INCLAIR: Inception-Based Longitudinal Clinical Anomaly Detection with Informed Reasoning
arXiv:2607.27487
INCLAIR is a framework that detects anomalies in longitudinal clinical records by scoring observations against multiple historical contexts, aggregating evidence, and producing natural‑language explanations with limited expert supervision.
Abstract
Detecting anomalies in longitudinal clinical profiles is clinically important but difficult: abnormal evidence is often sparse, patient histories have unequal length, and expert explanations are costly. We propose INCLAIR, a framework that scores each observation against multiple historical contexts, aggregates evidence at the profile level, and generates grounded natural-language explanations under limited expert supervision. Under stated within-profile exchangeability assumptions, the complete mean subsequence score takes an order- U-statistic form, yielding a variance decomposition and an incomplete-subset approximation that controls combinatorial inference cost independently of profile length. The same analysis shows that mean aggregation attenuates localized anomalies by a factor set by the anomaly support and profile length, motivating validation-selected top- pooling. Across three clinical datasets, INCLAIR consistently outperforms state-of-the-art baselines. We further validate practical relevance through a case study on longitudinal steroid profiles, comparing INCLAIR's predictions and explanations against domain-expert assessments supported by DNA analysis. The results show that INCLAIR enables clinically actionable anomaly detection under limited expert supervision.