💻 computer science

Predicting Expert-Rated Quality of Large Language Model Responses to Orthopaedic Patient Questions Using Computational Intelligence

This study evaluates the clinical quality of four large language models on orthopaedic patient questions through expert ratings, finding generally high performance with no statistically significant differences between models, while demonstrating that combining NLP and semantic features can weakly predict expert-rated accuracy despite limitations in automated quality prediction.

Original authors: Volkan Kızılkaya, Murat Aslanyürek, Canan Aslanyürek, Ali Levent, Mehmet Melih Asoğlu

Published 2026-10-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Volkan Kızılkaya, Murat Aslanyürek, Canan Aslanyürek, Ali Levent, Mehmet Melih Asoğlu

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the modern landscape of healthcare, patients increasingly turn to digital tools to understand their conditions, seeking clear answers to complex medical questions. Among these tools, large language models have emerged as powerful engines capable of generating human-like text, offering explanations that can be tailored to a reader's level of understanding. These systems work by analyzing vast amounts of written information to predict how words should follow one another, allowing them to construct coherent responses to inquiries ranging from general health advice to specific surgical procedures. However, the ability to generate fluent text does not automatically guarantee that the information is medically correct, complete, or safe for a patient to rely on. As these artificial intelligence systems become more integrated into patient education, a critical question arises: can we trust them to provide accurate guidance for serious conditions like knee injuries or joint replacements, and is there a way to measure their quality without a doctor reading every single response?

To address this, a team of researchers from Turkey set out to test four of the most advanced language models currently available: ChatGPT, Claude, DeepSeek, and Gemini. They focused their investigation on three common orthopaedic topics that frequently concern patients: injuries to the anterior cruciate ligament, tears in the meniscus cartilage, and total knee replacement surgery. The researchers compiled fifty distinct questions that a typical patient might ask about these conditions, covering everything from recovery times to the risks of surgery. They then submitted each question to all four models, generating a total of two hundred unique responses. To determine the quality of these answers, three orthopaedic specialists independently reviewed every single response. The experts were kept in the dark about which model produced which answer, ensuring their judgments were based solely on the content. They scored each response on how accurate the medical information was, how complete the explanation was, and whether it contained any misleading or incorrect details.

The results revealed that all four models performed remarkably well, generally producing responses that were medically accurate and comprehensive. The experts rated the answers highly, with most scores falling near the top of the possible range. Interestingly, while one model appeared slightly more accurate and another slightly more complete in the raw numbers, statistical analysis showed that these differences were not significant enough to declare one system superior to the others. In other words, for the broad categories of knee injuries and joint replacements, the models were essentially equal in their ability to provide high-quality information. However, the study also uncovered that performance was not uniform across every single question. When the researchers looked at individual questions, they found that some were much harder for the models to answer correctly than others. The most difficult questions tended to be those about total knee replacements, particularly inquiries regarding how long a patient stays in the hospital. These questions often depend on specific hospital protocols or individual patient health factors, making it difficult for a computer to give a single, universally correct answer. In these instances, the models sometimes provided answers that were too vague, overly detailed, or slightly off-topic, even if they were not outright dangerous.

A second major goal of the research was to see if a computer could predict how an expert would rate a response, simply by analyzing the text itself. The researchers used computational methods to break down the language of the responses, looking at factors like sentence length, the variety of words used, and how closely the answer matched the meaning of the question. They built a machine learning model to guess the expert scores based on these features. The results were mixed. While the computer model could predict the general trend of the scores better than random guessing, it struggled to distinguish between a very good answer and a perfect one. The experts had given most of the answers high marks, creating a situation where there was very little variation for the computer to learn from. Consequently, the automated system could not reliably identify the few lower-quality responses on its own. The study also found that the different models often produced answers that were semantically very similar to one another, meaning they used similar concepts and phrasing, yet this similarity did not guarantee that the answers were clinically perfect.

Ultimately, the study concludes that current artificial intelligence systems are capable of generating responses that are generally safe and informative for orthopaedic patients. They can serve as useful supportive tools to help explain medical concepts. However, they are not yet a replacement for professional medical judgment. The research suggests that while these systems are strong, they can still stumble on questions that require nuanced, context-specific knowledge, and their performance can vary depending on the specific topic. Furthermore, the attempt to create an automated quality control system to screen these answers before they reach a patient showed promise but remains limited. The findings indicate that while we can measure the quality of these responses to some degree using computer analysis, human experts are still essential for verifying the accuracy of medical advice, especially for complex or context-dependent situations. The technology is a powerful assistant, but it is not yet an independent authority.

Technical Summary: Predicting Expert-Rated Quality of LLM Responses in Orthopaedics

Problem Statement
Large Language Models (LLMs) are increasingly utilized in healthcare for patient education and clinical information access. However, their performance varies significantly across models, clinical domains, and specific question types. A critical gap exists in the systematic assessment of clinical quality dimensions—specifically accuracy, completeness, and misinformation—beyond simple linguistic fluency. Furthermore, while expert evaluation remains the gold standard, it is resource-intensive. The study addresses two primary challenges: (1) quantifying the comparative clinical quality of current LLMs (ChatGPT, Claude, DeepSeek, and Gemini) in the context of orthopaedic patient questions, and (2) investigating whether expert-rated quality scores can be automatically predicted using computational intelligence (Natural Language Processing, semantic representation, and machine learning) to enable automated pre-screening.

