Presentation (CTI)

OBJECTIVES: To identify and characterize neuropsychiatric symptoms in patients with Alzheimer’s disease (AD) using text mining of unstructured electronic health record (EHR) data.
METHODS: A retrospective study was conducted using the Japan Medical Data Survey (JAMDAS), a large EHR database from general practice settings in Japan. AD patients were identified using previously published ICD-10-based methods. Neuropsychiatric symptoms (NPS) were extracted from unstructured clinical text using a predefined terminology set developed through expert review and iterative validation in collaboration with developers of an industrial natural language processing platform (NLP). Patients were stratified by the presence of documented symptoms. Among patients with documented symptoms, symptom frequency and time from diagnosis to first documentation were summarized descriptively.
RESULTS: Among 175,886 AD patients, 30.2% had ≥1 documented NPS. Demographics were comparable across groups, while psychotropic medication use was more frequent among patients with documented NPS. At the patient level, agitation-related symptoms were identified in 39,038 patients (22% of all AD patients). At the documentation level, among 194,760 NPS records, agitation-related symptoms accounted for 65%, followed by depression/anxiety (20%) and hallucinations/delusions (14%), whereas apathy accounted for only 1%. Time from AD diagnosis to first documented NPS varied by symptom category: apathy tended to be recorded earlier after diagnosis, while other symptoms showed heterogeneous patterns. Notably, some agitation-related symptoms were documented as early as or earlier than depression, anxiety, or psychotic symptoms.
CONCLUSIONS: NLP-based analysis of unstructured EHR data enabled large-scale identification and characterization of NPS in routine clinical practice. Heterogeneity in documentation frequency and timing, together with patterns broadly consistent with expected treatment use, suggests that this approach captures underlying clinical processes rather than random variation or extraction artifacts. By enabling symptom-level assessment, this methodology may support more granular evaluation of symptom burden and inform real-world evidence generation and clinical decision-making.