Integrating Anomaly Detection and LLM-Based Explanation Generation in Clinical Data Dashboards.

This paper presents an integrated approach that combines unsupervised anomaly detection with large language model (LLM)-based explanation generation to enhance the interpretability of clinical study dashboards. Using data from the P4D (Personalized, Predictive, Precise, and Preventive Medicine for Major Depression) cohort study, an Isolation Forest algorithm was employed to identify atypical observations within multidimensional clinical data. A t-distributed Stochastic Neighbor Embedding (t-SNE) projection was used to assess the plausibility of detected outliers. To improve transparency for clinical stakeholders, several current-generation LLMs, including GPT-5, GPT-4, Gemini 2.5 Flash, and Llama 3, were evaluated for automatically generating concise, patient-specific explanations of anomalies. The evaluation followed a reasoning-based framework comparing model explanations with clinical assessments. Results show GPT-5 provided the most clinically helpful and format-consistent natural language outputs, while Llama 3 showed limitations in contextual precision. The integration of these components into the existing P4D Dashboard [1] for clinical study monitoring supports clinicians in identifying data inconsistencies and understanding their likely causes, thereby improving the usability and trustworthiness of automated clinical data analyses.