A Knowledge Graph and Large Language Model Framework for Personalized Clinical Decision Support in Bipolar Disorder and Borderline Personality Disorder: The KG- Psy Framework
Abstract Background: Differentiating between bipolar disorder (BD) and borderline personality disorder (BPD) is an intriguing clinical challenge that has attracted research attention, as the multiple co-occurring symptoms such as mood instability, impulsivity, and emotional dysregulation have made the boundary between the two disorders unclear. Recent research in computational psychiatry has shown promise in leveraging the power of knowledge graphs and large language models to enable accurate and interpretable decision-making. Methods: The study developed a hybrid framework, KG-Psy, that combines large language models with dynamic graphs. The system extracts clinical entities through GPT-5 while employing multi-step reasoning to find hidden connections within a psychiatric knowledge graph that weighs its evidence-based content. The system performs four main operations, which include extracting clinical entities, creating dynamic weighted knowledge graphs, using GPT-5 to trace multi-step cognitive pathways, and evaluating customized treatment options. The researchers used 310 anonymized psychiatric case reports to test their method. Results: KG-Psy achieved better discrimination between BD and BPD than other methods. The model achieved a maximum Precision (%) of 91.5, an improvement of 19%, 10.3%, and 7.4% over the KG-only, LLM-only, and Hybrid-other approaches, respectively. The system successfully discovered unknown cognitive pathways, which it used to create accurate pathway maps and provide treatment recommendations supported by numerical evidence. The results of ablation experiments showed that LLM reasoning, combined with the knowledge graph structure, led to high performance. Conclusion: The KG-Psy system introduces a novel approach to psychiatry inspired by large-scale language models and dynamic graphs that integrates data-driven linguistic analysis with established symbolic knowledge to enhance the capabilities of differential diagnosis and treatment development for complex mood and personality disorders. Future work could include validation on larger datasets and integration of multimodal clinical data.
Authors
- Reza Rahaeimehr (ORCID: https://orcid.org/0000-0003-0305-3661)
- Zahra Mahmoudpour Harris
- Saba Hesaraki (ORCID: https://orcid.org/0009-0003-3039-8793)
Institutions
- Islamic Azad University, Tehran (IR)
- Augusta University (US)
- Shariaty Technical College (IR)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-16
- DOI
- https://doi.org/10.21203/rs.3.rs-9621282/v1
- Primary Topic
- Artificial Intelligence in Healthcare and Education
- Type
- preprint