question, including the core issue, practical guidance, key training points, precautions, and necessary follow-up recommendations.UnderstandabilityWhether the response is clearly expressed, logically organized, patient-friendly, and easy for patients to understand and apply in practice. Operational definitions of the five rating domains The five-domain rating framework was developed with reference to previous evaluation methods for LLM medical question answering, patient education quality assessment, and understandability assessment, and was adapted to the phase-specific characteristics of post-ACLR rehabilitation [21, 26–29]. This framework is not a formally validated questionnaire; therefore, score differences should be interpreted as comparative signals based on expert ratings rather than as established thresholds of clinical significance. Expert raters and calibration The expert panel consisted of five orthopaedic clinicians, each with 5–10 years of clinical experience in sports medicine and familiarity with post-ACLR rehabilitation principles. The raters were aware of the study aim and rating task and participated voluntarily. Before formal rating, all raters reviewed a unified scoring guide and participated in pre-rating calibration. The calibration covered operational definitions of the five rating domains, scoring anchors, and typical response scenarios likely to cause disagreement. The raters then scored a small set of example responses that were not included in the formal analysis. Disagreements were discussed to standardize rating criteria. These pilot ratings were used only for calibration and were not included in the final statistical analysis. General scoring anchors are provided in Supplementary Table S2, and domain-specific scoring guidance is provided in Supplementary Table S3. For rehabilitation recommendations where reasonable differences may exist across guidelines or surgeon preferences, raters were instructed to prioritize mainstream guidelines, expert consensus, and risk-control principles. When a recommendation reflected reasonable clinical variation, it was not judged as incorrect solely because it differed from the rater’s personal preference. Instead, the rating focused on whether the response acknowledged the need for individualized assessment, rehabilitation phase boundaries, progression criteria, and safety warnings. The expert raters were not involved in question development. All formal rating responses were anonymized and randomly reordered, and raters were blinded to model identity during scoring. Study outcomes The primary outcome was the overall score for each model, defined as the equally weighted mean of the five domain scores for each response. Equal weighting was used to provide a transparent and reproducible summary metric for model comparison. Because Safety and Accuracy may have higher clinical priority, individual domain scores were also interpreted, and the overall score was not treated as the only indicator of response quality. Secondary outcomes included individual domain scores and phase-stratified overall scores. Inter-rater agreement and linear mixed-effects modelling were used as supplementary analyses of rating reliability and robustness of the findings. Statistical analysis Because Likert-type ratings are ordinal data, paired non-parametric methods were used for the main statistical inference. Means ± standard deviations were reported as descriptive results rather than as primary inferential estimates. Ratings from the five experts were first aggregated to the question-by-model level, and question was used as the paired block. All 30 questions had complete observations across the three models, satisfying the block structure required for the Friedman test. Overall scores and individual domain scores were compared among the three models using the Friedman test, with Kendall’s W reported as the omnibus effect size. When the omnibus test was significant, pairwise comparisons were performed using paired Wilcoxon signed-rank tests with Holm correction for multiple comparisons. Rank-biserial correlation was reported as the pairwise effect size. P values were reported as exact values when P ≥ 0.001 and as P < 0.001 otherwise. Inter-rater agreement was assessed using a two-way mixed-effects, consistency-type ICC(3,k). As a sensitivity analysis, a linear mixed-effects model was constructed using the original question-by-model-by-rater level overall scores as the dependent variable. Model type was treated as a fixed effect, and question and rater were included as random intercepts, with Doubao as the reference group. Phase-stratified analyses were exploratory and followed the same non-parametric comparison workflow. Because each rehabilitation phase contained only six questions, model-by-phase interaction testing was not treated as confirmatory. All tests were two-sided, and P < 0.05 was considered statistically significant. Statistical analyses were performed in R version 4.4.1. Ethics statement This study did not involve patients, patient data, biological samples, imaging data, or clinical intervention. The formal ethics approval and consent statement