Development of a Pilot Chatbot to Detect Interactions Between Monotherapy and Combination Therapy with Captopril Using a Local (LLM) in Cardiovascular Therapy

Ikhwan Yuda Kusuma (Universitas Harapan Bangsa)
Afriza Pujiati (Universitas Harapan bangsa)
Khamdiyah Indah Kurniasih (Universitas Harapan bangsa)
Siti Setianingsih (Universitas Harapan bangsa)
Bizhar Ahmed Tayeb (University of Szeged)
Nur Arifin Akbar (University of Palermo)
Muhammad Syaiful Aliim (Universitas Jenderal Soedirman)

Journal of Advanced Health Informatics Research
Vol. 4 No. 1 (2026)

Cardiovascular diseases (CVDs), including hypertension and heart failure, remain major contributors to global morbidity and mortality and frequently require long-term polypharmacy. The concurrent use of multiple medications increases the risk of drug–drug interactions (DDIs), which may reduce therapeutic effectiveness, increase toxicity, and contribute to adverse drug reactions. This study aimed to develop and validate a locally deployed large language model (LLM)-based chatbot for detecting drug–drug interactions involving captopril in both monotherapy and combination therapy settings. A development and validation study was conducted using DDI data obtained from DrugBank and Drugs.com. The chatbot was developed using the LLaMA-3 model integrated with LangChain, Ollama, and FastAPI and was evaluated through iterative testing and 5-fold cross-validation. System performance was assessed using accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), while usability was evaluated using the System Usability Scale (SUS) and Single Ease Question (SEQ) questionnaires completed by pharmacists. The chatbot demonstrated progressive performance improvement throughout development and achieved excellent performance during final validation involving 300 drug pairs, with an accuracy of 99.0%, sensitivity of 100.0%, specificity of 98.0%, PPV of 100.0%, and NPV of 92.0%, exceeding all predefined acceptance thresholds. Usability testing indicated only fair to moderate usability, with a mean SUS score of 65.0 and a mean SEQ score of 5.0, suggesting that further refinement of the user interface and workflow may be required. The locally deployed LLM-based chatbot demonstrated satisfactory diagnostic performance and preliminary feasibility for captopril-related DDI screening. Although the system showed promising classification performance, additional usability optimization and evaluation in real-world clinical workflows are needed before broader implementation can be considered as a pharmacist-supportive screening system.