A Human-in-the-Loop AI System for Analyzing Task Failure Factors in Infant Developmental Assessment: System Design and Preliminary Validation

Abstract

This study proposes an AI system that supports expert assessment by utilizing Generative AI to conduct a multiperspective analysis of factors underlying task non-attainment in developmental assessments, and preliminarily examines its behavior and effects on experts. Conventional assessment methods tend to record a child’s failure to complete a task simply as a “lack of ability.” In practice, however, there are cases where children possess latent abilities but fail to perform due to temperament, anxiety, or environmental factors. Overlooking this “Competence-Performance Gap” carries the risk of leading to inappropriate diagnoses and interventions. In the proposed system, the AI generates hypotheses regarding “why a task was not attained” from the perspectives of developmental psychology, early childhood education, occupational therapy, and speech-language pathology, thereby complementing the expert’s interpretation. We position this work as an exploratory case study (n=6 evaluations; 2 experts × 3 video clips from ESCS sessions), designed not to confirm false-negative reduction directly but to generate indirect process-level evidence through proxy indicators (AI-expert judgment agreement, hypothesis count change, confidence shift, and category shift of hypothesis labels). The observed patterns suggest that the AI and the experts held complementary perspectives. While the AI support did not alter the final judgments themselves, it was accompanied by consolidation of hypotheses (mean 3.3 → 2.0) and slightly elevated confidence in the majority of cases. Conversely, the issue of “over-interpretation” emerged as a challenge for the AI. It was also suggested that the detailed behavioral descriptions, introduced to suppress hallucinations, might have inadvertently encouraged the AI to focus excessively on localized actions. These findings are hypothesis-generating rather than confirmatory; direct verification of false-negative reduction requires a subsequent larger-scale study with longitudinal ground truth.