is provided in the Declarations section. Results Overall model performance A total of 90 model-generated responses were included, corresponding to 450 expert rating records. Overall scores differed significantly among the three models (Friedman χ² = 46.067, P < 0.001; Kendall’s W = 0.768). GPT-5.4 achieved the highest overall score (4.61 ± 0.13), followed by MiniMax-M2.7 (4.53 ± 0.19), whereas Doubao had the lowest score (3.86 ± 0.29). A Kendall’s W of 0.768 indicated a strong and relatively consistent effect size for overall score differences across the same standardized question set. Overall model performance is shown in Fig. 1, and detailed statistical results are presented in Tables 2 and 3. Item-level distribution of overall scores for the three large language models. Boxplots show the distribution of overall scores across the 30 questions. The centre line represents the median, boxes represent the interquartile range, and whiskers represent values within 1.5 times the interquartile range. Dots represent item-level overall scores for individual questions. Horizontal brackets indicate results of Holm-adjusted paired Wilcoxon signed-rank tests; * indicates P < 0.05, and *** indicates P < 0.001 Table 2Comparison of overall and domain-specific scores across the three models: Descriptive scores and omnibus testsOutcomeGPT-5.4DoubaoMiniMax-M2.7Friedman χ²Kendall’s WOverall P valueOverall score4.61 ± 0.133.86 ± 0.294.53 ± 0.1946.0670.768< 0.001Accuracy4.86 ± 0.113.54 ± 0.444.31 ± 0.3159.5130.992< 0.001Safety4.85 ± 0.133.99 ± 0.294.42 ± 0.1955.5900.926< 0.001Stage-fit4.81 ± 0.203.61 ± 0.504.37 ± 0.3759.5130.992< 0.001Completeness4.45 ± 0.283.21 ± 0.294.61 ± 0.1952.7030.878< 0.001Understandability4.07 ± 0.214.95 ± 0.094.92 ± 0.1950.6990.845< 0.001 Comparison of overall and domain-specific scores across the three models: Descriptive scores and omnibus tests Table 3Comparison of overall and domain-specific scores across the three models: Holm-adjusted pairwise comparisons, P value (rank-biserial r)OutcomeGPT-5.4 vs. DoubaoGPT-5.4 vs. MiniMax-M2.7MiniMax-M2.7 vs. DoubaoOverall score< 0.001 (+ 1.00)0.021 (+ 0.48)< 0.001 (+ 1.00)Accuracy< 0.001 (+ 1.00)< 0.001 (+ 1.00)< 0.001 (+ 1.00)Safety< 0.001 (+ 1.00)< 0.001 (+ 0.98)< 0.001 (+ 1.00)Stage-fit< 0.001 (+ 1.00)< 0.001 (+ 1.00)< 0.001 (+ 1.00)Completeness< 0.001 (+ 1.00)0.002 (-0.72)< 0.001 (+ 1.00)Understandability< 0.001 (-1.00)< 0.001 (-1.00)0.937 (-0.07)Note: Data are item-level mean scores presented as mean ± standard deviation for descriptive comparison. Inferential tests used paired non-parametric methods. Omnibus comparisons were performed using the Friedman test, with Kendall’s W reported as the omnibus effect size. Post hoc pairwise comparisons were performed using paired Wilcoxon signed-rank tests with Holm correction. Values in parentheses are signed rank-biserial correlations; positive values indicate higher scores for the first named model Comparison of overall and domain-specific scores across the three models: Holm-adjusted pairwise comparisons, P value (rank-biserial r) Note: Data are item-level mean scores presented as mean ± standard deviation for descriptive comparison. Inferential tests used paired non-parametric methods. Omnibus comparisons were performed using the Friedman test, with Kendall’s W reported as the omnibus effect size. Post hoc pairwise comparisons were performed using paired Wilcoxon signed-rank tests with Holm correction. Values in parentheses are signed rank-biserial correlations; positive values indicate higher scores for the first named model Post hoc pairwise comparisons showed that GPT-5.4 had a higher overall score than Doubao (Holm-adjusted P < 0.001; rank-biserial r = + 1.00) and MiniMax-M2.7 (Holm-adjusted P = 0.021; rank-biserial r = + 0.48). MiniMax-M2.7 also outperformed Doubao (Holm-adjusted P < 0.001; rank-biserial r = + 1.00). Although the difference between GPT-5.4 and MiniMax-M2.7 reached statistical significance, the magnitude of the difference was small. Comparison across individual domains Across the five domains of Accuracy, Safety, Stage-fit, Completeness, and Understandability, the omnibus comparisons among models were all statistically significant (all P < 0.001; Kendall’s W range, 0.845–0.992). Accuracy and Stage-fit had the highest Kendall’s W values. GPT-5.4 achieved the highest scores for Accuracy, Safety, and Stage-fit; MiniMax-M2.7 achieved the highest score for Completeness; and Doubao achieved the highest mean score for Understandability. The domain-specific scoring patterns are shown in Fig. 2. Domain-specific scores of the three large language models. Bar plots show mean item-level scores across the 30 questions for Accuracy, Safety, Stage-fit, Completeness, and Understandability. Error bars indicate standard deviations For Completeness, pairwise comparison showed that MiniMax-M2.7 scored higher than GPT-5.4 (Holm-adjusted P = 0.002; rank-biserial r = -0.72). For Understandability, there was no significant difference between Doubao and MiniMax-M2.7 (Holm-adjusted P = 0.937; rank-biserial r = -0.07). These results indicate that the ranking of the three models was not consistent across