Methodology
The study employed a rigorous, two-pronged approach involving clinical evaluation and computational modeling:

Dataset Construction:
- Questions: 50 patient-oriented questions were curated across three clinical categories: Anterior Cruciate Ligament (ACL) injuries (n=15), meniscal tears (n=15), and Total Knee Arthroplasty (TKA) (n=20).
- Generation: Four LLMs (ChatGPT, Gemini, Claude, DeepSeek) generated responses to all 50 questions, resulting in 200 unique responses. Data collection was standardized using browser automation to ensure consistent timing and prompt delivery.
- Expert Evaluation: Three orthopaedic specialists independently evaluated all 200 responses (600 total evaluations) in a blinded manner.
- Metrics: Responses were scored on Accuracy (0–4), Completeness (0–4), and Misinformation (0–2). Free-text comments were also collected to characterize specific issues.
Statistical Analysis:
- Model performance was compared using paired statistical methods (Friedman test, Wilcoxon signed-rank test with Holm correction) to account for the fact that all models answered the same questions.
- Analyses were conducted at the overall level, clinical category level, and individual question level.
- Inter-rater reliability was assessed using Krippendorff's alpha and Intraclass Correlation Coefficients (ICC).
Computational Intelligence & Machine Learning:
- Feature Extraction:
  - NLP Features: Word count, sentence count, lexical diversity, readability scores (Flesch-Kincaid, Flesch Reading Ease).
  - Semantic Features: Sentence-BERT (SBERT) embeddings were used to calculate Question-Response Cosine Similarity and inter-model semantic similarity.
  - Explainable Textual Features: Counts of numerical expressions, list structures, hedging, clinical warnings, and risk/treatment terms.
- Modeling: An RBF-Support Vector Regression (RBF-SVR) model was developed to predict the mean expert scores for Accuracy and Completeness.
- Validation Strategy: To prevent data leakage (where responses to the same question appear in both training and test sets), the study utilized nested grouped cross-validation. The grouping variable was Question_ID, ensuring all four LLM responses to a single question remained within the same fold. Hyperparameter tuning was performed via an inner loop, while final performance was evaluated on the outer test folds.

Key Results

Clinical Performance:
- All four LLMs demonstrated generally high clinical quality, with mean Accuracy scores ranging from 3.607 to 3.827 (out of 4) and low Misinformation scores (0.093–0.127 out of 2).
- No Statistically Significant Differences: Overall, there were no statistically significant differences among the models for Accuracy (p=0.070), Completeness (p=0.361), or Misinformation (p=0.451).
- Variability: Performance varied by clinical category. A significant overall model effect was observed for Accuracy in the TKA category (p=0.041), though no specific pairwise difference remained significant after Holm correction.
- Question-Level Insights: 13 questions were identified as "lower-performing" (mean Accuracy ≤ 3.667), with 7 belonging to the TKA category. The lowest-performing question (Q44) concerned hospital length of stay, a context-dependent variable.
- Semantic Similarity: Responses from different models to the same question showed high semantic similarity (mean SBERT cosine similarity: 0.813–0.850), yet this did not correlate directly with clinical accuracy.
- Expert Comments: The most frequent issue identified in free-text comments was "unnecessary, overly detailed, or off-topic content" (72 responses), rather than direct misinformation.
Machine Learning Prediction:
- Accuracy Prediction: The final RBF-SVR model achieved a Mean Absolute Error (MAE) of 0.261, representing an 11.6% reduction in error compared to a baseline mean-prediction model. However, the coefficient of determination ( $R^2$ ) was negative (-0.161), and the rank correlation (Spearman $\rho$ ) was weak but significant (0.202, p=0.004).
- Completeness Prediction: The model achieved an MAE of 0.294 with an $R^2$ of -0.106 and a non-significant rank correlation.
- Feature Contribution: Ablation analyses indicated that combining NLP features, semantic similarity, and explainable textual features yielded the lowest prediction errors and highest $R^2$ values in exploratory settings, suggesting that surface-level linguistic features alone are insufficient for quality prediction.

Significance and Claims
The paper posits that current LLMs are capable of generating high-quality responses for orthopaedic patient education, but their performance is not uniform; it fluctuates based on the clinical topic and, more critically, the specific nature of the question. The study emphasizes that high semantic similarity between models does not guarantee clinical accuracy, reinforcing the necessity of expert evaluation.

Regarding the machine learning component, the authors modestly claim that while linguistic, semantic, and explainable features provide a "measurable signal" for predicting expert scores, the current predictive capability is limited. The negative $R^2$ values and weak correlations indicate that automated quality prediction cannot yet replace expert evaluation. Instead, the proposed framework is positioned as a potential tool for automated pre-screening to identify potentially lower-quality responses, rather than as a standalone clinical decision-making or quality-assessment system. The study concludes that LLMs should serve as supportive tools in patient education, with expert verification remaining essential, particularly for context-sensitive questions.

Technical Summary: Predicting Expert-Rated Quality of LLM Responses in Orthopaedics

More like this