Jaehyun Lim

LumanLab Inc, Republic of Korea

Abstract Title:

A Three-Stage LLM Pipeline for Pediatric Development Screening from Caregiver Chat Responses

Biography:

Jaehyun Lim, MD, is the CEO of LumanLab Inc., Republic of Korea. His recent work focuses on turning caregiver free-text responses into structured developmental screening signals and building reproducible evaluation pipelines for LLM-based scoring. His interests include robust LLM evaluation, pediatric developmental screening workflows, and clinical deployment and validation of generative AI tools.

Research Interests:

Parent-reported behavior is useful for developmental screening, but responses are often brief or ambiguous. BeDevel is a chat-based screening service for children aged 18-42 months that converts caregiver responses into structured item-level assessments. Our system uses a three-stage LLM pipeline: response adequacy assessment, guided follow-up for insufficient responses, and structured scoring. Guided follow-up converts insufficient caregiver responses into evaluable inputs while preserving conversational flow. The assessment uses 18 predefined developmental items spanning social communication and behavior patterns. Children are assigned to age bands (18-23, 24-35, and 36-42 months), and each band applies a predefined subset of items from the 18-item catalog. Our system stores score, reasoning, confidence, and key observations, while tracking atypical patterns separately from risk-level counts. Risk levels are computed from non-optimal completed-item counts, and recommendation generation is linked to evaluation outcomes. Each assessment step is persisted through API-driven workflows so evaluation states and item-level outputs can be retrieved consistently across sessions. Response evaluation was performed using the Google gemini-3 model. The benchmark pipeline evaluates gemini outputs against expert labels using predefined criteria (Accuracy >= 0.85, Cohen's kappa >= 0.75, and Macro F1 >= 0.80). In benchmark results, the gemini evaluator on Item 1 (Name Recognition) achieved accuracy=0.9375 and kappa=0.872. Because outputs are structured at item level, evaluators can review per-item reasoning and key observations together with aggregate risk summaries and recommendations during follow-up decisions. The service is intended for screening support rather than diagnosis, and this implementation turns caregiver free-text input into structured screening outputs that support prospective clinical validation.