Introduction
Chronic pain represents a significant global health crisis, imposing substantial economic and social burdens. While effective intervention often necessitates specialist knowledge, Primary Care Physicians (PCPs) typically serve as the initial point of contact. However, a critical expertise gap exists; surveys indicate that over 70% of PCPs lack confidence in treating pain due to insufficient systematic training.
Existing Clinical Decision Support Systems (CDSS) frequently struggle with the high-dimensional, sparse, and voluminous nature of Electronic Health Record (EHR) data. Traditional rule-based or case-based reasoning methods often fail to generalize effectively or demand excessive computational resources for real-time clinical application. This research introduces the Agent-based Clinical Decision Support System (ACDSS), designed to bridge the gap between general practitioners and pain specialists by providing automated, real-time clinical guidance.
System Architecture and Methods
The core of the ACDSS is an AI Engine that utilizes a multi-step pipeline to transform raw clinical data into actionable insights:
- Unsupervised Feature Learning: The system adopts a "Deep Patient" approach, employing stacked denoising auto-encoders to process raw EHR data. This allows the model to learn a general-purpose mathematical representation of a patient's health state without requiring manual labeling for every attribute.
- NLP Label Extraction: Natural Language Processing (NLP) tools are used to detect clinical phrases and extract labels—such as specific diagnoses or medications—from free-text clinical notes to train task-specific modules.
- Task-Oriented Modules: The engine is divided into four specialized components: General Diagnosis (predicting pain types and complications), General Therapy (proposing physiotherapy), Prescription (recommending medications), and Patient Referral (optimizing specialist matching).
- Agent-Based Interface: The system functions as an intelligent agent, presenting visualizations to the PCP and iteratively updating its learning model based on physician feedback.
Experimental Setup
The study utilized a proprietary database of anonymized EHRs from the pain department of Tan Tock Seng Hospital in Singapore. The dataset included demographics, health state descriptions, pain symptoms, and free-text clinical notes for over 500 patients. To ensure privacy, textual data was abstracted into fixed-length parameter representations before being processed by the AI model.
Key Results and Discussion
The system's performance was evaluated using macro-averaging accuracy across its four primary tasks. As shown in the results, the system excelled in referral and therapy tasks but faced challenges with complex medication prescriptions.
- Task Performance: Task 4 (Patient Referral) achieved the highest accuracy at 0.81, followed by Task 2 (General Therapy) at 0.69. Task 1 (General Diagnosis) reached 0.60, while Task 3 (Prescription Recommendation) had the lowest performance at 0.50.
- The Prescription Gap: The lower accuracy in prescriptions suggests that medication decisions often rely on external factors—such as drug interactions or specific contraindications—not fully captured in the primary pain descriptions.
- Complexity Ceiling: Evaluation of Task 2 revealed a significant "complexity ceiling." While the recall rate remained near 1.0 for patients requiring only 1-2 types of therapy, it plummeted to 0.34 for multi-morbid patients requiring six or more interventions. This highlights the difficulty of modeling complex, overlapping clinical symptoms.
Conclusion and Future Work
The ACDSS demonstrates the feasibility of using unsupervised "patient representations" to support multiple clinical tasks simultaneously. While the system provides a layer of privacy through mathematical abstraction, its current limitations include a relatively small training cohort (500 patients) and a need for more comprehensive medical history to improve prescription accuracy. Future work will focus on scaling the dataset, integrating real-time specialist availability into the referral module, and validating the system's efficacy through long-term PCP adoption and human-in-the-loop feedback.
Summary updated