different rating domains. Phase-stratified analysis After stratification by rehabilitation phase, all phase-specific Friedman tests were statistically significant, suggesting that the general pattern of model differences was observed across different phases. However, after Holm correction, none of the pairwise comparisons within individual phases reached statistical significance, with adjusted P values ranging from 0.094 to 1.000. This pattern indicates phase-level omnibus differences, but the study was not designed to identify corrected pairwise advantages within individual phases. Thus, no single model could be interpreted as having a stable phase-specific advantage. Item-level phase-stratified performance is shown in Fig. 3. Phase-stratified item-level overall scores. The heatmap shows item-level overall scores for the three models across the 30 standardized questions, grouped by rehabilitation phase. Colours represent overall score values Inter-rater agreement and sensitivity analysis The mean ratings across the five expert raters showed good agreement [ICC(3,k) = 0.893]. A linear mixed-effects model based on all 450 original rating records showed findings consistent with the primary analysis after accounting for random effects of question and rater (Table 4). Table 4Sensitivity analysis based on the linear mixed-effects modelItemEstimateStandard error95% CIP valueEstimated marginal means Doubao3.8600.0393.783–3.937— GPT-5.44.6080.0394.531–4.685— MiniMax-M2.74.5250.0394.448–4.602—Fixed effects and pairwise comparisons GPT-5.4 vs. Doubao0.7480.0340.681–0.815< 0.001 MiniMax-M2.7 vs. Doubao0.6650.0340.598–0.732< 0.001 GPT-5.4 vs. MiniMax-M2.70.0830.0340.016–0.1500.0156Random-effect variance Variance of random intercept for question0.029——— Variance of random intercept for rater0.000135——— Residual variance0.088———Note: The linear mixed-effects model used overall score from the original rating records as the dependent variable, model type as a fixed effect, and question and rater as random intercepts, with Doubao as the reference group. The overall score was defined as the equally weighted mean of Accuracy, Safety, Stage-fit, Completeness, and Understandability. Pairwise P values were adjusted using the Holm method Sensitivity analysis based on the linear mixed-effects model Note: The linear mixed-effects model used overall score from the original rating records as the dependent variable, model type as a fixed effect, and question and rater as random intercepts, with Doubao as the reference group. The overall score was defined as the equally weighted mean of Accuracy, Safety, Stage-fit, Completeness, and Understandability. Pairwise P values were adjusted using the Holm method Compared with Doubao, GPT-5.4 had a significantly higher overall score (beta = 0.748, SE = 0.034, 95% CI 0.681–0.815, P < 0.001), and MiniMax-M2.7 also had a significantly higher overall score (beta = 0.665, SE = 0.034, 95% CI 0.598–0.732, P < 0.001). A statistically significant difference remained between GPT-5.4 and MiniMax-M2.7 (beta = 0.083, SE = 0.034, 95% CI 0.016–0.150, P = 0.0156). The random-intercept variance attributable to raters was small (0.000135), lower than the variance attributable to questions (0.029) and the residual variance (0.088), suggesting that the between-model differences were not primarily driven by systematic bias from individual raters. Overall, the sensitivity analysis supported the robustness of the primary findings. Discussion Main findings This study found distinct expert-rated profiles among the three models under standardized post-ACLR rehabilitation question-answering conditions. GPT-5.4 performed more consistently in Accuracy, Safety, and Stage-fit; MiniMax-M2.7 performed better in Completeness; and Doubao had the highest mean score for Understandability but weaker performance in clinically oriented domains. These findings suggest that LLM response quality should be judged by domain, rather than by overall score or fluency alone. For post-ACLR rehabilitation, clear language is useful, but clinical acceptability depends more on phase-appropriate advice, avoidance of premature progression, recognition of warning signs, and referral for professional reassessment when needed. These findings describe standardized expert-rated response quality rather than real-world clinical effectiveness. Comparison with existing literature Previous medical and orthopaedic evaluations have shown that LLM responses can be plausible and readable, but their performance varies by task, question type, and evaluation domain [30–33]. The present results similarly indicate that readability or conversational quality should not be treated as a proxy for clinical reliability. In orthopaedic patient education, Yuncu et al. showed that different LLMs had similar expert-rated accuracy and clinical applicability in patient education after total knee arthroplasty, while readability metrics differed [17]. Other orthopaedic studies also suggest that model performance varies by question type and evaluation metric [19–21]. The added value of the present study lies in its focus on post-ACLR rehabilitation and in the inclusion