JMIR Publications

ABSTRACT

Background:

Internet hospitals have emerged as a innovative telemedicine model in China, providing online services like consultations, prescription renewals, and report interpretations. These platforms generate a vast amount of authentic, patient-written text describing symptoms and needs. While Natural Language Processing (NLP) and Large Language Models (LLMs) offer powerful tools for text analysis, a significant gap exists in applying these techniques to systematically decode and categorize the specific service intents within real-world clinical texts from internet hospitals. Most existing research relies on data from social media or forums, not from actual clinical practice.

Objective:

This study aimed to leverage AI to analyze patient descriptions from an internet hospital platform with three primary objectives: 1.To develop and optimize a specialized large language model for understanding unstructured clinical narratives. 2.To analyze and categorize the specific service intents expressed in patients' medical descriptions. 3.To examine the distribution of these patient intents across different clinical departments to reveal demand patterns for strategic planning.

Methods:

The study retrospectively collected 16,457 de-identified patient consultation records from a tertiary hospital's internet hospital in China. A hybrid methodology was employed: First, an unsupervised learning approach was used to discover intent categories. Patient texts were converted into embeddings using the chinese-roberta-wwm-ext-large model, reduced in dimensionality with UMAP, and clustered using K-Means, which revealed eight distinct intent categories. Subsequently, a random sample of 5,000 records was manually annotated by three clinical experts based on these categories (achieving high inter-rater agreement, Fleiss' Kappa = 0.82). This labeled dataset was then used to fine-tune the pre-trained model, and its performance was compared against traditional machine learning models (e.g., SVM with TF-IDF features).

Results:

The analysis successfully identified eight coherent patient intent categories. The most prevalent needs were "Online Interpretation of Existing Medical Reports" (25.5%) and "Initial Consultation" (24.4%). The fine-tuned chinese-roberta-wwm-ext-large model achieved the best performance with a weighted F1-score of 0.7030, outperforming traditional models. A highly significant association was found between clinical departments and the types of patient intents (χ² = 8874.241, p < .01). Based on standardized residual analysis, four distinct operational models for internet hospitals were synthesized: Centralized Efficiency Departments (e.g., Dentistry, Dermatology): Strongly associated with "Initial Consultation" and "Online Test Booking." Long-Term Management Departments (e.g., Hematology, Gastroenterology): Primarily linked to "Prescription Renewal for Chronic Diseases." Balanced Departments (e.g., Nephrology, ENT): Managed a diverse mix of intents like "Report Interpretation" and "Online Follow-up." Low-Engagement Departments (e.g., Neurosurgery, Oncology): Primarily used for "Report Interpretation" and "Remote Guidance," serving an auxiliary role.

Conclusions:

This study demonstrates that a hybrid AI methodology effectively identifies and categorizes patient service intents within internet hospitals. The findings reveal a clear, efficiency-driven allocation of patient demand, with core services being report interpretation and initial consultations, and significant variation across departments. This provides a critical evidence base for hospitals to optimize resource allocation, design specialized operational models for different departments, and strategically develop their telemedicine services. Future work should focus on refining categories, integrating multimodal data (e.g., images), and modeling temporal consultation sequences for a more comprehensive understanding.