of Stage-fit as an independent domain in blinded expert rating. ACLR rehabilitation involves progressive advancement and risk control. Therefore, response quality depends not only on whether the content is complete and easy to understand, but also on whether it respects phase-specific safety boundaries [34]. The inconsistent performance of different models across different rating domains further suggests that a single overall score is insufficient to characterize model performance in patient education settings. Phase-sensitive rehabilitation and stage-fit The rehabilitation tasks after ACLR differ by phase. Early rehabilitation emphasizes protection and recognition of complications; the middle phase focuses on range of motion, strength, and load management; and the later phase emphasizes running, jumping, cutting, and functional assessment before return to sport [34, 35]. Responses that provide only general time-based advice without specifying progression criteria and safety boundaries may be clinically insufficient, even if the content appears complete [22, 23, 36]. In this study, Stage-fit showed strong discrimination among models and captured rehabilitation-specific differences that general quality ratings may not reflect. In questions related to return to running or return to sport, whether the response emphasized strength, swelling control, neuromuscular control, psychological readiness, and functional testing directly affected its clinical acceptability. It should be noted that each phase contained only six questions, and phase-specific pairwise comparisons did not remain significant after correction. Therefore, the present findings are better suited to support overall differences in model performance rather than to identify a stable superior model within a specific rehabilitation phase. Fluency-safety mismatch Doubao had the highest mean score for Understandability but relatively lower Accuracy, Safety, and Stage-fit. This reflects an important issue in patient-facing LLM applications: clear and natural language does not necessarily indicate clinical reliability. In rehabilitation settings, patients may place greater trust in responses that are friendly and well organized. However, if a response omits warning signs, underestimates progression requirements, or fails to advise professional reassessment, fluent expression may obscure potential risk [14, 32]. Therefore, in high-risk rehabilitation questions, Safety and Stage-fit should take priority over linguistic presentation. Role of clinician-patient contact Post-ACLR rehabilitation depends on ongoing clinical evaluation. A single-turn model response cannot obtain surgical details, concomitant injuries, changes in pain and swelling, range of motion, strength recovery, motor control, or psychological readiness, and it cannot verify the accuracy of patient-reported information [1, 2, 37]. These missing inputs are especially important when patients ask about progression, return to running, or warning symptoms. Thus, even well-rated responses should be viewed as text outputs generated under limited clinical context, not as substitutes for clinician-patient assessment. Whether such responses change patient behaviour, adherence, or outcomes remains unknown and requires studies involving patients. Equal weighting and clinical priorities This study used an equally weighted mean of the five domains to calculate the overall score, with the aim of providing a transparent and reproducible summary metric for comparison. However, equal weighting does not imply that the five domains have equal clinical importance. In real rehabilitation practice, Safety and Accuracy should generally take priority over Understandability or Completeness [22, 28]. Therefore, the overall score should be interpreted together with the individual domain scores, rather than as a single standalone quality judgment. Future studies may explore weighted scoring systems based on expert consensus, patient needs, or risk level. Because no validated minimal clinically important difference exists for this five-domain rating framework, small numerical differences, especially between GPT-5.4 and MiniMax-M2.7, should be interpreted cautiously and together with the domain-specific results. Practical implications Based on the design of this study, the practical implications should be limited to information support rather than decision-making [38, 39]. Potential use cases include drafting patient education materials, explaining general rehabilitation principles, helping patients prepare questions for follow-up visits, and improving educational wording after clinician review. The different rating profiles across models suggest that rehabilitation-oriented systems may require domain-specific safeguards, guideline constraints, phase-specific safety filters, and clinician review before use. Future studies may compare clinician responses, direct model responses, and clinician-reviewed model responses in terms of patient understanding, risk recognition, rehabilitation adherence, and clinical outcomes. Limitations This study has several limitations. First, it used a single-turn simulated question-answering design and did not evaluate interactive or longitudinal use; therefore, the findings cannot be generalized to real clinical communication. Second, patient-centred outcomes such as usability, behavioural impact, adherence, and clinical outcomes were not assessed, and expert ratings are not equivalent to patients’ perceived usefulness. Third, the question set was developed based on guidelines and clinical experience, but was not validated by patients, real consultation logs, or Delphi consensus. The equal allocation of 30 questions across phases also does not reflect the real-world frequency distribution of patient questions. Fourth, the five-domain rating framework was not a formally validated scale, equal weighting does not fully reflect clinical priorities, and expert ratings remain subjective. Future studies should formally validate this framework, including its construct validity, inter-rater reliability in larger rater panels, and ability to distinguish clinically meaningful differences in rehabilitation advice. Fifth, only three publicly accessible models were included, and all prompts were written in English; therefore, the findings cannot be generalized to Chinese or other language contexts. Sixth, the models were accessed through public web interfaces, and specific version identifiers, temperature, token limits, and backend settings could not be fully controlled. Seventh, each question-by-model combination was sampled only once, repeated-response stability was not tested, and multi-turn interactions were not analyzed. These factors should be considered when interpreting the scope and transferability of the findings. Conclusions In standardized single-turn post-ACLR rehabilitation question answering, the three LLMs showed distinct expert-rated performance profiles, with different strengths across Accuracy, Safety, Stage-fit, Completeness, and Understandability. These findings support evaluation of patient-facing rehabilitation information beyond linguistic fluency or clarity of expression. In phase-sensitive rehabilitation settings, medical accuracy, safety, and Stage-fit should remain central evaluation criteria. The present results should be understood as benchmark evidence under standardized single-turn conditions, not as evidence of clinical implementation or patient benefit. Future clinical use would require clinician review, phase-specific safety safeguards, explicit risk warnings, and patient usability evaluation. Supplementary Information Supplementary Material 1. Acknowledgements Not applicable. Abbreviations Anterior cruciate ligament Anterior cruciate ligament reconstruction Artificial intelligence Confidence interval Intraclass correlation coefficient Large language model Standard error Authors’ contributions Tiange Xia: Conceptualization; Methodology; Investigation; Data curation; Writing—original draft; Writing—review & editing. Yijie Zhang: Investigation; Data curation; Writing—review & editing. Shaoshuo Li: Methodology; Investigation; Data curation; Writing—review & editing. Yi Zhou: Investigation; Data curation; Writing—review & editing. Wenyu Tian: Investigation; Data curation; Writing—review & editing. Zhiwei Jiang: Investigation; Data curation; Writing—review & editing. Yang Shao: Conceptualization; Methodology; Writing—review & editing; Supervision. Jianwei Wang: Conceptualization; Methodology; Writing—review & editing; Supervision. Tiange Xia, Yijie Zhang, and Shaoshuo Li contributed equally to this work. All authors read and approved the final manuscript and agree to be accountable for all aspects of the work. Funding This research was funded by the National Natural Science Foundation of China (Grant Nos. 82405520 and 82274546) and the Jiangsu Province Traditional Chinese Medicine Science and Technology Development Program (Grant No. QN202322). Data availability The datasets generated and analyzed during the current study are included in this published article and its supplementary materials. Additional materials are available from the corresponding author on reasonable request. Declarations Ethics approval and consent to participate This study did not involve patients, patient data, biological samples, imaging data, or clinical intervention. Expert raters evaluated de-identified large language model outputs as part of a professional academic assessment. All raters were aware of the study aim and rating task and participated voluntarily. According to the authors’ institutional policy, this type of expert evaluation of non-human-generated text outputs did not require formal ethics approval or informed consent. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Tiange Xia, Yijie Zhang and Shaoshuo Li contributed equally to this work. Contributor Information Yang Shao, Email: wxzy074@njucm.edu.cn. Jianwei Wang, Email: wxzy006@njucm.edu.cn. References Associated Data Supplementary Materials Supplementary Material 1. Data Availability Statement The datasets generated and analyzed during the current study are included in this published article and its supplementary materials. Additional materials are available from the corresponding author on reasonable request.